homemaker-layout/DESIGN.md

homemaker — Design & Plan

Status: validated direction, pre-implementation. Reviewed against the Urb source 2026-06-12; review findings folded in (see §4.5 evidence note, §4.6 throughput arithmetic, §5 decision 6, §6 port-scope expansion, §7 re-scoped phases, §8). Audience: a fresh session that will break this into bd (beads) tasks (note: no beads database exists yet — run bd init first). Self-contained — assumes no memory of the originating conversation.

---

1. Purpose

homemaker-layout is a clean-room Python successor to the Perl Urb project (/home/bruno/src/urb). Urb models a building as a binary slicing tree and evolves layouts with mutation + crossover, scored against Christopher Alexander–style pattern fitness. Two long-standing problems motivate the rewrite:

1. It doesn't scale — beyond a few rooms, evolution never finds layouts an architect would consider obvious.
2. Local minima — even small programmes converge to poor optima.

The eventual goal is a 100% Python system. During bring-up, Perl Urb is kept as a throwaway fitness oracle behind the .dom file format.

---

2. Constraints that fix the representation

These come from the problem domain and are not negotiable; importantly, they vindicate the slicing tree rather than argue against it:

• Multi-storey with stacked walls. An upper storey retains the storey below, except additional divisions/undivisions. Load-bearing walls must stack ⇒ every cut is a full edge-to-edge guillotine cut. Urb already enforces this via Below-inheritance (an upper quad reads its geometry from the matching quad below).
• Quadrilateral rooms only (no L/Z shapes) — recursive bisection produces exactly this.
• No pinwheel / non-slicing layouts — undesirable for load-bearing construction and adaptability (cf. Brand, How Buildings Learn). This is the one class a slicing tree can't express, and we don't want it anyway.
• Plots are near-rectangular but general convex quadrilaterals (not axis-aligned). Geometry must handle skew; the slicing combinatorics are unaffected.

Conclusion: the slicing tree is the correct phenotype. The rewrite is about the genotype, the search, and the fitness shape — not about leaving the slicing class.

---

3. What we built this session (all committed)

Package src/homemaker_layout/:

• dom.py — .dom YAML ⇄ Node tree. Linkage (parent/below/position), wall_outer inset on load with raw-corner stash for byte-perfect round-trip, emit.
• geometry.py — faithful port of Urb's top-down geometry (Coordinate/Coordinate_a/_b/Area/Length) + Coordinate_Offset wall inset. Memoised (uncached recursion is exponential in depth).
• programme.py — parse patterns.config spaces: into per-code size/width/proportion/adjacency/level/count requirements.
• solver.py — bottom-up division-ratio solver (scipy least_squares). (Outcome: falsified as a standalone component — see §4.2.)
• oracle.py — Phase-1 fitness bridge: write .dom, run urb-fitness.pl, parse .score + .fails.

Experiments in experiments/: dump_areas.{py,pl}, resolve_ratios.py, refine_sweep.py, sweep_failtypes.py, optimize_fullfitness.py.

---

4. Empirical findings (the core of this document)

4.1 Geometry port — VALIDATED

Per-leaf areas computed in Python are byte-identical to Urb across all 35 programme-house .dom files, including the wall inset and multi-storey wall-stacking inheritance. (experiments/dump_areas.{py,pl}.) The infrastructure is trustworthy.

4.2 Bottom-up area-proxy sizing solver — FALSIFIED

The original hypothesis: give leaves target sizes, solve cut ratios bottom-up, let the EA search only topology. Tested by re-solving an evolved candidate's ratios from programme targets and scoring via the oracle.

• resolve_ratios.py on candidate-002: areas recovered accurately (errors collapsed, e.g. t1/t2/t3 from +1.4/+2.4/+4.8 → ~+0.05), and it fixed the original's size failure — but total fitness dropped (0.00737 → 0.00065, 4 fails) because it introduced shape/relational failures.
• refine_sweep.py (warm-start refine of all 34 candidates): 0/34 improved. Total failures 124 → 297 (equal-offset cuts) and 124 → 626 (independent-offset cuts).
• sweep_failtypes.py (failure-type histogram, equal-offset):

  type	area-dominant Δ	shape-aware Δ
  width	+82	+29
  proportion	+35	+7
  crinkliness	+18	+4
  adjacency	+18	+13
  size	−15	+15
  access	+29	+39
  total added	+173	+110

Why it fails: in Urb's fitness, every cut position is simultaneously a size knob and an adjacency/access/shape knob. A solver that optimises only size/shape is blind to access/adjacency and trades them away. Refining a co-evolved local optimum with a partial objective is structurally unable to win, and the 0.5^n failure penalty makes every new failure catastrophic while fixes are only linear. The proxy solver is strictly worse than optimising real fitness. Do not pursue it.

4.3 "Perpendicular" failures were an artifact — RESOLVED

Letting the two ends of a cut float independently produced skewed cuts and many perpendicular failures. Tying the two ends (equal offset, a == b, one DOF per cut) produces near-perpendicular walls on these near-rectangular plots and yields zero perpendicular failures. Equal-offset cuts are the only mode to use. This also halves the variable count and matches the slicing model.

4.4 DOF / over-determination — partially real, not fatal

A topology with R rooms has ~R−1 cut DOF but ~2–3 size/shape constraints per room, so a fixed topology can be over-determined: you cannot always hit area + width + proportion for every room at once (heavy shape weighting traded straight into size, §4.2 table). This limits any single-objective sizing pass — but it is not fatal, because optimising the full objective still found large gains (§4.5). The earlier "infeasibility" worry was overstated.

4.5 Full-fitness frozen-topology optimisation — VALIDATED ✅

Drive the equal-offset ratios with Nelder-Mead against the real oracle fitness (whole objective, no proxy), topology frozen (experiments/optimize_fullfitness.py):

candidate	DOF	original	optimised	gain	fails
2f45907 (best evolved)	7	0.012617	0.015684	×1.24	2→2
candidate-002 (MCP-refined)	6	0.007375	0.012319	×1.67	2→2
c964435 (MCP baseline)	6	0.003667	0.005836	×1.59	3→3

Every design improved 24–67%, none added a failure. Headroom widens on weaker designs. Because the optimiser sees the whole objective (including the 0.5^n penalty), it never trades into a new failure — the cliff that destroys the proxy solver protects the full-objective optimiser.

Implications:

• There is large, unclaimed geometry headroom above every EA design — even the best. Urb's EA under-optimises geometry: source inspection confirms slide() (Mutate.pm:256-269) re-randomises the cut position uniformly across the span — Urb has no fine-tuning geometry operator at all, which fully explains the headroom.
• A full-objective geometry inner loop is genuinely valuable (the proxy solver is not).
• The EA/search should therefore own topology; geometry is delegated to the inner loop. This is the memetic architecture (§5).
• Corroboration for §4.3: Urb's own mutations use equal offsets (Divide($division, $division)) — equal-offset cuts match how every corpus design was generated.

4.6 Oracle throughput (measured)

urb-fitness.pl scores many .dom files per invocation, so the Perl startup (~0.65 s) amortises across a batch and cached fields (e.g. occlusion) persist. Measured on the 35-file corpus: 0.99 s/dom batched vs 1.65 s/dom for a single-file call. The cost is assessment-dominated (~1 s/dom of actual work), so startup amortisation gives ~40% — useful but bounded.

Consequences:

• Batching only helps when evaluations are submitted together — favour population/parallel-evaluating optimisers (CMA-ES, differential evolution, island EA, pattern search) over inherently sequential ones (Nelder-Mead), both inner loop and outer search, so a whole generation scores in one oracle call.
• Do the arithmetic before scoping topology search on the oracle. §4.5 used ~200 inner evaluations per topology ⇒ ~3 min/topology at 1 s/dom. A run comparable to urb-evolve (pop 128 × 768 generations) is years of oracle time; even 32 topologies × 100 generations with a trimmed 50-eval inner loop is ~2 days. Therefore:
  • The oracle supports Phase 1 fully and Phase 2 only as a small-scale proof (tens of topologies, budgets counted in oracle calls).
  • A native Python fitness is effectively a gate for topology search at any real scale — not merely a later optimisation. (It also brings independence, penalty reshaping, and large programmes.)
  • Warm-starting the inner loop from the parent's optimised ratios (Lamarckian inheritance, §5 decision 6) is the main lever for cutting the per-topology cost — with high-locality moves most cuts survive a mutation, so an order-of-magnitude reduction is plausible. Measure this in Phase 1.

4.7 Occlusion-disabled re-baseline (measured 2026-06-12)

With the §6 descope in place (URB_NO_OCCLUSION=1 patch in Urb), the corpus re-baseline (experiments/rebaseline_no_occlusion.py): all 35 scores change (mostly up, ×1.0–×1.24 — daylight terms pin to 1), exactly one failure-set change (458aa8b8 gains two crinkliness fails — expected mechanism: no shading discount on external wall area), batched oracle ~8% faster (0.92 s/dom). New inner-loop reference gains (deterministic seed, budget 400, accept_innerloop.py bars): 2f45907 0.01304→0.02128 (×1.63), candidate-002 0.00808→0.01373 (×1.70), c964435 0.00400→0.00674 (×1.68, fails 3→2); ~35 oracle calls per topology. All Phase-2+ work uses the flag; flag-off numbers above are historical.

4.8 The 0.5^n failure penalty is a first-order pathology

Multiplicative 0.5^n over failure count (a) makes the landscape a cliff (no gradient across the huge zero-feasibility region), (b) rewards fewer flags over better geometry (the original outscored better-sized solved designs purely on flag count), and (c) is representation-independent. Reshaping it (additive / soft / multi-objective Pareto) is a high-leverage change that helps Urb today and homemaker tomorrow.

4.9 Penalty reshaping decision: lexicographic outer search (measured 2026-06-14)

experiments/penalty_reshape.py, URB_NO_OCCLUSION=1, programme-house.

Inner-loop protection (nm_search, budget 80, 3 files × 3 seeds = 9 runs): All runs show n_fails ≤ x0_n_fails. 0/9 regressions. The 0.5^n cliff in the native fitness scalar is unchanged and continues to protect the inner loop.

Outer-search comparison (budget 3000, 3 seeds, seed = 2f45907):

scheme	seed	best	fails	note
lex	0	0.01781	2
lex	1	0.01793	2
lex	2	0.01785	2
scalar	0	0.01781	2	(same outcome)
scalar	1	0.01890	3	trapped by high-score 3-fail design
scalar	2	0.02632	2	(different topology path)

lex mean: 0.01786 / 2.00 fails. scalar mean: 0.02101 / 2.33 fails.

Key result (seed 1): scalar promoted a 3-fail design whose raw score (×0.125 penalty) beat the pool's 2-fail candidates — exactly the §4.8 pathology. Lexicographic comparison (-n_fails first, then fitness) is immune: any 2-fail design beats any 3-fail design regardless of raw score. Within a homogeneous fail tier both schemes are identical (seeds 0 and 2 agree in serendipitous runs where scalar also stays in the 2-fail tier).

Decision: lexicographic. 0.5^n stays in the fitness scalar (inner loop unchanged). Outer search uses (-n_fails, fitness) as comparison key.

4.10 Deceptive level-fix valley and compound operators (measured 2026-06-14/15)

Context: programme-house, Phase 3 native fitness + Phase 4 lex search, seed warmstart-2f4.dom (best Phase-3 result, 2 fails at score 0.032). Goal: reach ≤ 1 fail, beating the Perl optimiser (2–3 fails).

The deceptive valley. The 2-fail state has l1 (living room, min 27 m², required level 0) on level 1. The obvious repair is level_fix: swap l1 with a leaf on level 0. But every single-step level_fix move creates 5+ new fails because the displaced room (t3, the WC) is dropped into an arbitrary slot that violates adjacency, size, and access constraints simultaneously. The lex comparator (-n_fails, fitness) correctly rejects these — but the result is that the 2-fail state appears completely surrounded by ≥ 5-fail states, and the search stalls. This is a textbook deceptive valley: the fitness gradient points away from the global optimum.

Compound operator. mutate_level_compound_fix (added operators.py) escapes the valley by doing two things atomically:

1. Move l1 to level 0 by swapping it with the largest leaf there (the circulation C node, because C is generic and can absorb the swap without producing a new structural failure).
2. Re-insert the displaced t3 by dividing the sibling of that C node (so t3 lands adjacent to C, satisfying the adjacency requirement).

The new split gets division=[0.25,0.25] (giving t3 ≈ 3.4 m², barely in range) and rotation=0 (t3 on the left, adjacent to the C sibling).

The warm_x0 initialization bug. The compound operator sets specific ratios on a newly-created split node. But driver.py was initialising the NM inner loop from parent.ratios, which has no entry for the new node (it was a leaf). warm_x0 defaulted the new node to 0.5, giving t3 ≈ 6.8 m² — a size fail — so NM started at 3 fails instead of 1. Lex then always rejected the compound child; level_compound_fix was completely invisible to the outer search for ~12 000 evals (until warm_x0 was fixed).

The correct fix distinguishes genuinely-new split nodes from stale hidden nodes that become visible after structural mutations (e.g. swap can flip a b.below pointer, revealing pre-writeback division values from a different topology). Only use the child's explicit ratio for node (li, path) if the matching node in the parent was not already divided; everything else falls through to parent.ratios or defaults to 0.5. Fix in driver.py lines 259–267.

Results (50 000 evals each, pop 8, child_budget 80, 4 workers):

seed	event	eval	fails	score
warmstart-2f4	seed	200	2	0.032
warmstart-2f4	level_compound_fix fires	12 280	1	0.000122
warmstart-2f4	level_retype 0/ll<->1/l	17 880	1	0.00497
warmstart-2f4	final	50 040	1	0.00518
compound3-raw	seed (1-fail hand-built)	200	1	0.000118
compound3-raw	level_retype 0/ll<->1/l	18 360	1	0.00383
compound3-raw	final	50 040	1	0.00523

Perl optimiser reference: 2–3 fails.

The two-C topology breakthrough. After level_compound_fix fires, the topology is: level 0 = ll(l1), lr(t2), rl(C), rrl(t3), rrr(O) — but now l1 is at level 0 (correct) and t3 is adjacent to rl(C) (staircase). However l1 is occupying ll, and rl(C) is the staircase core — so t3-adj-C is satisfied via rl, but there is no second C to satisfy staircase independently. Score ≈ 0.000157 (1 fail).

At eval ≈ 18 000, level_retype 0/ll<->1/l (swap the type of ll on level 0 with l on level 1) creates a TWO-C configuration at level 0: ll(C), lr(t2), rl(C), rrl(t3), rrr(O), with l1 moving to level 1. The score jumps 25× to ≈ 0.005. Why two C nodes work:

• ll(C) (bottom-left, 23 m²) satisfies t3-adj-C via geometric contact at the l/r zone boundary with rrl(t3).
• rl(C) (top-right, 8.5 m²) satisfies staircase adjacency via tree adjacency to rrr(O) (its right sibling when r.rotation=3).

Both constraints are simultaneously met because binary-tree sibling adjacency and cross-zone geometric adjacency provide independent paths.

Why 0 fails is geometrically impossible on this programme + plot. l1 needs min 27 m² at level 0. The only space large enough is ll (≈ 23 m², the entire left half of level 0). Putting l1 at ll removes the t3-adj-C provider. The alternative — dividing ll into lll(l1)+llr(C) — gives llr a proportion of ≈ 6:1 (width ≈ 0.73 m), failing both the proportion and width constraints. 0 fails is not achievable on this programme+plot with a binary slicing tree representation; 1 fail is the geometric optimum.

---

5. Validated architecture

Memetic search, full objective throughout:

            ┌─────────────────────── topology search (OUTER) ───────────────────────┐
            │  genome = slicing topology + per-leaf type assignment + per-floor       │
            │           divide/undivide deltas (base floor is master)                 │
            │  operators = high-locality topology moves (see §6)                      │
            │                                                                          │
            │   for each proposed topology:                                            │
            │      ┌──────────── geometry inner loop ────────────┐                     │
            │      │ optimise equal-offset cut ratios (1 DOF/cut) │                     │
            │      │ against the FULL fitness (derivative-free /  │                     │
            │      │ gradient), to convergence                    │                     │
            │      └──────────────────────────────────────────────┘                    │
            │      score = best full-fitness over inner loop                            │
            └──────────────────────────────────────────────────────────────────────────┘
                                   fitness: NATIVE Python (fast), reshaped penalty

Key decisions, all evidence-backed:

1. Geometry = inner optimisation against full fitness (§4.5), not an area proxy (§4.2). Equal-offset cuts, one DOF per free branch (§4.3).
2. Search owns topology only. The base-floor tree is the primary genome; per-floor deltas are a small secondary genome (multi-storey constraint as a regulariser, via Below-inheritance).
3. Prefer population/batch-evaluating optimisers so the batched oracle is efficient (§4.6). A native Python fitness (faithful to Urb, validated against the oracle on the 35-file corpus) gates topology search at scale (§4.6 arithmetic); the oracle suffices for the inner loop and a small-scale topology-search proof only.
4. Reshape the failure penalty (§4.8) — additive/soft or multi-objective — so the search has a gradient and isn't dominated by flag-count. Caution: the 0.5^n cliff is what protects the inner loop from trading into new failures (§4.5); reshaping must not lose that property. Candidate resolutions: keep the cliff inside the inner loop only, lexicographic ordering (failure count first, score second), or genuine multi-objective Pareto. Decide in Phase 4 with measurements.
5. Representation upgrade (later): canonical slicing encoding (normalized Polish expression / skewed slicing tree, Wong–Liu) for redundancy-free, high-locality topology moves; bottom-up shape feasibility checks. Defer until the inner loop + native fitness are in place.
6. Lamarckian geometry inheritance. A child topology's inner loop warm-starts from the parent's optimised ratios (cuts that survive the topology move keep their values; new cuts get heuristic defaults). This is the main cost lever for the memetic loop (§4.6) and a standard memetic design choice (Lamarckian vs Baldwinian — we write the optimised geometry back into the genome). Validate the warm-vs-cold speedup in Phase 1.

What we are not doing: the bottom-up area-proxy solver; independent-offset cuts; non-slicing representations (sequence-pair/B*-tree — excluded by §2).

---

6. Component plan

component	status	notes
dom.py (I/O + linkage)	✅ done	round-trips byte-perfect; keep
geometry.py (port + cache)	✅ done, validated	the trusted geometry kernel
programme.py	✅ done	extend as fitness needs grow
oracle.py (Perl bridge)	✅ done	throwaway; the validation reference
solver.py (area proxy)	⚠️ keep as artifact	falsified; do not build on it
geometry inner loop	❌ to build	full-objective ratio optimiser (DOF = free branches); batch/population so the oracle batches; warm-start support (§5.6)
topology genome + operators	❌ to build	base tree + per-floor deltas; high-locality moves
search driver	❌ to build	memetic EA / SA over topology; small-scale on oracle, full-scale needs native fitness
native fitness	❌ to build	gates topology search at scale (§4.6); port + validate vs oracle; scope is larger than the term list — see below
penalty reshaping	❌ to design	additive/soft or multi-objective; must preserve inner-loop cliff protection (§5.4)
canonical encoding (Polish expr.)	❌ later	representation upgrade once core lands

Urb fitness terms the native port must reproduce (all couple to geometry): size, width, proportion, adjacency, access/inaccessible, crinkliness, perpendicular, level, staircase volume/count, public access, circulation & outside ratios, min internal area. Source of truth: /home/bruno/src/urb/lib/Urb/Dom/Fitness/ProgrammeDriven.pm and the Storey/ Building/Leaf/Base submodules.

Port scope beyond the term list (found by source review — budget for these):

• Daylight + occlusion subsystem — DESCOPED (decision 2026-06-12). Occlusion is orthogonal to building a scalable optimiser. Instead of porting Urb::Misc::Sun/Urb::Field::Occlusion/CIESky, disable it in Urb behind an env flag (quality_daylight → 1 everywhere; Crinkliness/Area_Outside pins the CIEsky_vertical illumination factor to 1 — simple crinkliness = unweighted external wall area / floor area). The boundary-overlap geometry (Dom->Walls) stays in scope; the sky model does not. The native fitness ports simple crinkliness only; a Python occlusion subsystem is rebuilt post-Phase-5 once optimisation is fully native. Flipping the flag changes every score — re-baseline the corpus, the §4.5 table, and gate bars at one clean boundary, and run the Phase-2 urb-evolve benchmark under the same flag.
• The cost denominator. Fitness is value/cost: per-leaf area costs, interior/exterior wall edge costs, boundary costs (Leaf.pm:194-251, Storey.pm:122-147). Cost couples to geometry too.
• Structural failures not in the term list: "edge too long" (>8 m, two variants), "unsupported covered outside", "covered outside above ground", "level N not connected".
• Missing-space failure stacking (ProgrammeDriven.pm:192-212): a missing space generates 2 base failures plus one per size/width/proportion/adjacency/ level requirement — up to ~7 failures. Penalty reshaping (Phase 4) must preserve this hierarchy or the search will happily drop rooms.
• Two-phase graph build: adjacency/level/vertical checks run on the unmerged tree; graphs are rebuilt after Merge_Divided for storey processing (ProgrammeDriven.pm:83-103). Easy to get subtly wrong; the 35-file validation gate will catch it, but anticipate it.
• Known stub to decide on (fidelity-vs-fix, §8.1): has_vertical_connection (ProgrammeDriven.pm:399-423) matches any leaf of the target type anywhere on the level below — no spatial-overlap check. A faithful port reproduces the bug; decide explicitly.

---

7. Phased roadmap

• Phase 0 — diagnostics (done): geometry port validated; proxy solver falsified; full-fitness geometry headroom validated; oracle throughput measured (~1 s/dom batched).

• Phase 1 — geometry inner loop (on batched oracle): full-objective ratio optimiser; use a population/batch optimiser so a generation scores in one oracle call. Reproduce/exceed the §4.5 gains. Integrate as optimise(topology, x0=None) -> (geometry, fitness). Two cheap experiments belong here: (a) warm-vs-cold start — quantify the §5.6 speedup; (b) optimiser bake-off — DOF is only ≈ rooms−1, so batched multi-start pattern search may beat CMA-ES on simplicity; measure, don't commit blind. Gate: match §4.5 gains at materially lower oracle-call budget.

• Phase 2 — topology search, small-scale proof (on batched oracle): base-tree + per-floor-delta genome, high-locality operators, memetic driver wrapping the Phase-1 inner loop. Explicitly small (§4.6 arithmetic): tens of topologies, budgets counted in oracle evaluations, not generations. Compare against urb-evolve from the same seeds/programmes at equal oracle-call budget (urb-evolve has diversity injection/culling baked in, so generations are not comparable). Gate: memetic loop beats equal-budget urb-evolve. Scaling up waits for Phase 3.

  Gate result (homemaker-py-way, 2026-06-13, URB_NO_OCCLUSION=1, budget 2000): experiments/benchmark_vs_urbevolve.py; urb-evolve scores unchanged, memetic scores corrected (patterns.config missing from re-score cwd in first run, fixed in same session).

  seed	system	best@1000	final@2000	fails
  init.dom	memetic	8.84e-10	3.37e-09	18
  init.dom	urb-evolve p16	9.10e-06	9.36e-05	6
  init.dom	urb-evolve p128	4.83e-09	3.27e-05	6
  c964435	memetic	7.65e-03	7.65e-03	2
  c964435	urb-evolve p16	4.00e-03	4.00e-03	3
  c964435	urb-evolve p128	4.00e-03	4.00e-03	3
  2f45907	memetic	2.13e-02	2.13e-02	2
  2f45907	urb-evolve p16	1.30e-02	1.30e-02	2
  2f45907	urb-evolve p128	1.30e-02	1.30e-02	2

  Verdict: 2/3 seeds → REVIEW.

  • Seeded designs (c964435, 2f45907): memetic beats urb-evolve by 1.91× and 1.63×; topology search adds value over the inner-loop-only reference (crossover finds a better topology at eval 372 for c964435).
  • Blank-slate (init.dom): memetic stalls at 18 fails after 2000 evals; urb-evolve reaches 6 fails. The 0.5^n cliff means each fail adds ~2× penalty; 12-fail gap = ×4096. Root cause: single-seed topology mutation chain builds structure one room at a time; urb-evolve's random-population initialisation explores broader topology diversity upfront. Not a regression — this is a scope gap: blank-slate construction is harder than seeded improvement, and addressed separately (random multi-start bootstrap, or Phase 4 penalty reshaping which flattens the fail cliff).
  • The memetic loop is confirmed correct and competitive on the realistic use case (seeded designs). Phase 3 (native fitness) unblocks scaled runs where this gap will also narrow.

• Phase 3 — native Python fitness (gates scaled topology search): first disable occlusion/daylight in Urb behind an env flag and re-baseline (§6 descope note); then port Urb's programme-driven fitness — the §6 "port scope beyond the term list" items (simple crinkliness, cost denominator, structural failures, failure stacking, two-phase graph build). Validate score + failure set against the flagged oracle across the 35-file corpus (float tolerance, identical failure sets). Swap behind the same interface; retire the oracle. Then re-run Phase 2 at scale.

  Gate result (homemaker-py-ccw, 2026-06-13, URB_NO_OCCLUSION=1, budget 20000): experiments/run_search_scaled.py; native fitness only, no oracle. pop_size=16, child_budget=80, seed_budget=300. 71.8 evals/s, 279.8s elapsed.

  programme-house, seed c964435 vs Phase-2 and urb-evolve references:

  seed	system	budget	best	fails
  c964435	memetic Phase-2 (oracle)	2000	7.65e-03	2
  c964435	urb-evolve p16	—	4.00e-03	3
  c964435	urb-evolve p128	—	4.00e-03	3
  c964435	memetic Phase-3 (native)	20000	1.04e-02	2

  Verdict: PASS.

  • Best 1.04e-02 beats Phase-2 oracle run (7.65e-03) by 1.36× and urb-evolve p128 (4.00e-03) by 2.60×; both at 2 fails.
  • Winning topology found at eval 10357 via rotate 1/ll — unreachable within the Phase-2 budget of 2000.
  • Population diverse: 16 members, all at 2 fails (top 15), range 5.99e-03–1.04e-02.
  • Throughput 71.8 evals/s vs ~0.5 evals/s for the batched oracle (≈140× speedup).
  • harbor-house (16 rooms, oracle-impossible): run attempted, results below.

  harbor-house (16 rooms, budget 10000): seed 2b51b05 (best corpus design, 48 fails raw):

  system	budget	best	fails	evals/s
  oracle	—	impossible	—	—
  memetic Phase-3 (native)	10000	3.73e-18	49	15.8

  Search found 3.73e-18 vs seed inner-loop baseline 8.73e-19 (4.3× lift). 638 topologies in 633s. 49-fail landscape: still many fails, but topology search is finding structure (best 3 population members all at 49 fails). The 16-room programme is qualitatively beyond the oracle's capability — this run is only possible with native fitness.

• Phase 4 — penalty reshaping (done, homemaker-py-yg5, 2026-06-14): Decision: lexicographic outer-search comparison (see §4.9). Inner loop unchanged — still uses raw 0.5^n fitness scalar (cliff protection preserved, §5.4). Outer search compares individuals by (-n_fails, fitness): fewer fails always beats more fails; within a tier, compare by score. Implemented in driver.search(use_lex=True). _CHILD_INNER_KW stale sigmas entry also removed (NM default has no sigmas parameter).

• Phase 5 — representation upgrade: canonical slicing encoding (Polish expression) + bottom-up shape feasibility; scale to larger programmes.

Each phase has a concrete go/no-go gate; do not advance on faith.

---

8. Risks & open questions (decisions for the next session)

1. Native-fitness fidelity vs simplification. Port Urb's fitness exactly (maximise comparability) or take the opportunity to clean up known issues (the 0.5^n cliff, the t3 width-default contradiction below, the has_vertical_connection no-overlap stub — §6)? Recommend: port faithfully first (bugs included), validate, then reshape in Phase 4.

2. Programme contradictions exist. e.g. t3 (3 m² WC) inherits the 4 m width_inside default (Fitness/Base.pm:60) — geometrically impossible; the original "passes" only by failing size instead. Confirmed in source. Need a sane width default scaled to area, or per-room widths.

3. Inner-loop optimiser choice — RESOLVED (homemaker-py-d0s, 2026-06-13). Bake-off over 3 files × 4 methods × 3 seeds at budget 200 (experiments/bakeoff_innerloop.py), cold-start, URB_NO_OCCLUSION=1:

   method	x@40	x@80	x@200	s/eval	oracle calls	fails+
   Nelder-Mead	1.45	1.50	1.56	2.05	200	0
   CMA-ES	1.09	1.32	1.41	1.69	18	0
   compass	0.71	0.92	1.48	1.69	12	3
   compass-ms	0.71	0.92	0.92	1.44	13	4

   Decision: keep CMA-ES (already the default) for the Perl oracle era. Nelder-Mead wins quality per eval (+x0.15 at @200) but is inherently sequential — 200 Perl invocations vs 18 for CMA (§4.6 batching matters). Compass stalls on narrow-valley landscapes (2f45907: x0.62 vs x1.30) and introduces fail regressions 3/9 runs. Multi-start compass wastes budget on phase splits.

   Phase 3+ note: once native fitness replaces the oracle, oracle-call count disappears. Revisit Nelder-Mead then — its quality advantage is real. Gradient-based (autograd through native fitness) is also an option.

4. Search algorithm for topology. Memetic GA (keep crossover — now meaningful, since a subtree = a contiguous region) vs simulated annealing (the floorplanning workhorse with M1/M2/M3 moves on Polish expressions).

5. Penalty reshaping vs inner-loop protection — RESOLVED (homemaker-py-yg5, 2026-06-14). Lexicographic outer-search comparison (§4.9). Inner loop unchanged.

6. Other continuous DOF are out of scope for Phase 1 — deliberately. Floor-to-floor height is an Urb mutation (Mutate.pm:279-291, bounded 2.7–3.6 m) and feeds cost and stair fit; stair riser/width similar. Cut ratios dominate. Revisit (+1 DOF per storey) if Phase 2 plateaus.

7. End-state confirmed: 100% Python; Perl oracle is scaffold only.

---

9. How to reproduce (for the next session)

cd /home/bruno/src/homemaker-layout
# deps: pyyaml numpy scipy (shapely networkx for later phases)

# geometry port vs Urb (must be identical):
for d in /home/bruno/src/urb/examples/programme-house/*.dom; do
  diff <(perl -I/home/bruno/src/urb/lib experiments/dump_areas.pl "$d") \
       <(python3 experiments/dump_areas.py "$d") || echo "MISMATCH $d"
done

python3 experiments/resolve_ratios.py        # proxy solver (falsified)
python3 experiments/sweep_failtypes.py       # failure-type histogram
python3 experiments/optimize_fullfitness.py 200   # full-fitness headroom (validated)

Oracle invocation (see oracle.py): cwd = the .dom's directory (so patterns.config is found), perl -I<urb>/lib <urb>/bin/urb-fitness.pl <file>, env DEBUG=1 to defeat the skip-if-newer cache; reads <file>.score and <file>.fails.

---

10. Key gotchas discovered (carry forward)

• Wall inset: the .dom plot is the outer boundary; Urb insets the root by wall_outer on load (Urb::Dom::_deserialise, Dom.pm:458) and offsets back out on save. geometry.offset_quad mirrors it; dom.py stashes raw corners in node_file. Skipping this makes all areas ~14% too large.
• Multi-storey Below-inheritance: an upper quad's coordinates come from the matching quad below; a cut is "owned" by the lowest storey where its path is divided (solver.free_branches selects these). Walls stack for free.
• Geometry must be cached — the pull-based recursion is exponential in depth otherwise (geometry._cache, cleared on dom.load and after each solver mutation).
• Equal-offset cuts (a == b) ⇒ perpendicular walls, 1 DOF/cut. Independent offsets are wrong.
• 0.5^n cliff dominates fitness; it punishes new failures catastrophically (good for the inner loop, brutal for search gradient).
• Oracle ≈ 1 s/dom batched (1.65 s single; assessment-dominated, startup ~0.65 s amortises across a batch). Submit many .doms per call and prefer population optimisers; native fitness is a later speed/scale win, not a gate.

---

11. Phase 6 — topology-search quality for full / multi-storey programmes

Epic: homemaker-py-c4c. Status: scoped 2026-06-17, pre-implementation. This section is the experiment ledger for the epic; each subsection is stubbed now and filled in by the session that runs the experiment (record the command, the numbers, and a one-line verdict, in the style of §4).

11.0 Diagnosis (why this phase exists)

The delivered speedups landed in the two layers that were never the bottleneck. The native fitness (~140× over the oracle, §7 Phase 3) and the geometry inner loop (~1.6×, §4.5/§4.7) both operate within a fixed topology: the inner loop polishes geometry inside a failure tier and, by design, the 0.5^n cliff stops it ever changing the failure count (§4.5: 0-fail-change across the headroom table). But final design quality is dominated by failure count, which is almost entirely a topology property. So faster fitness and better geometry do not move the number an architect would notice.

Topology search on full programmes is the weakness:

• blank-slate programme-house (init.dom): memetic stalls at 18 fails; urb-evolve reaches 6 (§7 Phase 2 verdict).
• harbor-house (16 rooms): out1.dom = 74 fails, generated.dom = 130 fails, both at ~machine-epsilon score; failures dominated by missing-room stacking (each missing room stacks critical + size + width
  • adjacency + level, §6).

Smoking gun: operators.mutate_divide (operators.py:71) types each new leaf at random from programme-codes + C + O. Nothing makes the required programme spaces a constructive invariant, so on a large programme required rooms simply go missing → catastrophic 0.5^n stacking, and the search is a random walk over type assignments with a flat-and-catastrophic gradient in the high-fail regime.

Causal frame for the fixes. The base-floor tree is the master genome; upper storeys are divide/undivide deltas (Below-inheritance); the programme partitions rooms by required level (harbor: 10 on L0, 4 on L1, 2 free). So construction and search should follow the genome's dependency order — credible base floor first, upper floors as deltas, with each floor's required-room set known from the programme. Do not hard-freeze the base when adding floors: that recreates the §4.2 partial-objective trap at the topology level (a base optimised purely as a ground floor can be a bad substrate — the vertical core must stay aligned and load-bearing walls must stack).

11.1 Premise experiment: single-storey harbor (homemaker-py-c4c.1) — DONE

Built examples/harbor-house-l0/ from harbor by retaining only the 10 space codes explicitly marked level: 0 (cr1, ef1, da1, k1, ws1, m×3, la1, st1, me1, of×2 → 13 room instances), pruning adjacencies to the retained codes, and setting single-storey constraints (storey_minimum: 1, storey_limit: 1). The straddling anonymous spaces n/t (no explicit level key) were dropped so the set is an unambiguous single floor. Seeded from the bare plot (init.dom).

• Expectation / decision rule: near-zero fails ⇒ bottleneck is multi-storey coupling (staging is the lever); still stalls (esp. missing) ⇒ per-floor construction itself is the bottleneck (§11.2 required first).

• Command (reproduce):

  URB_NO_OCCLUSION=1 python3 experiments/run_search_scaled.py \
    examples/harbor-house-l0 20000 0 \
    examples/harbor-house-l0/init.dom examples/harbor-house-l0/generated.dom
  

• Result: 20000 native evals across 250 topologies (234 s, 85 evals/s). Best 33 fails, fitness 2.25e-12 — deep in the 0.5ⁿ high-fail penalty regime, with the whole 16-member population stuck at 33–35 fails. The smaller budget-300 smoke run sat at 40 fails; full budget only crept 40 → 33. Not near zero. Fail histogram of the best generated.dom:

  count	category
  13	missing (all 3 m meeting rooms never constructed: required/critical + per-instance size/width/adjacency sub-checks)
  6	adjacency (ws1→c, k1→da1, da1→c, da1→k1, me1→c, la1→c)
  4	access
  4	size
  2	edge too long
  2	crinkliness
  1	proportion
  1	too few stairs — single-storey artifact (staircase_min floored to 1 by the fitness or 1 default; constant across runs)
  33	total

• Verdict: per-floor CONSTRUCTION is the bottleneck, not multi-storey coupling. Even on a single floor with only 13 rooms and zero delta/core-alignment complexity, the search cannot assemble the required room set: the dominant category (13/33 = 39 %) is missing — the counted anonymous space m×3 is entirely absent — and the remaining fails are downstream adjacency/access/size consequences of a room set the mutation operators never managed to construct. This matches the §11.0 prediction's "still stalls (esp. missing)" branch: §11.2 programme-aware construction + missing-room repair is the prerequisite, and staging alone (§11.3) will not rescue it. §11.3 stays blocked on §11.2.

11.2 Programme-aware construction + missing-room repair (homemaker-py-c4c.2) — DONE

Two changes (operators.py, wired in driver.py):

1. constructive_topology — bootstrap seeder that makes the required room set a constructive invariant. It sizes each storey to its required rooms (partitioning by level; level-free rooms distributed round-robin over a shuffled order), plus one circulation C and one outside O per storey, grows the slicing tree to that leaf count, and assigns the types. Stochastic (random splits/rotations, shuffled type→leaf assignment) so a bootstrap batch is still a diverse population. Replaces the random random_topology bootstrap whenever the programme has required spaces.
2. mutate_place_missing — repair operator. Detects a required-but-absent space (graph.check_space_counts) and inserts one by dividing a host leaf into [room | remainder]. Lex-safe host ranking (cf. §4.10): generic O leaves first (unbounded, nothing displaced), then other non-required leaves, circulation/stairs only as last resort; a required room is never displaced. Forced onto the room's required storey when the programme constrains its level. Weight 2.0 in the mutation mix (noops cheaply once complete).

• Gate: missing-type failures collapse to ~0; net-fail improvement vs the blank-slate baseline; no regression on the seeded programme-house 1-fail optimum (§4.10).

• Commands (reproduce):

  # A/B at identical budget+seed (old = git HEAD before this change):
  URB_NO_OCCLUSION=1 python3 experiments/run_search_scaled.py \
    examples/harbor-house 20000 0 examples/harbor-house/init.dom out.dom
  # §4.10 regression: warmstart-2f4 seed, 50000 evals, pop 8, 4 workers
  

• Result (harbor-house, 20000 native evals, seed 0, identical config):

  metric	OLD (random bootstrap)	NEW (constructive)
  seed best fails	163	139
  final total fails	133	105
  missing fails	103 (77 %)	12 (11 %)
  missing-records	22	2
  dominant remaining	missing	crinkliness 27, size 23, access 13, edge 12

  Constructive seeding alone gives a 24-fail head start at the seed (163 → 139) and the run ends at 105 vs 133 (−21 %), with the missing stack collapsed 103 → 12. §4.10 regression: PASS — the warmstart-2f4 seed still reaches a 1-fail population (whole pop 1f at 50 040 evals; place_missing noops harmlessly when the set is complete).

• Verdict: construction works and is necessary, but reframes the bottleneck. Making the required set a constructive invariant removes the catastrophic missing-room stacking that dominated the blank-slate baseline (77 % → 11 % of fails). But a complete 36-room harbor design then carries a large quality-fail load — crinkliness/size/access/edge-too-long packing of two fully-populated floors — that the current geometry inner loop + topology operators reduce only partway in 20k evals. So total fails improve but stay high. The dominant categories are now exactly what §11.4 (graded objective, to navigate the dense quality-fail regime) and §11.3 (staging — build one credible floor at a time instead of cramming both) target; §11.3 is unblocked by this result. A concrete next seeder refinement (filed): the type→leaf assignment is currently random, ignoring adjacency — clustering each room near its required c/neighbour at construction time should cut the adjacency (8) and downstream access (13) fails directly.

  Note on the baseline: DESIGN cited a "74-fail out1.dom", but the on-disk out1.dom is untracked and was overwritten by a prior experiment (it now re-scores to 37 fails; the committed out1.dom.fails of 74 lines belongs to the superseded .dom). The honest, reproducible comparison is therefore the identical-config A/B against the pre-change code (133 fails), not the stale out1.dom number.

11.3 Staged per-floor search (homemaker-py-c4c.3) — DONE

Searches the genome in causal dependency order (driver.search_staged), two stages composed from the existing driver.search:

1. Stage 1 — base floor (40 % of budget). A single-storey programme is auto-derived to a tempdir (programme.write_stage1_programme): the full patterns.config filtered to the storey-0 room set (programme.partition_rooms_by_storey), level: keys dropped, adjacencies pruned to surviving refs, storey_limit/staircase forced to 1. The base is searched on that reduced programme but ranked with a substrate-readiness bonus — key (-n_fails, fitness·(1 + W·readiness)), W=1 — so it is selected as a good substrate, not merely a good ground floor (anti-§4.2). graph.substrate_readiness = core_factor · capacity: full credit for a reserved C leaf ≥ STAIR_MIN_AREA (vertically-alignable core), times min(1, usable_base_area / required_upper_area) (enough divisible footprint for the upper set).
2. Stage 2 — upper floors as deltas (remaining budget). The best base is lifted (operators.lift_base_to_storeys) into a full multi-storey design that preserves the base storey and its inherited core and instantiates each upper storey's required room set by construction (the Stage-2 analog of §11.2 seeding). Deltas are searched with the base kept mutable at low probability (base_p=0.15, threaded through the exploratory ops; place_missing/core_* stay unbiased — repair and core-maintenance must reach the base).

• Gate: staged beats single-stage on harbor at equal budget; reserved-core + readiness prevent the bungalow trap (stage 2 does not carve a core from scratch); no programme-house regression.

• Commands (reproduce, URB_NO_OCCLUSION=1, 20000 evals, seed 0):

  python3 experiments/run_search_scaled.py examples/harbor-house 20000 0 \
    examples/harbor-house/init.dom scratch/ab_single.dom        # single-stage
  python3 experiments/run_staged_search.py examples/harbor-house 20000 0 \
    examples/harbor-house/init.dom scratch/ab_staged.dom         # staged
  

• Result (harbor-house, 20000 native evals, seed 0, identical config):

  metric	single-stage	staged
  total fails	105	95
  crinkliness	27	18
  edge too long	12	8
  proportion	6	4
  width	4	2
  size	25	26
  access	13	18
  missing	8	8
  adjacency	2	2

  Single-stage reproduces the §11.2 baseline exactly (105 fails); staged ends at 95 (−10, −9.5 %). The gain is concentrated in the packing fails staging targets — crinkliness 27→18 and edge-too-long 12→8 — at a small cost in access (+5). Anti-bungalow: confirmed. Every core_divide/core_undivide in the Stage-2 winning lineage is a noop — the core is inherited from Stage 1 and is never carved from scratch. Programme-house regression: PASS — single-storey programmes fall through to plain search; the warmstart-2f4 seed (50000 evals, pop 8, 4 workers) still reaches a whole-population 1-fail optimum (§4.10).

• Verdict: staging helps, modestly, and is the right structural frame. Building one credible, substrate-ready floor first — then upper floors as constructed deltas with an inherited core — beats cramming both floors simultaneously (95 vs 105) without touching the inner loop. The remaining load is the dense quality-fail regime (size/access/crinkliness on two fully-populated floors) that §11.4 (graded objective) targets: with missing already collapsed (§11.2) and the floors now assembled in dependency order, the lever left is navigation within the high-fail plateau, where lex-by-count gives near-zero gradient.

11.4 Graded high-fail objective (homemaker-py-c4c.4) — DONE (negative)

Premise (from Phase 4, §4.9): lexicographic-by-total-count (-n_fails, fitness) gives ~zero selection signal in the high-fail regime because the 0.5^n cliff flattens fitness to ~machine-epsilon, so neighbours at ~49–105 fails look indistinguishable. Proposed fix: a continuous proximity key beneath fail-count and above fitness — (-n_fails, grade, fitness).

Implementation (kept, default-off). fitness._leaf_grade reads each failing per-leaf quality factor (perpendicular/proportion/size/width/crinkliness/access) as proximity-to-satisfaction f / FAIL_THRESHOLD ∈ [0,1) and sums it; Fitness.score_with_grade returns it alongside score/fails. The scalar fitness and the fail count are untouched, so the inner-loop 0.5^n cliff (§5.4) is unaffected — inner-loop 0/9-regression check: PASS (re-ran §4.9 part 1, run_inner_loop_protection, 0/9 regressions). The grade is read once per child off the already-optimised tree in driver._evaluate (one extra native eval, ~1/child_budget) and used only in the outer comparator key, behind search(..., use_grade=True) / search_staged(..., use_grade=True) (default False; threaded to Stage 2 only — Stage 1 keeps its readiness key, §11.3). Structural fails (missing/adjacency/edge-too-long/level/…) score 0 grade, so the missing-space hierarchy (§6) is preserved: grade can never reward dropping a room.

• Commands (reproduce, URB_NO_OCCLUSION=1, 20000 evals):

  USE_GRADE=0 python3 experiments/run_staged_search.py examples/harbor-house 20000 <seed> \
    examples/harbor-house/init.dom scratch/st_lex.dom        # lex baseline
  USE_GRADE=1 python3 experiments/run_staged_search.py examples/harbor-house 20000 <seed> \
    examples/harbor-house/init.dom scratch/st_grade.dom      # lex + grade
  

• Result (harbor-house, staged, 20000 native evals, total fails at budget):

  seed	staged lex	staged lex+grade
  0	95	99
  1	96	98
  2	106	102
  mean	99.0	99.7

  Grade wins 1/3 seeds, loses 2/3, and is slightly worse on the mean — within seed-noise, no escape from the plateau. Single-stage seed 0 is a dead heat (105 = 105). Stage-1 is identical by construction (grade off there); the divergence is entirely in Stage 2, where the grade run stalls early (seed 0: last improvement at 13600/20000 evals, stuck at 99) while lex keeps reducing the count (99→95).

• Why it fails — the premise is falsified by measurement. The cliff is constant within a fail-tier (0.5^n, n fixed), so within a tier reported fitness is value/cost × const and still spans ~6 orders of magnitude (seed-0 Stage-2 history: 1.2e-37 → 4.6e-31 all inside the same descending fail count). The outer comparator only ever compares within a tier (−n_fails dominates across tiers), so lex's secondary fitness key already carries a strong, well-graded signal — exactly the gradient §11.4 assumed was missing. Inserting grade above fitness displaces that working signal: the population fills with high-grade (shallow-fail) incumbents and the fail-reducing restructurings — which transiently deepen other fails and so look worse on grade — are no longer selected. Placing grade below fitness instead would be near-inert (fitness ties are measure-zero in a continuous objective). Either way there is no lever: the high-fail plateau is a topology basin, not a comparator-resolution problem.

• Verdict: reject the graded objective; lexicographic (-n_fails, fitness) stands. The §11.3 staged 95-fail result remains the harbor best. The remaining load is genuinely structural (escaping topology basins), which is what §11.5 (structural niching + restarts) and the 9gp canonical-encoding capstone target — not outer-comparator reshaping. The use_grade flag and score_with_grade are kept default-off for reproducibility and possible reuse (e.g. as a diversity signal under §11.5 rather than a selection key).

11.5 Topology diversity: structural niching + restarts (homemaker-py-c4c.5) — DONE (negative)

Premise (epic diagnosis): the population dedups on the fitness scalar (driver.admit, abs(fitness) within 1e-9) and so has no structural diversity preservation — proposed as the root cause of the blank-slate gap (§7 Phase 2: memetic 18 fails vs urb-evolve 6), a single mutation chain losing to urb-evolve's upfront random-population diversity.

Implementation (kept, default-off). A cheap structural topology signature (genome.signature) string-encodes each storey's tree shape + cut orientations

• leaf types, routed through encode so dead inherited fields canonicalise; it is ratio-invariant (same topology, different geometry → same signature). Two diversity mechanisms, both behind flags on search/search_staged: niche_by_signature holds at most one individual per signature in the population (structural niching, keeping the better of a collision) in place of the fitness-scalar guard; restart_patience=<evals> does a soft restart on stagnation (keep restart_elite incumbents, refill with fresh constructive/random seeds — urb-evolve's upfront diversity as a soft restart). SearchResult gained n_distinct_signatures / diversity_history / n_restarts to quantify diversity over time.

• Commands (reproduce, URB_NO_OCCLUSION=1, 20000 evals):

  NICHE=0 python3 experiments/run_search_scaled.py examples/programme-house 20000 <seed> \
    examples/programme-house/init.dom scratch/ph_before.dom   # legacy dedup (before)
  NICHE=1 python3 experiments/run_search_scaled.py examples/programme-house 20000 <seed> \
    examples/programme-house/init.dom scratch/ph_niche.dom    # structural niching
  NICHE=1 RESTART_PATIENCE=2000 python3 experiments/run_search_scaled.py \
    examples/programme-house 20000 <seed> examples/programme-house/init.dom scratch/ph_restart.dom
  # harbor (staged): swap run_staged_search.py, seed examples/harbor-house/init.dom
  

• Diversity (the secondary criterion) — MET. Niching takes the final population from ~4–6 / 16 distinct topologies (legacy dedup) to 16 / 16; restarts raise distinct topologies seen by ~30 % (≈105–138 → ≈164–186 on programme-house). The signature machinery works exactly as designed.

• Fail count (the gate) — NOT MET. Blank-slate programme-house, total fails at budget (lower is better):

  seed	before (legacy)	niche	niche + restart
  0	11	14	12
  1	11	11	14
  2	15	13	13
  mean	12.3	12.7	13.0

  Harbor-house (staged, seed 0): legacy 95 (reproduces §11.3 exactly), niche 94, niche+restart 108. Across both programmes niching is a tie within seed noise and restarts are strictly worse; nothing approaches the ≤ 6 gate.

• Why it fails — the premise is falsified by measurement. More structural population diversity does not buy lower fails: the legacy dedup already holds 14/16 distinct topologies on harbor (Stage-2 starts from lifted bootstraps), so it was never the diversity bottleneck the epic assumed. Maximal diversity (16/16) with the fixed tournament pressure just diffuses effort — the fitness-scalar dedup's smaller effective population exploits a basin slightly harder. Restarts throw away converging Stage-2 work and regress hardest. The high-fail plateau is a reachability problem (operators + encoding cannot reach the low-fail basins), not a population-management one — the same conclusion §11.4 reached from the comparator side.

• Verdict: reject niching/restarts as defaults; the legacy fitness-scalar dedup stands. niche_by_signature / restart_patience are kept default-off for reproducibility and reuse, and genome.signature is the cheap stand-in that the canonical Polish encoding (homemaker-py-9gp) supersedes. With §11.3–§11.5 all landed, the residual load is genuinely structural: the principled lever is the canonical encoding (associativity collapse (a|b)|c == a|(b|c)) plus richer topology operators, not outer-loop selection/population reshaping.

11.6 Adjacency-aware constructive seeding (homemaker-py-s44) — DONE (positive)

Premise (follow-up to §11.2): constructive_topology instantiated every required room but typed the leaves at random, so rooms landed stranded from circulation. On harbor the seed carried ~29 adjacency-to-c + ~27 per-leaf access + level-inaccessible fails (≈ 56 of the seeder-controllable load; the remaining size/width/proportion/crinkliness fails are geometry, the inner loop's job). The programme confirms the shape: of 16 harbor spaces all 16 require adjacency to c, so the dominant lever is connect every room to circulation.

Implementation (operators._assign_adjacency_aware, default-on). A single circulation leaf cannot border a dozen rooms, and a slicing tree guarantees adjacency only between siblings — so adjacency must be read from the geometric leaf graph, not the tree. The seeder now spends ~one extra leaf per three rooms on circulation, builds the type-independent geometry.leaf_graph, and picks a greedy connected dominating set of circulation leaves (start at the highest-degree leaf, extend along the frontier by most-newly-dominated): every room leaf ends up bordering a connected circulation spine, so adjacency-to-c and access are satisfied by construction at the seed geometry. Rooms are placed on dominated leaves (constraint-hardest first), outside O on the most peripheral leaf; room order and tie-breaks stay stochastic so a bootstrap batch is diverse. Threaded through driver.search(seed_adjacency_aware=True); adjacency_aware flag on constructive_topology (env ADJ in run_search_scaled.py) for the A/B.

• Commands (reproduce, URB_NO_OCCLUSION=1, 20000 evals, single-stage):

  ADJ=0 python3 experiments/run_search_scaled.py examples/harbor-house 20000 <seed> \
    examples/harbor-house/init.dom scratch/hh_adj0.dom        # random assignment (before)
  ADJ=1 python3 experiments/run_search_scaled.py examples/harbor-house 20000 <seed> \
    examples/harbor-house/init.dom scratch/hh_adj1.dom        # adjacency-aware (after)
  

• Seed quality (harbor, 10 seeds, raw seed before optimisation): adjacency-to-c 29.2 → 12.2, per-leaf access 26.6 → 8.3, level-inaccessible 0.4 → 0.2 (≈ 56 → 21 seeder-controllable fails). Geometry fails rise at the raw 0.5-split seed (more, smaller leaves) but are recovered by the inner loop.

• End-to-end (total fails at budget, single-stage, lower is better):

  seed	harbor before	harbor after	prog-house before	prog-house after
  0	105	100	11	10
  1	115	85	11	8
  2	110	87	15	10
  mean	110.0	90.7	12.3	9.3

  Harbor −19.3 fails (−17.5 %), programme-house −3.0 (−24 %). ADJ=0 seed 0 reproduces the §11.2 single-stage 105 baseline exactly (clean control). Notably the adjacency-aware single-stage harbor (mean 90.7, best 85) now beats the §11.3 staged best of 95 — the first Phase-6 fail-count reduction from seeding rather than search machinery.

• Verdict: keep adjacency-aware seeding as the default. It is the first lever in Phase 6 to move the fail count on both programmes. The win is the dominant adjacency-to-c / access load; secondary adjacencies and the staged lift_base_to_storeys upper floors are picked up in §11.7 (homemaker-py-ld5).

11.7 Adjacency-aware lift + secondary adjacencies (homemaker-py-ld5) — DONE (positive)

Two gaps left by §11.6: (a) lift_base_to_storeys — the staged Stage-2 seeder — still typed upper-floor leaves at random, so staged search did not get the adjacency win; (b) secondary adjacencies (k1↔da1, da1↔o, ~4 harbor rooms) were ignored.

Implementation. _assign_adjacency_aware gained a fixed_circ parameter: the dominating-set search is seeded from given circulation leaves, so on an upper floor the spine grows off the inherited vertical core rather than from scratch (preserving the §11.3 anti-bungalow core-alignment invariant). Room placement is now constraint-ordered: codes with the most non-c adjacency requirements are placed first, each onto the open slot that satisfies the most of its requirements against already-typed neighbours (circulation + rooms placed so far), clustering k1↔da1, da1↔o, etc. lift_base_to_storeys(reqs=…, adjacency_aware=True) grows a per-floor circulation budget and calls it with the core as fixed_circ; threaded through search_staged(seed_adjacency_aware=True) (ADJ env in run_staged_search.py).

• Seed quality (harbor lift, 8 seeds, raw seed): adjacency-to-c 16.1 → 7.6, access 16.2 → 7.2 on the lifted upper floor.

• End-to-end (harbor, staged, 20000 evals, total fails at budget):

  seed	staged before (ADJ=0)	staged after (ADJ=1)
  0	95	97
  1	96	78
  2	106	81
  mean	99.0	85.3

  ADJ=0 reproduces the §11.4 staged lex baseline exactly (95/96/106, mean 99.0 — clean control). Staged adjacency-aware is −13.7 fails (−14 %) and is now the best harbor configuration overall: staged baseline 99.0 → single- stage adjacency-aware (§11.6) 90.7 → staged + adjacency-aware lift 85.3 (best 78, seed 1). Staging and adjacency-aware seeding compose: the credible Stage-1 base and the core-seeded upper spine each contribute.

• Verdict: keep adjacency-aware lift + secondary clustering as defaults. Harbor is now ~85 fails, down from the 95/105 plateaus that opened Phase 6. The residual is geometry- and shape-bound (size/proportion/crinkliness on the denser, more-circulation layouts), which is the canonical-encoding / shape-feasibility territory of homemaker-py-9gp.

11.8 Topology diversity × selection pressure, co-tuned (homemaker-py-6zy) — DONE (negative)

Premise (loose end from §11.5): structural niching was A/B'd against the legacy fitness-scalar dedup with selection pressure held fixed at a binary tournament (driver._tournament, k=2). §11.5's own mechanism note named the coupling as the reason for its null — "Maximal diversity (16/16) with the fixed tournament pressure just diffuses effort" — i.e. diversity and pressure are coupled but were varied as if independent: niching widens the population, but k was never sharpened to convert the extra exploration back into exploitation. This issue isolates that coupling — sweep tournament size jointly with niching to test whether sharper selection turns the 16/16 structural diversity into lower fails. The project had already pivoted to the canonical encoding (homemaker-py-9gp); this is a falsification check so the lever is not silently lost, not an expected win (§11.4/§11.5 both located the plateau in reachability).

Implementation (knob only; default-off behaviour unchanged). Exposed tournament_k: int = 2 on search / search_staged, threaded into both _tournament call sites (crossover pair + mutation parent) and all three internal search() calls of the staged path; reuses the §11.5 genome.signature / niche_by_signature machinery unchanged. The experiments harness reads HOMEMAKER_TOURNAMENT_K (mirrors NICHE) in run_search_scaled.py / run_staged_search.py; experiments/run_6zy_ab.sh runs the joint grid (RESUME-able).

• Commands (reproduce, URB_NO_OCCLUSION=1, 20000 evals; blank-slate seed init.dom to match §11.5):

  # grid: NICHE ∈ {0,1} × HOMEMAKER_TOURNAMENT_K ∈ {2,3,4}
  NICHE=0 HOMEMAKER_TOURNAMENT_K=2 python3 experiments/run_search_scaled.py \
    examples/programme-house 20000 <seed> examples/programme-house/init.dom scratch/out.dom
  # harbor (staged): run_staged_search.py, seed examples/harbor-house/init.dom
  bash experiments/run_6zy_ab.sh        # full grid → scratch/6zy/summary.tsv
  

• Diversity (mechanism check) — confirmed biting. niche=on holds the final population at 16/16 distinct topologies at every k; niche=off sits at 4–11/16. The pressure knob is genuinely varied (k=2,3,4). So both levers are live — the null below is not a machinery artefact.

• Fail count (the gate) — no cell beats the baseline. Blank-slate programme-house, total fails at budget over 5 seeds (0–4), mean (sd):

  niche \ k	k=2	k=3	k=4
  off	4.80 (1.60)	6.40 (2.50)	6.00 (2.00)
  on	6.20 (1.72)	7.00 (1.41)	6.60 (1.85)

  The legacy (off, k=2) cell is the best of the six (4.80); every higher-pressure row and every niche=on row is equal-or-worse (6.0–7.0). All differences sit within ~1 sd at 5 seeds, so the grid is a wash — but the central tendency is unambiguous: sharpening k and adding niching both slightly hurt, the opposite of the rescue the premise hypothesised. Harbor-house (staged, seed 0) reinforces it: niche=on is uniformly worse than off at every k (k2 72→83, k3 77→82, k4 67→75); within the niche=on row higher k helps monotonically (83→82→75) but never catches the niche=off row, and the best cell overall (off, k=4 = 67) is a single-seed wiggle within noise of the off, k=2 = 72 baseline.

• Why it fails — the coupling is real but points the wrong way. Sharper selection does not convert the extra structural diversity into lower fails; if anything the 16/16 niched population at high k over-commits the larger spread to a handful of basins and loses the occasional lucky low-fail draw the smaller fitness-scalar population stumbles into. §11.5's "diffuses effort" diagnosis survives co-tuning: the bottleneck is reachability (operators + encoding cannot reach the low-fail basins), so reshaping selection/population pressure cannot recover what the search space does not expose — the same conclusion §11.4 reached from the comparator side and §11.5 from the diversity side.

• Verdict: §11.5 null is robust to selection pressure — reject k>2 and niching as defaults; binary tournament + fitness-scalar dedup stand. tournament_k is kept (default-2) as a reusable knob alongside niche_by_signature. With §11.4/§11.5/§11.8 all negative on the outer loop, the residual is confirmed structural: the principled lever is the canonical encoding + richer topology operators (homemaker-py-9gp), not selection or population management.

12. Phase 7 — scaling validation & residual reduction (post-c4c)

Epic: homemaker-py-leu. Status: opened 2026-06-19. Continuation of the closed Phase 6 (§11). Phase 6 evidence located the leverage in construction / seed quality (§11.6/§11.7 wins) rather than search machinery (§11.4/§11.5 both regressed); the harbor residual is now geometry/shape-bound at ~85 fails. This section is the experiment ledger for Phase 7, same discipline as §11: each subsection records the command, the numbers, and a one-line verdict.

12.1 Larger-than-house benchmark: maple-court (homemaker-py-leu.1) — DONE

Why. Harbor (16 programme entries, 2 storeys) was the biggest real programme in examples/. homemaker-py-9gp's headline claim is scaling >16 rooms and its acceptance criterion demands "a larger-than-house programme" to measure on — so a bigger benchmark is a prerequisite, not optional. Proportion-aware seeding (leu.2) and re-scoped 9gp are both measured against this baseline.

The benchmark. examples/maple-court/ — a three-storey assisted-living / co-housing facility: 26 distinct programme entries / 52 room instances across 3 required storeys (storey_minimum: 3), ~1015 m² target internal area on a ~790 m²/floor plot. It mirrors harbor's structure deliberately — a dominant adjacency-to-c load on nearly every room plus a handful of secondary adjacencies (da1↔k1, da1↔o, lr1/ws1/lo1/gh1/gy1 ↔ o), anonymous interchangeable room families (m×3, t×6, n×4, r×12, em×2, py×2, tt×4), and staircase_min/max: 2. Code letters avoid the generic c/o/s leading-letter trap (those are reserved in fitness.py/graph.py for circulation/outside/sahn): no room code starts with c/o/s, so harbor's quirk of typing Common Room / Storage / Office as quasi-generic (cr1/st1/of) is not reproduced. init.dom is a single O footprint; storeys are built by the search from storey_minimum, exactly as harbor.

Baseline (current default search: adjacency-aware seeding + staged, §11.7). Reproduce (URB_NO_OCCLUSION=1, 20000 evals, staged, ADJ=1 default):

URB_NO_OCCLUSION=1 python3 experiments/run_staged_search.py \
  examples/maple-court 20000 <seed> examples/maple-court/init.dom scratch/mc_s<seed>.dom

seed	total fails	best lineage
0	145	rotate 0/rrlr
1	158	core_undivide noop
2	152	swap 0/rrlllr
mean	151.7

Each run executed exactly 20000 native evals across 250 topologies (~36 min, ~9.1 evals/s) and re-scored native-consistent (→ OK). The best layout (seed 0, 145 fails) was saved as examples/maple-court/generated.dom with its .fails (superseded in §12.2 by the proportion-aware 126-fail layout). The single-stage harness (run_search_scaled.py) also accepts the programme unchanged. The score prints near-zero (0.5^145 fail cliff) — the fail count is the yardstick.

• Verdict: benchmark established at mean 151.7 fails (best 145). As expected for a programme ~3× harbor's room count, the absolute fail floor is well above harbor's ~85; this is the scaling yardstick leu.2 (proportion-aware seeding) and the re-scoped 9gp are measured against. The residual character is the same geometry/shape family flagged at the close of §11.7.

12.2 Proportion-aware constructive seeding (homemaker-py-leu.2) — DONE (positive)

Premise (follow-up to §11.6/§11.7). The constructive seeders grow geometry with uniform [0.5, 0.5] cuts before types are assigned, so the raw seed is "more, smaller leaves" of equal area: a room with a large programme target comes out too small, a small room too big, and the inner loop must recover all of size/width/proportion from scratch. With the adjacency load now cut by seeding (§11.6/§11.7), this geometry residual is the dominant remaining term. Attacking it at the seed — in the proven construction direction — is far cheaper than the 9gp encoding rewrite.

Implementation (operators._size_divisions_from_targets, flag seed_proportion_aware, env PROP, default-on per the A/B below). After the adjacency-aware type assignment (§11.6/§11.7, left exactly as is), each leaf carries a target area — a sized room's programme size; circulation/outside absorb the plot slack (floored at 0.4 × mean room area so a circulation leaf never shrinks below door-width and undoes the §11.6 adjacency win). Because division=[f, f] cuts off left area-fraction f (rotation-independent — verified), bottom-up subtree-target sums compose multiplicatively to give every leaf area ∝ its target. Area alone regressed the raw seed, though: choosing only the cut fraction to hit a target area slices thin slivers with terrible aspect (proportion/width/edge-too-long fails swamp the size gain — measured below). So each cut also picks the rotation (the two distinct cut directions) that makes its two children squarest; rotation depends on realised parent geometry, so the pass runs top-down. Both ratio and rotation derive from the target dims; neither touches topology or type assignment. Threaded through driver.search/search_staged(seed_proportion_aware=…).

• Raw-seed fails (10 seeds, single-stage constructive, before optimisation), area-only vs area+rotation:

  family	harbor before	area-only	area+rot
  geometry	123.0	135.9	99.9
  access/adj	19.1	23.8	20.4
  total	144.1	162.1	123.7

  Area-only makes geometry worse (slivers); area+rotation drops the geometry family on every programme — harbor 123.0 → 99.9 (−19 %), programme-house 13.1 → 8.7 (−34 %), maple-court 200.5 → 164.1 (−18 %). Access/adjacency regresses slightly (rotation shifts the leaf graph the adjacency assignment was computed against): harbor +1.3, prog-house +2.4, maple +3.4 — far smaller than the geometry gain. The size family in particular falls as intended (harbor size 31.4 → 22.0), and proportion flips from a regression to a win (21.3 → 12.8) once rotation is co-chosen.

• End-to-end (total fails at budget, 20000 evals, 3 seeds, PROP=0 vs PROP=1; harbor & maple-court staged):

  seed	harbor PROP=0	harbor PROP=1	maple PROP=0	maple PROP=1
  0	97	72	145	126
  1	78	81	158	148
  2	81	69	152	134
  mean	85.3	74.0	151.7	136.0

  Harbor −13 % (best 69, was 78), maple-court −10 % (best 126, was 145). PROP=0 reproduces the §11.7 staged harbor (85.3) and §12.1 maple baseline (151.7) exactly — clean controls. Proportion-aware seeding is the first Phase-7 lever to move the fail count on the larger-than-house benchmark.

• A storey-count bug surfaced (homemaker-py-cq1). programme-house has storey_minimum: 2 but all rooms level: 0, and n_storeys_required only read level: keys — so the constructive seeder built a 1-storey seed for a 2-storey programme and search_staged fell through to plain search. Fixed (programme.storey_minimum/n_storeys_for; driver.search passes min_storeys to the seeder; search_staged routes on max(level-derived, storey_minimum)). No-op for harbor/maple (level-derived already ≥ storey_minimum); independent win on programme-house (single-stage baseline 8.0 → 5.0 with a correct 2-storey seed).

• programme-house regresses, but it is a convergence-speed artifact, not a worse optimum. On the 6-room programme proportion-aware seeding loses at 20000 evals on every path tested (single-stage 1-storey 8.0→11.7, single-stage 2-storey 5.0→8.3, staged 2-storey 4.3→6.0). The mechanism is a deeper local optimum: the equal-area PROP=0 seed has badly-proportioned leaves, so undivide moves — the route to programme-house's simpler optimum — are accepted as improvements; the well-fitted PROP=1 seed makes undivide an immediate fitness drop (merging two good leaves yields one bad one), walling off the restructuring path. A budget sweep (staged, storey-fixed) shows this is reachability speed, not an asymptotic trap:

  budget	PROP=0 (s0/s1)	PROP=1 (s0/s1)
  20000	4 / 5	8 / 6
  60000	2 / 2	4 / 3
  150000	2 / 0	1 / 10

  PROP=1 reaches 1 fail (seed 0, 150k — beating PROP=0's 2; best-known is 2), so it is not trapped; the gap narrows with budget and crosses over. (Staged splits budget by fraction, so runs at different budgets evolve different Stage-1 bases and are not nested — hence the high variance, e.g. PROP=1 seed 1 swinging 3→10.) The same "deeper basin" that helps where the constructed topology is roughly right (large programmes, scarce budget) delays convergence where the seed must be restructured (small programmes).

• Verdict: keep proportion-aware split sizing, default-on (seed_proportion_aware default True, env PROP=1). It is a measured win on both larger programmes — harbor −13 %, the maple-court scaling benchmark −10 % — exactly the regime Phase 7 targets and the basis the re-scoped 9gp is measured on. The only regression is a small-programme convergence-speed effect that washes out with budget (PROP=1 reaches the known floor), with no evidence of an asymptotic penalty, so default-on is not paid for by a worse optimum anywhere. The win is rotation-and-ratio sizing from target dims; the bare ratio is not enough (area-only regressed). Area sizing assumes total target ≈ plot area; choosing the cut direction for aspect is what makes it pay.

12.3 Re-scoped 9gp: shape feasibility + reachability moves (homemaker-py-9gp)

Re-scoped capstone of the epic (2026-06-19): the original canonical-Polish- expression rewrite was justified partly by a niching signature, but §11.5 falsified niching and genome.signature already supplies the cheap stand-in. The two surviving, evidence-supported parts are landed here as operators on the existing decoded Node tree — no Polish-expression rewrite — each measured independently against the §12.2 leu.2 baseline (maple-court staged 136.0, harbor 74.0). A true canonical encoding is revisited only if the M3 measurement proves associativity valuable at scale.

9gp.1 — shape-feasibility pre-filter (scaling lever). operators. predicted_shape_fails(root, reqs, fit) lays a topology out at its proportion- aware target geometry (reusing _size_divisions_from_targets, §12.2 — the squarest layout the inner loop warm-starts from) and counts the size/width/proportion/crinkliness fails the native fitness reports: a cheap lower-bound proxy for the best shape the topology can reach. driver._evaluate calls it before the inner loop and prunes (1 feasibility eval instead of ~80 inner-loop evals) when the predicted shape fails both exceed a tunable threshold and are ≥ the incumbent's total fails — the second guard makes the proxy safe (a topology whose shape floor is still below the incumbent is never discarded). Pruned individuals are tagged pruned/…, counted as explored topologies but never bred from or ranked, so budget flows to feasible topologies. Seed/bootstrap/restart batches are never filtered (construction invariants must survive). Threaded as search(…, feasibility_filter, feasibility_max_shape_fails) through search_staged; default OFF so the §12.2 controls reproduce exactly (test_feasibility_filter_off_matches_baseline). Env: FEAS=1 MAXSHAPE=<n>.

9gp.2 — M3 Wong-Liu re-association move (reachability lever). operators. mutate_reassociate adds the associativity move (a|b)|c ↔ a|(b|c) on two same-orientation live cuts (both directions, for reversibility): a pure- topology move that preserves the leaf set and types but reaches tree shapes the existing set cannot. M1 (operand swap) is mutate_swap and M2 (single-cut orientation complement) is mutate_rotate; associativity was the missing canonical-slicing move attacking the reachability bottleneck §11.4/§11.5 both fingered. Only live cuts (below is None, as mutate_rotate) are restructured, so dead inherited fields are untouched and encode re-anchors deltas; the two restructured cuts default to 0.5 and the inner loop recovers their ratios. Registered in MUTATIONS; default OFF via enable_reassociate (forces its mutation weight to 0 so the baseline is byte-identical). Env: REASSOC=1.

• Implementation status (this session): both land with unit tests (tests/test_operators.py: reassociate preserves the leaf multiset, changes the signature, noops on perpendicular cuts, stays canonical on the harbor corpus; predicted_shape_fails is non-negative, pure, deterministic. tests/test_driver.py: filter-off reproduces the baseline trajectory; filter-on prunes at 1 eval/topology and never admits a pruned individual). Full suite green (211 passed). A short smoke run on maple-court confirms both paths execute under the real native fitness.

• Calibration (predicted shape-fail floor of the constructive seeds). Over 8 proportion-aware constructive seeds, predicted_shape_fails is maple 121–163 (mean 135.6) and harbor 72–90 (mean 84.6) — essentially equal to the final achieved total fail counts (maple 126–148, harbor 69–81). So the shape floor at the best achievable geometry already accounts for almost the whole residual: independent confirmation of §11.7 that the Phase-7 residual is geometry/shape- bound. MAXSHAPE was set below the incumbent range (maple 100, harbor 55) so the pred ≥ incumbent safety guard is the dominant prune gate (experiments/ run_9gp_ab.sh).

• A/B sweep (DONE — negative). maple-court + harbor, seeds 0/1/2, 20000 evals, staged, total fails at budget:

  programme	seed	baseline	reassoc	feas	combined
  maple-court	0	126	131	129	131
  maple-court	1	148	141	151	142
  maple-court	2	134	146	140	144
  maple-court	mean	136.0	139.3	140.0	139.0
  harbor	0	72	83	82	81
  harbor	1	81	81	80	81
  harbor	2	69	70	69	70
  harbor	mean	74.0	78.0	77.0	77.3

  The baseline controls reproduce the §12.2 leu.2 means exactly (maple 136.0, harbor 74.0) — a clean control, so the negative is real. Every variant is neutral-to-slightly-worse on every programme: reassoc +3.3/+4.0, feas +4.0/+3.0, combined +3.0/+3.3 (maple/harbor). The feasibility filter did prune and explore more topologies in several runs (maple s1/s2 combined 342/319, s2 feas 317 vs the baseline 250) — but the extra topologies did not lower the fail count, and M3 reassociate never produced a win despite reaching new tree shapes.

• Verdict: keep both default-OFF; the Phase-7 residual is NOT reachability- or feasibility-bound. This is the third independent negative on search machinery (§11.4 graded objective, §11.5 niching+restarts, now §12.3 M3 moves + shape pruning), against four positives all from construction/seed quality (§11.2, §11.6, §11.7, §12.2). The associativity move reaches new topologies but they are not better; the shape filter saves budget on topologies whose shape floor already matches the incumbent, but — precisely because the floor ≈ the achieved total (calibration above) — there is no lower-fail basin for that saved budget to find. The geometry/shape residual is intrinsic to the constructed layouts, not a search-reachability deficit. A full canonical Polish-expression rewrite is not justified: its one measurable promise here (associativity reachability) was tested directly and did not pay.

• Residual diagnostic (where the shape fails actually live, maple-court, 6 constructive seeds). A per-leaf breakdown — to test, not assume, what the next lever would be — overturns the obvious "shape-aware placement" guess:

  signal	measured	reading
  plot utilisation (target/plot area)	0.44 (0.28–0.54)	NOT density/area-bound — ample slack
  failing leaves / total	~68 / 73	shape fails are uniform, not concentrated
  dominant factors	crinkliness 346, size 242, proportion 121, width 102	perimeter/area + undersize, both granularity effects

  Because nearly every leaf fails (not a few mismatched ones), the residual is not a room→leaf placement mismatch — there are no well-shaped leaves to place demanding rooms into. The mechanism is over-granular construction: 73 small leaves for 52 rooms at 44 % utilisation gives every leaf a high perimeter/area ratio (crinkliness) and rooms below their target area (size). So the measured candidate lever is construction granularity / leaf shape (fewer, larger leaves; merge or share leaves across same-class rooms; a coarser spine), NOT shape-aware placement and NOT more search machinery. This is a hypothesis with a measured motivation, filed as homemaker-py-c3g — it is unproven and must be A/B'd against the §12.2 baseline before adoption, same discipline as every lever above. It may also be that 52 distinct rooms simply cannot be well-shaped as 52 leaves at this density, i.e. the residual is the geometry floor of the slicing representation; the experiment is what decides.

12.4 Construction granularity A/B (homemaker-py-c3g) — DONE (null) + a noise finding

The c3g hypothesis tested directly. The cheap raw-seed probe (circ-per-room divisor circ_divisor, env CIRCDIV, default 3) confirmed the mechanism but also its catch: a coarser spine lowers the shape floor (maple 135→110, harbor 83→66 as div 3→∞) yet raises access/adjacency by as much, leaving the raw total floor flat-to-worse (maple 198→210, harbor 121→134). div=3 already sits near the total-floor minimum. Because §12.3 showed shape is the hard residual and access/adjacency are cheap to repair, the open question was whether that trade pays end-to-end.

• End-to-end A/B (20000 evals, staged, total fails at budget; div=3 reuses §12.3):

  programme	div=3 (baseline)	div=6	div=8
  maple-court	136.0	137.0	134.3
  harbor	74.0	75.3	—

  Per-seed: maple div6 143/122/146, div8 132/138/133; harbor div6 65/76/85. Every arm is within ±1.7 of baseline — inside the noise floor (below) — with a huge per-seed spread (maple div6 122–146). Null result: coarsening the spine does not pay end-to-end. The raw-probe prediction held — the shape-floor gain is cancelled by access/adjacency damage that is not free to repair after all.

• A reproducibility finding surfaced en route (homemaker-py-xcy, P2 bug) — later RE-DIAGNOSED and FIXED (2026-06-22). The div=3 control gave 129 vs §12.3's 126 for the same maple seed 0. The first diagnosis blamed operators._assign_adjacency_aware iterating id()-ordered Python sets of Nodes — this was wrong. That function already ends every max/min with a unique leaf-idx tiebreak, and its set unions are used only for membership, so order never leaks: constructive_topology(seed=0) is byte-identical across processes for every example programme (stable sha1, e.g. maple e688f744326b). The "sig hashes 4480 vs 16064" was a measurement artifact — Python's builtin hash() of a string is salted per process (PYTHONHASHSEED), so an identical signature hashes to different ints run-to-run (reproduced 51920/5342/59970 for one identical string). Use genome.signature equality or a stable hash, never builtin hash(), to compare topologies. The real cause was parallel-only: driver._run_batch admitted futures via concurrent.futures.as_completed, i.e. in completion order, and admit() is order-sensitive (accrues n_evals per result; keeps the first individual of an equal-key tie as best). A long parallel run diverged 167 vs 161 fails (maple seed 0) — the true source of the ±3..6 "noise". Fix: iterate the futures in submission order (for f in futs: f.result(); all still run concurrently), reproducing the serial admission sequence. After the fix two workers=4 runs are byte-identical (162 fails). Serial (workers=1) was already byte-for-byte reproducible. Implication for the §11/§12 ledger: per-seed numbers are reproducible only at a fixed worker count. Serial≠parallel is expected (children/iteration = 1 vs n_workers changes batch granularity, hence the search), not nondeterminism. Any A/B that compared runs at different worker counts — or any pre-fix parallel run — conflated this with a real effect; sub-±3 effects (the §12.3 +3-4 negatives, the §12.4 ±1.7) should be re-run at a single fixed worker count before being trusted as magnitudes.

• Verdict: keep circ_divisor=3 default; the granularity lever is null. Together with §12.3 this closes the residual-reduction question for now from both sides: neither search machinery (§12.3) nor construction granularity (§12.4) moves the maple/harbor geometry residual beyond noise. The weight of evidence is that the residual is the geometry floor of the slicing representation at this room density — 52 distinct rooms as 52 adjacency-connected leaves inherently incur ~135 shape+access fails. Further progress, if wanted, needs either the determinism fix (to even see sub-±3 effects) or a representational change beyond the slicing tree — not another seed/search tweak at this scale.

13. Phase 8 — lowering the geometry/shape floor (homemaker-py-erc)

Phase 8 runs DIAGNOSTICS FIRST to decide which floor-lowering lever to invest in, then the construction/inner-loop experiments in dependency order. §12.3/§12.4 established the floor is real (search machinery and circulation-granularity both null); the open question is what about the floor — per-leaf slicing tax, or fixable cuts — and where the slack hides (util 0.44 yet rooms undersize).

13.1 Diagnostic A: per-leaf shape-fail vs density/granularity (homemaker-py-erc.1) — DONE

GATES leaf-sharing (erc.3) vs compactness-cuts (erc.5). Reads only; no A/B, no baseline reproduction. Builds the §12.2 constructive seed (adjacency- and proportion-aware), lays it out at the proportion-aware TARGET geometry — the squarest geometry the inner loop warm-starts from, exactly as operators.predicted_shape_fails — then counts size/width/proportion/crinkliness fails per leaf. Script: experiments/diag_leaf_shapefail.py (seeds 0/1/2).

View 1 — cross-programme density sweep (per-leaf rate = shape-fails ÷ leaves):

programme	rooms	leaves	l/room	util	shape	/leaf	siz/lf	wid/lf	prp/lf	crk/lf
programme-house	6	9.0	1.50	0.83	8.0	0.889	0.000	0.519	0.222	0.148
harbor-house-l0	13	13.0	1.00	0.31	19.0	1.462	0.231	0.154	0.487	0.590
harbor-house	37	45.0	1.22	0.50	87.3	1.941	0.519	0.378	0.296	0.748
maple-court	52	73.0	1.40	0.54	134.3	1.840	0.562	0.224	0.251	0.804

Per-leaf shape-fail SATURATES at ~1.8–1.9 once the programme is non-trivial: the tiny 6-room case is the only outlier (0.89, no size fails, high util 0.83), and the three larger programmes cluster at 1.46→1.94 with no dependence on leaves-per-room (which barely moves, 1.0–1.5). Cross-programme "density" here is confounded by plot/room-mix/util (util swings 0.31→0.83), so this view alone cannot separate "intrinsic per-leaf tax" from "more leaves, worse cuts".

View 2 — synthetic granularity sweep, maple-court, room set FIXED, leaf count varied via the c3g circ_divisor knob (the controlled test):

circ_div	leaves	l/room	util	shape	/leaf	siz/lf	wid/lf	prp/lf	crk/lf
2	81.0	1.56	0.46	139.0	1.716	0.477	0.169	0.226	0.844
3	73.0	1.40	0.54	134.3	1.840	0.562	0.224	0.251	0.804
4	68.0	1.31	0.44	126.7	1.863	0.495	0.294	0.289	0.784
6	65.0	1.25	0.47	126.0	1.938	0.554	0.303	0.262	0.821
9	63.0	1.21	0.50	116.3	1.847	0.481	0.280	0.339	0.746

With the programme held fixed, the per-leaf shape-fail rate is FLAT as leaf count varies (1.72–1.94, no monotone trend; if anything a slight rise as you coarsen, since the survivors are bigger but still fail). Crucially TOTAL shape fails track leaf count almost linearly (139 → 116 as leaves 81 → 63), and crinkliness — the dominant factor (crk/lf ≈ 0.75–0.84) — is itself flat per leaf. Each leaf carries a roughly fixed ~1.8 shape-fail tax regardless of how finely the same plot is sliced. The target layout already picks the squarest-aspect cut direction (_size_divisions_from_targets chooses rotation for squarest children), so leaves are already near-optimally shaped and STILL fail at ~1.8/leaf — there is little compactness headroom left to recover at fixed leaf count.

VERDICT — per-leaf shape-fail is FLAT vs slicing density (controlled view 2) → the floor is INTRINSIC to per-leaf slicing, not to cut quality. By the diagnostic's decision rule this prioritises leaf-sharing (erc.3 — fewer leaves for the same rooms is the only lever that moves the floor) and deprioritises compactness-aware cuts (erc.5 — cuts are already squarest and still pay the tax; little headroom at fixed count). Note this is not the §12.4 circ_divisor null: that lever removed CIRCULATION leaves and the shape gain was cancelled by access/adjacency damage; leaf-sharing removes ROOM-leaf count (multi-room leaves) without disturbing the circulation spine, so the access penalty that killed c3g need not apply. Recommendation: close/deprioritise erc.5, advance erc.3.

13.2 Diagnostic B: undersize-despite-slack localization (homemaker-py-erc.2) — DONE

GATES plot-fill construction (erc.4) vs the inner-loop slack-expansion term (erc.6). The §12.3 paradox: plot utilisation ≈ 0.44 (over half the plot "empty") yet rooms are UNDERSIZE. Where is the slack stranded, and at which stage should it be spent? Reads only. Builds the §12.2 constructive seed (whose geometry already sits at the proportion-aware TARGET ratios — the inner-loop warm start, so it is the "before" state), measures per sized-room leaf achieved-vs- target area and a plot accounting, then runs innerloop.optimise (nm, budget 80 = the bootstrap child budget) and re-measures. Script: experiments/diag_slack_localization.py (harbor-house + maple-court, seeds 0/1/2).

programme	state	sizeF	util	tgtFill	ā/t	%und	%ovr	sized%	circ%	out%
harbor-house	BEFORE (target)	23.3	0.50	0.50	1.43	43	12	50	46	4
harbor-house	AFTER (innerloop)	21.7	0.49	0.50	1.40	54	16	49	46	4
maple-court	BEFORE (target)	41.0	0.54	0.44	1.46	42	15	54	43	3
maple-court	AFTER (innerloop)	37.3	0.53	0.44	1.46	42	19	53	44	3

(util = sized-room area ÷ plot; tgtFill = Σ room targets ÷ plot; ā/t = mean achieved/target over sized leaves; %und/%ovr = leaves below 0.9× / above 1.1× target.)

The "56 % empty plot" is a misreading. Sized rooms already occupy ~50–54 % of the plot and hold 1.4–1.5× their aggregate target area (util > tgtFill); the other ~46 % of the plot is circulation, not claimable void (out/uncovered is only 3–4 %). So rooms are over-provisioned in total — there is no unused plot to hand them.

The size fails are pure MALDISTRIBUTION, set by SLICING POSITION not by need. The median room sits right at target (a/t ≈ 1.0), but a long undersize tail (p25 ≈ 0.35, min 0.05) starves while a few giant leaves balloon (max 6.8× harbor, 14.7× maple). Decisively, the same room type with the same target lands at both extremes — harbor r (target 10 m²) appears at 68 m² (6.8×) and 2.3 m² (0.23×); maple n (target 60 m²) appears near target and at 2.7 m² (0.05×). A leaf's area is dictated by its depth/position in the binary slicing tree (ratios multiply down the ancestry), essentially independent of its target; _size_divisions_from_targets sets each local cut proportionally but cannot defeat the multiplicative depth effect. This is the same root cause as §13.1 (the binary-slicing structure), now seen on the size axis.

The inner loop cannot repair it. Over budget 80 the size fails move only −1.6 (harbor) / −3.7 (maple), util is flat-to-down, and %undersize is flat-to- worse (43→54 harbor). On a frozen topology the equal-offset ratio DOF cannot shrink a 14× leaf to feed a starved one without trading into shape fails (the 0.5ⁿ cliff, §4.5, blocks it), and the symmetric size Gaussian (quality_size is gaussian(area, 1, target, σ)) gives no net reward for redistribution.

VERDICT — the slack is depth-driven maldistribution inside the room set, not unclaimed plot, and the inner loop (frozen-topology ratios) provably cannot move it. This falsifies plot-fill construction in the "claim the empty plot" sense (erc.4 as scoped — rooms are already 1.4× over aggregate target; the empty- looking plot is circulation) and deprioritises the inner-loop slack-expansion term (erc.6 — wrong DOF: ratios on a frozen tree cannot undo a depth-set 14× leaf, and the blocker is position not a missing expansion reward). The fix must live UPSTREAM of the inner loop, where leaf area is actually decided: construction that balances tree DEPTH so equal-target rooms land at comparable depth / caps giant leaves (re-scope erc.4 from "plot-fill" to depth-balanced / giant- splitting construction), reinforcing §13.1's call to advance leaf-sharing (erc.3) for the starved tail. Recommendation: re-scope erc.4, deprioritise erc.6.

13.3 Experiment: leaf-sharing / multi-room leaves (homemaker-py-erc.3) — DONE

The lever §13.1 named as the only one that moves the floor: collapse same-code rooms into fewer, larger shared leaves so the per-leaf ~1.8 shape tax is paid once per group instead of once per room. Unlike c3g (§12.4) this removes ROOM-leaf count, not circulation, so the access/adjacency penalty that sank c3g need not apply.

Mechanism — explicit, type-guarded per-leaf multiplicity. A construction stamps leaf.share = k and leaf.share_type = code on each shared leaf (operators._share_rooms groups a sized, multi-instance code into runs of ≤ N = leaf_share_factor; _leaf_mult_from_plan stamps the survivors and _size_divisions_from_targets sizes them to k × target). The fitness honours k only while leaf.type == leaf.share_type (graph.leaf_share), so any retype/undivide silently invalidates a stale share — the mutation operators need no resets, and a small leaf can never retype its way into claiming rooms it does not provide. Two scoring sites, both gated by a default-OFF leaf_sharing key (controls reproduce the §12.2 baseline exactly — 214 tests pass with it off):

• graph.check_space_counts counts coverage (Σ per-leaf k) against req.count, so one shared leaf satisfies several same-code rooms with no missing fail;
• fitness.quality_size centres the size Gaussian on k × target (σ scaled by k). quality_proportion/quality_width need no change — a proportionally-scaled leaf keeps its aspect and only gets wider.

Design history: the first cut recovered k from area (round(area/target)) to avoid genome state, but the §13.2 depth maldistribution left shared leaves below k × target, so round undercounted and 17–44 missing fails leaked back (harbor share3+il: 87.3 total, 16.7 missing; the inner loop could not close it — frozen-topology ratios, §13.2). Switching to explicit share (an undersize shared leaf is present → a light size fail, not a heavy missing fail) closes the leak. Because the phenotype tree is never rebuilt from the genome in the hot path (genome.decode is unused; operators edit dom.Node trees in place), the two Node fields survive the whole search via deepcopy without threading through GNode/encode/decode; .dom serialisation emits share only on a live shared leaf.

Floor probe (experiments/diag_leaf_sharing.py, harbor + maple, seeds 0/1/2) — build the §12.2 seed both ways, score at the seed geometry and again after innerloop.optimise (nm, budget 80) under the same objective. Averaged fails:

programme	mode	leaves	total	missing	size	crink
harbor	OFF +il	45.0	120.3	0.0	21.7	33.7
harbor	share2 +il	31.7	86.0	0.0	15.3	22.0
harbor	share3 +il	25.7	73.3	0.0	12.7	17.7
maple	OFF +il	73.0	194.7	0.0	37.3	58.3
maple	share2 +il	52.0	145.7	0.0	25.7	41.3
maple	share3 +il	47.0	133.0	0.0	21.0	39.3

The floor moves and the leak is closed — share3 cuts the achievable floor −39 % harbor (120.3 → 73.3) / −32 % maple (194.7 → 133.0) with zero missing fails, and the missing did not re-emerge as size fails (size still falls, 22→13 harbor / 37→21 maple). The drop is exactly where §13.1 predicted: shape factors fall with leaf count (harbor leaves 45→26, crinkliness 34→18). Larger leaf_share_factor helps monotonically here (share2 → share3), bounded by leaf_share_max (default 4).

Verdict — leaf-sharing is the floor-mover §13.1/§13.2 called for: −32…−39 % on the achievable floor, no missing-fail leak. The flag is threaded through the staged driver (driver.search/search_staged → constructive_topology / lift_base_to_storeys) and exposed for the A/B via LEAFSHARE/LEAFSHAREFAC in run_staged_search.py (which injects the objective into the inner-loop and final-score fitness, both arms on one programme dir). Smoke-tested end-to-end (harbor, staged, leaf_sharing+factor 3: re-score OK).

End-to-end A/B (experiments/run_leafshare_ab.sh, staged search, 20 000 native evals, seeds 0/1/2, leaf_share_factor=3 vs the default-OFF baseline, final native re-score):

programme	baseline (s0/1/2)	mean	leaf-share f3 (s0/1/2)	mean	Δ
maple-court	129 / 148 / 134	137.0	78 / 89 / 92	86.3	−37 %
harbor-house	72 / 81 / 69	74.0	50 / 52 / 49	50.3	−32 %

VERDICT — leaf-sharing is the first lever to move the Phase-8 floor, and it moves it decisively: −37 % maple / −32 % harbor end-to-end. The default-OFF baseline arm reproduces §12.2 exactly (maple 137.0 vs 136.0, harbor 74.0 vs 74.0), so the gap is the lever, not drift; and the separation is total — every share run beats every baseline run on the same programme (maple worst-share 92 < best-baseline 129; harbor 52 < 69). Fewer leaves also make each eval cheaper, so the share arm runs ~35 % faster at equal budget. This is the §13.1/§13.2 prediction realised: the per-leaf ~1.8 shape tax is intrinsic, so collapsing 52→47 / 45→26 room-leaves is what lowers the floor — and the explicit type-guarded multiplicity (vs the area-derived first cut) is what lets the gain survive without a missing-fail leak. Scoreboard update: this is the 5th win from construction/seed quality and the first floor-mover of Phase 8; it confirms §12.3's thesis that only lowering the geometry floor (not search machinery) can help. Follow-ups: surface leaf_sharing on the homemaker-evolve CLI / as a patterns.config key for production use, sweep leaf_share_factor/max_share, and test the erc.4 depth-balancing synergy (shared leaves at correct absolute area) now that the leak is closed.

13.4 Experiment: depth-balanced construction (homemaker-py-erc.4) — DONE (modest)

The lever Diagnostic B (§13.2) called for. B localized the size fails to depth-driven maldistribution: a leaf's area is the product of cut fractions down its ancestry in the binary slicing tree, so the same-target room lands at 0.05× and 14.7× by slicing position, and the inner loop (frozen topology) provably cannot move it. The fix must live in construction, where leaf area is decided.

Mechanism — depth-balanced tree growth. _grow_leaves grew the tree by splitting a random leaf each step → a random caterpillar whose leaves sit at wildly different depths. The depth_balanced flag instead always splits a shallowest current leaf (operators._leaves_with_depth), growing a near-complete binary tree so all leaves land at comparable depth. The proportion-aware sizing pass (_size_divisions_from_targets) then hits each target with cut fractions near their proportional value instead of compounding fmin/fmax clamp error down a deep spine. Type-agnostic and topology-only — it changes which leaf is split, not the type assignment or the proportional sizing — so it composes with adjacency-aware seeding and leaf-sharing unchanged. Default OFF (214 tests pass with it off); threaded through constructive_topology / lift_base_to_storeys → driver.search/search_staged, exposed via DEPTHBAL in run_staged_search.py.

Floor probe (experiments/diag_depth_balance.py, harbor + maple, seeds 0/1/2) — build the §12.2 seed OFF vs balanced (vs balanced+share3 as the erc.7 preview), score at the seed geometry and after innerloop.optimise (nm, budget 80). dDep = leaf-depth spread (max−min); maxR/minR = max/min achieved/target over sized leaves; %und = fraction below 0.9×target. Averaged:

programme	mode	leaves	total	size	crink	%und	maxR	minR	dDep
harbor	OFF +il	45.0	120.3	21.7	33.7	54.2	12.0	0.1	7.0
harbor	bal +il	45.0	106.0	21.0	31.3	25.0	8.3	0.2	1.0
harbor	bal+sh3 +il	25.7	65.3	11.7	17.3	29.0	4.1	0.3	1.0
maple	OFF +il	73.0	194.7	37.3	58.3	42.3	16.4	0.0	6.7
maple	bal +il	73.0	173.0	37.3	61.7	22.4	6.2	0.2	1.0
maple	bal+sh3 +il	47.0	113.7	22.3	38.7	17.7	7.9	0.4	2.0

The depth spread collapses (7→1) and the giant leaf is tamed — maxR 12.0→8.3 harbor / 16.4→6.2 maple, %undersize 54→25 / 42→22 — at equal leaf count (45 / 73, no rooms removed). The achievable floor drops −12 % harbor (120.3→106.0) / −11 % maple (194.7→173.0) purely from tree shape, with zero missing-fail leak. Most of the total drop is in width/proportion (the giants were the wide, wrong-aspect leaves), not the soft size Gaussian (size barely moves). Crucially it is additive with leaf-sharing: bal+sh3 beats §13.3's share3-alone floor (harbor 65.3 vs 73.3, maple 113.7 vs 133.0) — balancing places the survivors of sharing at correct absolute area, exactly the synergy erc.7 was filed for.

End-to-end A/B (experiments/run_depthbal_ab.sh, staged search, 20 000 native evals, seeds 0/1/2, DEPTHBAL=1 vs default-OFF baseline, leaf-sharing OFF in both arms, final native re-score):

programme	baseline (s0/1/2)	mean	depth-bal (s0/1/2)	mean	Δ
maple-court	129 / 148 / 134	137.0	142 / 126 / 119	129.0	−5.8 %
harbor-house	72 / 81 / 69	74.0	67 / 77 / 71	71.7	−3.2 %

VERDICT — depth-balancing is a real but MODEST standalone lever: −5.8 % maple / −3.2 % harbor, much smaller than the −11/−12 % the seed-floor probe predicted, and the arms OVERLAP (maple balanced worst 142 > baseline best 129; harbor balanced 77 > baseline 69) — not the total separation leaf-sharing showed (§13.3, every share run beat every baseline). The default-OFF baseline reproduces §12.2 exactly (maple 137.0 vs 136.0, harbor 74.0 vs 74.0), so the comparison is clean and the small gap is the lever, not drift. The 20k search erodes most of the seed-floor advantage: the random-caterpillar arm partly catches up via divide/undivide mutations over the budget, so an 11 % lower seed floor realises only ~5 % at convergence. This is the mirror image of the §12.3/§11 thesis — seed quality helps, but here the search recovers enough of the gap that depth-balance alone is marginal, unlike the structural leaf-count cut of §13.3 which the search cannot undo (you cannot mutate 26 leaves back up to 45 cheaply).

Its real promise is the additive floor with leaf-sharing: the probe showed bal+sh3 beats share3-alone by a wide margin (harbor 65.3 vs 73.3, maple 113.7 vs 133.0) because balancing places the survivors of sharing at correct absolute area. The decisive end-to-end test is therefore erc.7 (depth-balance × leaf-sharing synergy + factor sweep), not depth-balance in isolation. Recommendation: keep depth_balanced (default OFF, no test/runtime cost, same leaf count), advance erc.7 to test whether the additive seed floor survives to convergence when stacked on the share lever that the search cannot erode. Scoreboard: a 6th construction/seed lever, but the first Phase-8 lever whose end-to-end gain is materially smaller than its seed-floor gain — a useful calibration of how much seed-floor reduction the staged search actually banks.

13.5 Experiment: leaf-sharing × depth-balancing synergy (homemaker-py-erc.7) — DONE (synergy confirmed)

The decisive test the §13.4 floor probe set up. Depth-balancing was only MODEST standalone (§13.4: −5.8 % maple / −3.2 % harbor, overlapping arms) because the 20k search erodes a tree-shape seed advantage via divide/undivide. But the probe showed bal+sh3 beats share3-alone at equal leaf count (harbor 65.3 vs 73.3, maple 113.7 vs 133.0) — additive on the leaf-COUNT cut the search cannot erode (you cannot mutate 26 leaves back up to 45 cheaply). Question: does that additive seed-floor advantage survive to convergence once stacked on the share lever that the search can't undo?

Setup (experiments/run_synergy_ab.sh, staged search, 20 000 native evals, seeds 0/1/2, final native re-score). Both arms hold LEAFSHARE=1 at factor 3 (the §13.3 winner). The control arm is share-alone (DEPTHBAL=0) and must reproduce §13.3; the experiment arm adds DEPTHBAL=1 (depth-balanced grow). One programme dir per programme — run_staged_search.py injects leaf_sharing into the whole pipeline so both arms score under the same relaxed objective.

programme	share-alone db0 (s0/1/2)	mean	bal+share db1 (s0/1/2)	mean	Δ
maple-court	78 / 89 / 92	86.3	76 / 85 / 86	82.3	−4.6 %
harbor-house	51 / 52 / 49	50.7	41 / 41 / 38	40.0	−21.1 %

The control arm reproduces §13.3 exactly (maple 86.3 = 86.3, harbor 50.7 ≈ 50.3), so the comparison is clean and the gap is the lever, not drift.

VERDICT — the synergy is REAL and SURVIVES to convergence, unlike depth-balance alone. Harbor is decisive: −21 %, every seed improves by 10–11 fails, and the arms are non-overlapping (bal+share worst 41 < share-alone best 49) — the total separation §13.4-standalone never reached. Maple is modest but uniform: −4.6 %, every seed improves (−2 / −4 / −6), ranges overlapping. This is the mirror image of §13.4: there the seed-floor advantage washed out because the search could erode tree shape; here depth-balancing rides on top of the leaf-COUNT cut that the search cannot erode, so balancing the survivors of sharing onto their correct absolute k×target area banks. The probe prediction held — bal+sh3 beats share3-alone end-to-end, not just at the seed.

Factor sweep (experiments/run_sharefactor_sweep.sh, leaf_share_factor 2/4 under bal+share, seeds 0/1/2, vs the factor-3 bal+share above):

programme	factor 2	factor 3	factor 4
maple-court	92.7	82.3	83.3
harbor-house	53.0	40.0	39.7

Factor 3 confirmed as the robust default once depth-balancing is stacked. Factor 2 regresses on both (maple +10.4, harbor +13.0) — too little sharing leaves more, smaller rooms for the depth-balance to fix. Factor 3 and 4 are statistically tied (maple f3 wins by 1.0, harbor f4 wins by 0.3 — both inside seed noise, ranges overlap), so factor 4 buys nothing material and gives up maple while risking larger shared leaves. leaf_share_max (scoring cap, default 4) already credits every multiplicity at factor ≤4 with zero missing-fail leak (final re-score OK in all runs), so it needs no separate sweep at the chosen factor 3.

Recommendation: make depth_balanced + leaf_sharing (factor 3) the default Phase-8 stack (both default OFF today, no test/runtime cost). Scoreboard: the first Phase-8 lever combination whose end-to-end gain (harbor −21 %) exceeds either lever alone (share −32 %→ this stacks a further −21 % on top; depth-balance −3 % alone), confirming the §13.4 thesis that levers the search cannot erode compound where shape levers do not.

13.6 Experiment: interior-O courtyard / light-well seeding (homemaker-py-ld2) — DONE (positive on dense floors)

The construction lever aimed at the erc crinkliness residual directly. The adjacency-aware seeder placed ONE O on the most PERIPHERAL leaf — where the adjacent rooms already have plot facade, wasting the daylight source — while the landlocked rooms (no facade, no uncovered-O neighbour → area_outside ≈ 0 → crinkliness ≈ 0 → fail) get nothing. This arm instead seeds O as INTERIOR light wells (the most-landlocked leaves first, greedily spread so each illuminates a fresh room set) and scales their count with the room count.

Seed diagnostic first (the epic mandate). Decomposing every crinkliness fail in the bal+share seed by side of the gaussian: all are UNDER-exposed (crink < 0.62, landlocked) — zero over-exposed slivers (crink > 21.7). So the residual is genuine under-daylighting, validating the premise (and correcting the epic's loose "high perimeter/area" wording — the failing leaves are starved, not over-walled). The naive default outside_divisor=6 was null (too few/small wells; harbor seed 147→142, crinkliness even rose). Sweeping the divisor found odiv=3 seed-optimal: harbor seed fails 147→129 (−18), maple 219→206 (−14), landlocked fails down — at the cost of more leaves (harbor +4, maple +8). Because it ADDS leaves it carries the §13.4 wash-out risk, so the convergence A/B decides.

Setup (experiments/run_interioro_ab.sh, staged search, 20 000 native evals, seeds 0/1/2, final native re-score). Both arms hold the default stack LEAFSHARE=1 (factor 3) + DEPTHBAL=1. Control is interior-OFF (peripheral O) — must reproduce §13.5 bal+share; experiment adds INTERIORO=1 (odiv=3).

programme	peripheral off (s0/1/2)	mean	interior odiv=3 (s0/1/2)	mean	Δ
maple-court	77 / 85 / 86	82.7	74 / 78 / 89	80.3	−2.8 %
harbor-house	41 / 43 / 38	40.7	28 / 39 / 35	34.0	−16.4 %

The control reproduces §13.5 (maple 82.7 ≈ 82.3, harbor 40.7 ≈ 40.0), so the gap is the lever, not drift.

VERDICT — positive on the DENSE floor, marginal elsewhere. Harbor is the win the issue targeted (it named "harbor-house ~19 rooms/floor" as where the single peripheral O is wasted): −16.4 %, every seed improves (−13 / −4 / −3), arms nearly non-overlapping (interior worst 39 ≈ control best 38). Maple is −2.8 %, within seed noise — two seeds improve, one regresses (+3), ranges overlap. This is the §13.4 pattern: the seed advantage (harbor −18, maple −14) survives roughly a THIRD on harbor but mostly washes out on maple, because a dense floor has enough landlocked rooms that the daylight gain outweighs the added-leaf tax, whereas on the sparser maple the +8 leaves nearly cancel it. Unlike depth-balance-alone (§13.4) which washed out entirely, interior-O holds on the dense floor.

Recommendation: make interior_outside (odiv=3) a default-ON Phase-8 lever (default OFF today). Harbor is decisive and maple is net-neutral (mean still −2.8 %, no programme regresses on mean), so the flip is strictly ≥ on both means and matches the dense-programme target. Follow-up homemaker-py-* flips the default (mirroring pll after erc.7). outside_divisor left at 3 (seed-optimal joint); a finer odiv sweep under convergence is low-prior given maple's marginal response.

§13.7 High-budget harbor floor probe — 71d go/no-go (homemaker-py-71d.1)

The whole Phase-8 construction stack is now default-ON (leaf-sharing factor 3, depth-balanced, interior-O odiv=3, circ_divisor 3, proportion-aware). Cumulative floor vs the §12.2 leu.2 baseline (all under the §13.3 leaf-share-relaxed objective, staged, seeds 0/1/2): maple 136.0 → 80.3 (−41 %), harbor 74.0 → 34.0 (−54 %) — the entire drop from construction levers, zero from search machinery, exactly the epic's thesis.

This probe decides 71d (failure-directed topology-repair operator). 71d's premise: the pre-stack harbor 3M-eval plateau (3m.dom, re-scores to 27 fails) is dominated by 13 crinkliness fails, characterised as landlocked rooms (area_outside == 0 → crink == 0 → quality_uncrinkliness hits the if not crink: return 0.0 branch, fitness.py:355 → guaranteed fail for ALL ratios), repairable only by topology — specifically interior O courtyards / facade access. That fix has since shipped DEFAULT-ON (interior_outside, §13.6), so the premise needs re-measuring on the current stack.

Setup (experiments/probe_harbor_floor.py, harbor-house, full default stack, seed 0, 500 000 native evals, staged, SERIAL — the leaf-share relaxed objective is injected by a parent-process fitness.load_config monkeypatch that does NOT reach ProcessPoolExecutor workers, so every §13.x floor run is serial; see homemaker-py-x3b for the production CLI wiring). The probe re-scores the best and splits each crinkliness fail into landlocked (area_outside == 0, 71d's ratio-invariant target) vs under-exposed (0 < crink < target, reachable by ratios/seeding).

metric	old 3M plateau (pre-stack)	full default stack, 500k
total fails	27	20
crinkliness	13	4
landlocked crinkliness	~13	2
top residual class	crinkliness	edge-too-long (6)

Final residual histogram (20 fails): 6 edge-too-long, 4 crinkliness, 4 size, 2 proportion, 2 width, 2 level-not-connected. Re-score OK (relaxed config consistent end-to-end).

VERDICT — NO-GO on 71d as scoped; interior-O already dissolved its target. The landlocked-crinkliness block 71d was built to repair collapsed from ~13 to 2 of 20 — because interior-O seeding is 71d's named fix (interior O courtyards) and now does it by default. Crinkliness is no longer the dominant class; the residual is small and spread across edge-too-long / size / proportion / width / connected, with no concentrated ratio-invariant block for a targeted repair operator to attack. A deterministic repair operator remains a genuine new operator class (not refuted by the §11.4/§11.5/§12.3 search-machinery losses), but its expected value is now low: its highest-leverage target is gone, and what remains is diffuse. Recommendation: close 71d (and prerequisites 7u5/jrb/u8x) as superseded-by-construction; the floor 71d targeted was lowered by interior-O, not by search machinery — consistent with the epic scoreboard. The deprioritised P4 levers erc.5 (compactness cuts — Diag A: floor is leaf-count not cut-quality, and leaf-sharing over-delivered) and erc.6 (inner-loop slack — Diag B: wrong DOF) close wont-fix on unmet revisit conditions, completing the epic.

Caveat (honest): single seed, 500k not 3M, relaxed config vs the old strict standalone 27 — so the 20-vs-27 total is not a clean apples-to-apples. The robust signal is the composition collapse (crinkliness 13→4, landlocked 13→2), which the §13.6 three-seed data corroborates (interior-O reliably cuts harbor landlocked fails). Follow-up observation, not part of this verdict: edge-too-long is now the single largest harbor class (6) — a candidate seed for any future floor work, distinct from the crinkliness regime Phase-8 addressed.

13.8 Experiment: share-aware edge-too-long cap (homemaker-py-hph) — DONE (positive, harmless)

§13.7's follow-up observation (edge-too-long = harbor's top class, 6 fails) is the seed. Dissection first (experiments/diag_edge_too_long.py on the 500k probe best): the 6 fails are only 2 distinct locations. (1) DOMINANT ~4/6: leaf lllr is a share=3 leaf — one quad holding 3 rooms (247 m², edges 15–17 m, aspect 1.2, NEARLY SQUARE). Its walls exceed the flat 8 m cap purely because it aggregates 3 rooms — a leaf-sharing REPRESENTATION ARTIFACT, not a design flaw. §13.3 relaxed size/missing for shared leaves (quality_size centres on k×target) but edge_cost (fitness.py) and outside_edge_cost still used a flat 8 m regardless of leaf.share — the same §13.3 leak on a different measure. (2) ~2/6: leaf llll, a 1.2 m × 16.7 m sliver (aspect 14) — a REAL narrow-room pathology, already independently caught by width/proportion; its edge-too-long is the wall it shares with lllr. No corridors involved.

Fix. New Fitness._edge_cap(*leaves) scales the 8 m cap by the largest type-guarded leaf_share (graph.leaf_share, §13.3's helper) among the adjoining leaves, mirroring quality_size's k×target; non-shared leaves keep the flat cap. Used by both edge_cost (interior wall, max share of the two leaves) and outside_edge_cost (one leaf). Gated behind a new share_edge_cap config knob (SHAREEDGE env), default OFF, so the §13.x controls reproduce. On the probe best the lever clears all 6 edge-too-long (20→14 total fails); the llll sliver stays flagged via width/proportion.

Setup (experiments/run_shareedge_ab.sh, full Phase-8 default stack LEAFSHARE=1/fac3 + DEPTHBAL=1 + INTERIORO=1/odiv3, staged, 20 000 native evals, seeds 0/1/2, final native re-score). Control SHAREEDGE=0 (flat cap) — must reproduce §13.6/§13.7; experiment SHAREEDGE=1.

programme	flat cap off (s0/1/2)	mean	share-aware on (s0/1/2)	mean	Δ
maple-court	74 / 78 / 89	80.3	73 / 78 / 71	74.0	−7.9 %
harbor-house	28 / 41 / 35	34.7	27 / 39 / 27	31.0	−10.6 %

The control reproduces §13.7 (maple 80.3 exactly, harbor 34.7 ≈ 34.0), so the gap is the lever, not drift.

VERDICT — positive and HARMLESS; recommend default-ON. Both programmes improve on the mean with zero regressions across all 6 seeds: harbor every seed (−1/−2/−8), maple two flat/down + one −18 (seed2). The asymmetry of magnitude (maple's big seed2 swing) is search noise, but the direction is structural: the lever only ever removes a false-positive fail on an aggregate shared leaf — it cannot add one (non-shared leaves are untouched), so it is monotone-harmless on the objective. This is unlike the §13.4-family construction levers that trade leaves for fails; there is no tax to wash out. Recommendation: flip share_edge_cap default-ON for leaf-sharing runs (it is the §13.3 relaxation completed on the wall measure), mirroring the pll/interior_outside default flips. A follow-up issue flips the default + rebaselines the §13.x floor numbers (harbor 34.7→31.0, maple 80.3→74.0 become the new full-stack baseline). Repro: experiments/diag_edge_too_long.py, experiments/run_shareedge_ab.sh.

13.9 Flip share_edge_cap default-ON + rebaseline §13.x floor (homemaker-py-rq2) — DONE

Acting on the §13.8 recommendation. Fitness.__init__ now defaults the share-aware edge cap to self._leaf_sharing when share_edge_cap is unset: under leaf-sharing the cap is ON, mirroring the pll bal+share and §13.6 interior_outside default flips. An explicit share_edge_cap=False still reproduces the pre-flip control arm, so the §13.8 A/B and any §13.x control stay reproducible (run_staged_search.py now pins conf["share_edge_cap"] = share_edge explicitly in both arms; the SHAREEDGE override is preserved). Non-sharing runs (every example patterns.config, where leaf_sharing is absent) are untouched — a control re-score of programme-house reproduces bit-for-bit.

New §13.x full-stack floor (Phase-8 default stack, staged, 20 000 evals, seeds 0/1/2): maple-court 80.3 → 74.0, harbor-house 34.7 → 31.0 — the share-aware arm from §13.8 becomes the baseline. test_edge_cap_flat_when_lever_off_even_with_sharing now pins share_edge_cap=False; test_edge_cap_defaults_on_under_leaf_sharing guards the flip. 222 tests pass.

13.10 Productionise leaf-sharing: per-code share + CLI wiring (homemaker-py-x3b) — DONE

Make the §13.3 lever a first-class, programme-author-controllable feature instead of an experiment-only env var + monkeypatch. Three pieces:

1. Per-code grain (SpaceReq.share). patterns.config spaces accept an optional share: N → SpaceReq.share (int, default 1 = not shareable; a has_share flag distinguishes an explicit share: 1 from the default). operators._share_grain(req, leaf_share_factor) resolves each code's grain from the global selector:

• leaf_share_factor == 0 — per-code opt-in: a code shares iff it sets share: N≥2; this is the safe default-on philosophy (sharing off unless the author asks, per space).
• leaf_share_factor ≥ 2 — global mode: every sized code shares at the factor, with an explicit share overriding (share: 1 opts a code OUT, share: N sets that code's grain to N). Reproduces the §13.3 experiment with no edits to example programmes (so §13.3/§13.9 baselines stay reproducible).

Only sized codes are ever shareable (an unsized c/o/s absorbs slack — no target to centre k rooms on). _share_rooms now groups per resolved grain.

2. End-to-end conf injection. The §13.3 scoring sites gate on a leaf_sharing conf key, but example patterns.config files don't set it — the experiment harness monkeypatched fitness.load_config to inject it. Productionised cleanly: load_config(dir, overrides=None) merges run-level keys last, and driver.search / innerloop.optimise / NativeEvaluator / _fitness_for thread conf_overrides={"leaf_sharing": True} through both the inner-loop scorer and the off-tree grade/feasibility scorer when sharing is on. So the whole pipeline scores under the relaxed objective the shared seed targets, with no monkeypatch and no on-disk edits. (share_edge_cap's §13.9 default-ON-under-sharing derivation in Fitness.__init__ rides along automatically.)

3. CLI. homemaker-evolve gains --leaf-sharing/--no-leaf-sharing (default ON, HOMEMAKER_LEAF_SHARING) and --leaf-share-factor N (default 3, HOMEMAKER_LEAF_SHARE_FACTOR), threaded to driver.search.

Default-OFF parity holds: overrides=None leaves load_config byte-identical and _share_rooms is never reached. Smoke-checked end-to-end on harbor-house (sharing on 37 fails vs --no-leaf-sharing 95 at budget 160). 233 tests pass.

14. Island model: multi-run recombination (homemaker-py-psk) — DONE (null)

Lever (user-proposed). Perl Urb ran the search many times and kept the best, because independent runs settle into different local minima. The Python tool is deterministic per --seed, so the analog is an island model with synchronous migration: run N independent seeds to convergence (Phase A), then PRIME a fresh population with those N converged elites and run a second, crossover-heavy phase (Phase B) to recombine basins. Distinct from §11.5 (c4c.5), which injected fresh random/constructive seeds for raw diversity and landed null — here the migrants are fully-converged elites, high-quality building blocks, so the "diversity does not help" result does not directly refute it. The one untested sub-mechanism: can crossover stack wins across independent basins (run A solved cluster X, run B solved cluster Y, child inherits both)?

Design (experiments/run_island_ab.py). Three numbers per programme, all leaf_sharing OFF so controls track the §12.2 baselines (maple 136 / harbor 74), all on equal actual eval budget (the staged search has a hard ~pop·child·2 bootstrap floor, so we account r.n_evals, never the request):

• bestN@A — best-of-N over Phase A (the FREE reference; these N runs happen anyway — the legitimate descendant of Urb's multi-run habit).
• island — Phase B result: a population primed from the N Phase-A elites via the existing seed_factory+bootstrap path (no new representation), evolved at p_crossover=0.7. Total budget = Phase A + migration.
• bestN@T — best-of-N over N independent runs at the same total per seed (the "N+ longer independent runs" control). THE BAR: island must beat it.

A default-off child_probe hook (driver.search) instruments the deciding mechanism: for every crossover child it records whether the spliced child beats max/min(parent fails). Parent fails are appended to the child lineage as |pf=a,b (only when the probe is set) so the signal survives the ProcessPoolExecutor pickle round-trip an id(root) key cannot.

Result (N=4, master_seed 0, 28160 actual evals/arm, 4 workers):

programme	bestN@A	island	bestN@T	verdict	crossover beat-min-parent
harbor	73	68	67	loses by 1 (within noise)	1 / 65
maple	134	124	116	loses by 8 (decisive)	3 / 63

Verdict: NULL / negative. The island model does not beat best-of-N at equal total budget. On harbor it ties-to-loses inside the parallel noise band; on maple it loses clearly (124 vs 116) — a single longer independent run reached 116 while the migration phase, given the same budget, stalled at 124. The migration phase buys nothing a longer independent run does not.

The mechanistic probe explains why (the deciding diagnostic). Crossover across independently-converged elites almost never synthesizes: of ~64 crossover children only 1/65 (harbor) and 3/63 (maple) beat the better parent, with a best fail-drop of just 2 and 5. This confirms the issue's alignment hypothesis: operators.crossover is area-matched subtree exchange, but two independently evolved trees encode similar arrangements at different paths/areas (the encoding is non-canonical — 9gp closed negative), so the splice is mostly disruptive, not combinatorial, and the inner loop re-solves ratios at the boundary (spliced quality not preserved). The null is therefore mechanistic, not budget.

Noise caveat (carry forward). Phase A is unaffected by the probe, yet harbor seed 2 scored 71 then 73 on byte-identical re-runs — parallel/BLAS non-determinism, the same ±2-3 effect §12.4 flagged. Sub-±3 verdicts under n_workers>1 are noise; both arms here ran at the same worker count so the comparison stays fair, and maple's −8 is safely outside the band.

This is the third search-machinery null after §11.4 (graded objective) and §11.5 (niching+restarts) / §12.3 (M3 + shape filter), against four construction/seed wins (§11.6, §11.7, §12.2, §13.x). best-of-N at the Phase-A budget remains a free, worthwhile habit; a dedicated migration phase is not worth its budget. The residual stays geometry/shape-bound. NOT gated on canonical encoding (9gp closed); the child_probe hook is kept default-off for reuse.

15. Leaf-sharing output honesty: unfold + polish auto-finish (homemaker-py-3l6) — DONE

Bug. Leaf-sharing (§13.3/§13.10, default ON) is a fitness-evaluation knob: a shared leaf of code X with share=k is credited as satisfying k programme entries, its size Gaussian re-centred on k*target. So the evolve inner objective rewards genomes that under-materialise the programme (fewer, larger rooms), but the winning .dom written to disk is that un-materialised genome. Re-scored by the canonical homemaker-fitness (sharing OFF), the un-materialised copies become missing required space (critical) fails. Measured on harbor-house (init.dom, 3M budget): internal best 1.03e-05 but canonical 6.73e-29, 90 fails (15 critical). The default silently optimised an objective the canonical scorer does not credit and wrote a catastrophically worse building than its reported internal fitness implied.

Investigation (homemaker-py-yaa). Four fixes were scoped (make no-sharing the default; re-score-and-warn on write; materialise shared leaves on write; anneal the grain to 0 mid-run). yaa characterised the transferability of a sharing-phase solution to the honest objective and reached a conclusive result:

• Naive warm-start from a raw sharing seed stalls (harbor 8.66e-08, 70 fails) — place_missing/divide cannot dig out the ~15-room count deficit fast enough.
• Warm-start + unfold (operators.unfold_shared_leaves at the transition) catches the direct no-sharing route (4.19e-06, 15 fails, 0 critical), matching the --no-leaf-sharing baseline (nols-2 4.19e-06). Bruno's key idea confirmed: the sharing phase's transferable value is the adjacency/topology skeleton, and the sole blocker to reusing it is the materialisation (count) deficit — not the k*target sizing mismatch. Unfold pays that deficit down.

operators.unfold_shared_leaves(root). Replaces every live shared leaf (share>1, share_type==type) with a balanced binary subtree of k same-code leaves splitting its footprint, sizes each for squarest proportion, and clears the share stamps. Footprint (plot area) is preserved; the adjacency skeleton is otherwise untouched. Returns the number of extra leaves created.

Fix — auto-finish before write (driver.polish_finish). Rather than unfold on write alone (honest room count but un-polished proportion/width/size on the fresh children), the finish runs yaa's proven unfold-then-polish as an automatic terminal phase. When a run used --leaf-sharing, before write:

1. deep-copy the best, unfold_shared_leaves it (materialise the deficit);
2. warm-start a leaf_sharing=False search (bootstrap=False) from the unfolded genome for --polish-budget evals — local search under the honest objective cleans up the newly materialised rooms.

The returned best.fitness is then the canonical score (sharing OFF ⇒ internal == canonical), and eval/topology/history accounting is stitched onto the sharing run with the two phases tagged share:/polish: (the objectives are not comparable, so the histories are concatenated, not merged). --polish-budget (env HOMEMAKER_POLISH_BUDGET): -1 = auto = budget//2, 0 = unfold + single rescore only (no search). An interrupt forces polish_budget=0 so a stopped run still writes an honest output without triggering a long extra phase.

The default stays --leaf-sharing ON: its ~35 %-faster topology search (§13.3) is retained, and the output is made honest by the finish instead of by disabling the lever. Option 1 (no-sharing default) and option 2 (warn-only) from the bug were therefore not needed; the annealing option is its own follow-up (Schedule B, homemaker-py-kpu) — the single-transition finish here is its proven precursor, and unfold_shared_leaves is the primitive it will reuse at each grain step.

Verification. harbor-house, budget 3000 + polish 1500: the reported polish fitness 4.79788e-27 matches the canonical homemaker-fitness byte-for-byte, 0 critical fails (the missing-room criticals are gone — the 15 shared-leaf copies are materialised). Small budget so absolute quality is low, but the honesty — the point of the bug — is restored. Tests: driver.polish_finish ×3 (unfold+rescore stitching, polish-search accounting, no-best noop); 254 pass.

16. In-run leaf-share grain annealing — Schedule B (homemaker-py-kpu) — DONE (negative)

Premise. §15's finish crosses the sharing→off objective cliff in a single hard transition (unfold every shared leaf at once, then polish). Schedule B (yaa's still-open option) instead ramps the grain down within one continuous run — e.g. 4 → 3 → 2 → off — carrying the whole population across each step. Graduated non-convexity: the coarse early grain fixes gross topology/adjacency on a small effective problem (few, large rooms); each step materialises a little more and refines per-room size/proportion/width; no single fitness cliff is crossed at once. The question (kpu): does a graduated ramp beat the single hard unfold transition (yaa's warm-chain 4.19e-06) and the direct no-sharing baseline (5.14e-06)?

8iv settled the unfold primitive first (NEGATIVE). kpu originally "wanted" the circulation-aware unfold from homemaker-py-8iv (route the materialised subtree's access through interior children). 8iv built and A/B-tested it and it lost to the plain balanced-grid unfold_shared_leaves (slice 41 fails vs grid 25 at 150k evals, grid leading throughout). So Schedule B reuses the existing grid unfold at every grain step — no slicing reintroduced; access is left to local search on the squarer grid seed (which yaa showed reaches 4.19e-06).

Mechanism (driver.search_annealed). One phase per descending grain in grain_ladder (default (4, 3, 2)), then a de-share polish:

1. Phase 0 (grain = ladder[0]): a normal search — constructs the population at leaf_share_factor = cap with the evaluator's leaf_share_max capped to cap (new max_share override, threaded through _overrides_for/_fitness_for/ _evaluate).
2. Each grain step (cap lowered): before resuming, unfold every population leaf whose share exceeds the new cap — operators.unfold_shared_leaves(root, above=cap) — so the leaves the lower cap would under-credit become real rooms instead of fresh missing fails; the rest stay collapsed for the next step. The whole population is then handed to the next search via the new seed_pop argument (each root re-optimised and re-scored under the lower cap), preserving topology/adjacency continuity rather than restarting from a single best.
3. Finish (grain off): unfold all remaining shared leaves (above=1) and run a leaf_sharing=False search (or a single rescore when polish_budget <= 0 / on interrupt), so the returned best.fitness is the honest canonical score exactly as §15 guarantees (verified: annealed output re-scored by homemaker-fitness matches the reported best byte-for-byte).

budget is split evenly across the sharing phases; polish_budget funds the finish. Phases are stitched with cumulative eval/topology accounting and a grain-tagged history (g4:/g3:/g2:/polish:) — objectives differ across grains so histories are concatenated, never merged. CLI: homemaker-evolve --anneal-grain 4,3,2 (implies sharing; self-finishing, so the §15 finish is not applied on top).

Verification (plumbing). harbor-house, budget 900 (300/phase) + polish 300, 4 workers: unfolds 33 → 22 → 9 leaf-copies across the ramp, population carried (anneal-seed/* lineages), honest share-free output whose reported best 3.36672e-29 matches homemaker-fitness byte-for-byte. (Fails rise at this toy budget — 64 leaves materialised with almost no recovery budget — so absolute quality is meaningless here; the head-to-head below runs at the baselines' budget.) Tests: unfold_shared_leaves(above=) grain-cap selectivity; search(seed_pop=) population seeding; search_annealed phase stitching / honest finish / degenerate-ladder fallback; 258 pass.

Head-to-head (DONE — NEGATIVE). harbor-house, init.dom, seed 0, pop 16, child 80, grain 4,3,2, budget 1.5M (500k/phase) + polish 1.5M = 3M total (workers 4, ~22h), matched to the yaa baselines. Result: 1.26e-08, 23 fails (canonical homemaker-fitness byte-for-byte). Both targets beat it decisively:

route	fitness	fails
direct --no-leaf-sharing	5.14e-06	15
yaa warm-chain (single hard unfold)	4.19e-06	15
Schedule B (graduated 4→3→2→off)	1.26e-08	23

~400× worse fitness, +8 fails. Verdict: the graduated grain ramp loses to the single hard sharing→off transition (§15). The trajectory shows why — each grain step spikes the fail count as its unfolded leaves acquire independent shape fails (phase-end fails 19 → 21 → 27, then the final de-share unfold 27 → 36), and the per-phase budget re-polishes a partially-materialised state that the next step materialises further, so the coarse-grain gains (19 fails at grain 4) do not carry forward. Splitting the budget across three intermediate materialisations left the polish phase starting from a deeper hole (36 fails) than the warm chain's single clean transition, and 1.5M polish evals recovered only to 23 — short of the 15 both baselines reach. Graduated non-convexity is falsified for this materialisation cliff: the transferable value is the sharing-phase topology skeleton (yaa), and it is best cashed in once, at full grain, not annealed. (Caveat: this run used workers=4 vs the baselines' workers=1; the ~400×/+8-fail gap is far larger than worker-count trajectory noise, so the direction is robust.)

The machinery is retained (search_annealed, --anneal-grain, unfold_shared_leaves( above=), search(seed_pop=), the max_share evaluator override) — it is correct, tested, and honest, and the seed_pop / grain-cap primitives are reusable — but the default finish stays §15's single-transition unfold+polish. Tests: 258 pass.

17. Finish-time global cell→room collapse (homemaker-py-94g) — DONE (positive)

Motivation — label-relative fails. A layout's leaf carries a room type, and many of a good layout's residual fails are label-relative: a cell fails size / width / proportion only because the room assigned to it wants dimensions it lacks — relabel that cell to a room it fits and the fail vanishes; a wrong-level fail is likewise a labelling error. On the harbor-house best layout (evolved-3M-nols-3, 15 fails) ~11 of 15 are label-relative. This is separable from the geometry-intrinsic fails §13 chased at the shape floor — long-thin useless cells (width/proportion/ crinkliness) and not-connected — which no relabelling can fix because the cell's geometry, not its label, is wrong. The collapse targets only the former.

Mechanism (Fitness.collapse_global). A one-shot, finish-time pass that relabels the whole building's room cells in one optimal assignment — the 9o5 per-class collapse (interchange superposition) generalised from one equivalence class to a global N inside-leaves ↔ M required-rooms matching (_best_assignment: brute-force under the class cap, else Hungarian). SUPPLY = leaves whose type is an assignable room code; DEMAND = every such code expanded by its required count. Constraints, each landed after an empirical correction (below):

• c/o/s partition. Assignable codes exclude any starting c/o/s. check_space_counts (graph.py) skips those as circulation/outside/sahn — including room codes that collide with the convention (cr1 Common Room, st1/st2 Storage). Those leaves are the circulation/structure skeleton and must never be relabelled; the collapse uses the same partition the scorer counts against.
• Hard level. A leaf may take a room only if its storey matches the room's required level (a −1e12 forbid penalty), so the collapse never adds a wrong-level fail.
• Adjacency relaxation. Geometry is fixed at finish time, so each leaf's graph neighbours are fixed and only labels move. Required adjacencies become a labelling relaxation: warm-started from the evolved labels, each pass is a linear assignment over the base value plus a bonus for each of a code's adjacencies satisfied by the current neighbour labels, iterated to a fixpoint (Jacobi/WFC-style). Only room↔room adjacencies can break — adjacencies to c/o are invariant since those leaves are never relabelled.
• Threshold objective. The base per-cell value is either continuous fit (sum(usage_quality*area), as 9o5) or — the default — the count of size/width/ proportion factors that pass (≥ FAIL_THRESHOLD), with continuous fit only as a tiebreak. A satisfied adjacency and a passing factor carry the same unit weight (_COLLAPSE_FAIL_W), so the collapse minimises (adjacency + size/width/proportion) fails jointly.
• Public-access pin. The building-level "no outside public access" check is existential (∃ a public street-edge outside leaf with an l/c/k neighbour) — invisible to the per-leaf objective. When the sole provider is an l/k room neighbour (no circulation fallback), that leaf is pinned (kept, its demand slot decremented) so the collapse cannot drop the check.

Two corrections found by measurement. A naive first cut (level-only, per-leaf, continuous fit) went 15→46 fails. Diagnosis killed two hypotheses: (1) the count explosion was not a merge effect (merge_divided merges only outside/sahn siblings, never rooms) but the c/o/s partition bug above — pulling cr1/st1/st2 into the assignment shredded the circulation skeleton; fixing the partition took +31→+1. (2) The residual +1 was the continuous objective shuffling a size fail from one leaf to another (pushing one just over the 0.1 threshold and another just under); the threshold objective optimises the fail count directly and removes it.

Keep-better + wiring. Fitness.collapse_finish scores baseline and collapsed on throwaway copies (scoring merges in place) and keeps the collapse only if the fail count does not increase — a strictly monotone safety belt. driver.collapse_best applies it to a SearchResult's best, canonically re-scoring and tagging lineage +collapse. evolve.py runs it after the §15 sharing finish behind --collapse/--no-collapse (default ON). Standalone homemaker-collapse <file.dom> (collapse_cmd.py) applies it to an existing layout, writing <stem>.collapsed.dom.

Verification. Sweep over 6 harbor-house evolved layouts (total fails, base 195): adj_off/quality 192, adj_on/quality 185, adj_off/threshold 181, adj_on/threshold 171. The default (adjacency=True, objective="threshold", public-access pin) is monotone across all 6 (never worse than baseline; keep-better guard is a belt, not needed here) — best layout 15→12, and e.g. 32→26, 90→82. The residual on the best layout is geometry-/building-bound, not label slack: the collapse searches labels only, never geometry, so it cannot touch long-thin cells or not-connected — those are spun out to homemaker-py-7fm (shape reshape) and homemaker-py-qi6 (circulation placement). Running the collapse inside search per-eval (rather than finish-time) is homemaker-py-qpk, gated on the 9o5 landscape-flattening risk (§13 / homemaker-py-xi7) and its own A/B. Tests: tests/test_collapse_global.py ×6 (demand-set relabel, level hard constraint, c/o/s exclusion, no-op safety, keep-better/unmerged); 267 pass.

18. Graded circulation-connectivity signal (homemaker-py-qi6) — DONE (negative)

Motivation — the binary fail is flat. After the §17 collapse, the residual fails on the harbor-house set are dominated by level N not connected (2 of the best layout's 12; also on 5 of the 6 sweep layouts). That fail comes from connected_circulation (graph.py): remove every non-circulation vertex from a storey's adjacency graph and require the remaining circulation cells (C stairs plus the cr/st room-codes that collide with the c/s prefix) to form ONE connected component. On the evolved layouts they instead fragment into 4–7 components per storey.

Why finish-time repair fails (measured, negative). The obvious §17-style companion — a finish-time pass that re-types boundary cells to circulation to bridge the components, kept only if the fail count does not rise — was prototyped (Steiner-MST bridge set per disconnected storey, keep-better guard) and measured on the 6 layouts: 195 → 560 fails (+365). The not connected fail is binary (one fail per storey regardless of fragmentation), but each storey needs 3–7 bridge cells, and every needed-room→circulation conversion triggers a missing-room fail cascade (2–5 fails) that dwarfs the single connectivity fail it clears. Keep-better reverts every one → no-op. Conclusion: connectivity cannot be bought at finish time when every cell is a needed room; it must come from the outer search allocating connected circulation topology. But the binary fail gives the search zero gradient — a 7-component storey scores identically (both in fail count and in the 0.5^n scalar) to a 2-component one — so the search cannot tell it is making progress.

Mechanism — a graded proximity on the same channel §11.4 built. graph.circulation_connectivity(G) returns the fraction of circulation cells in the largest connected circulation component ∈ [0,1] (1.0 = a single connected spine, lower = more fragmented, 0.0 = no circulation), measured on the same circ subgraph the fail uses so the two agree at the connected endpoint. Summed over storeys it is the graded proximity scalar Fitness.score_with_grade already carries for the outer comparator, gated by the conn_grade conf flag: when on it replaces the §11.4 leaf quality-proximity on that channel (a distinct, better-motivated use — §11.4 was rejected because within a fail-tier the 0.5^n scalar is NOT flat there and grade merely displaced a working signal; connectivity is the opposite case, genuinely flat under the binary fail). Like §11.4 it leaves the scalar fitness and fail count byte-identical (verified) — it is only the secondary key (-n_fails, grade, fitness) (driver use_lex and use_grade), strictly beneath fail-count so the §6 missing-space hierarchy and the §5.4 inner-loop cliff are untouched. Among equally-failing neighbours the search now prefers the one whose circulation is closer to one component, restoring the gradient toward connected topologies.

Wiring. conn_grade threads through _overrides_for/_fitness_for/_evaluate and the search signature; enabling it implies the grade key. evolve.py exposes --conn-grade (env HOMEMAKER_CONN_GRADE, default OFF); the grade is read off the optimised tree, one extra native eval per child.

Build. Signal, fitness wiring, CLI, and 9 tests landed (tests/test_conn_grade.py: pure-graph fraction contract, non-circ cells ignored, monotone under (dis)connection, and the score/fail- count-invariance of the flag). 276 tests pass.

A/B verdict (measured, 2026-07-22, qpk protocol, experiments/run_qi6_ab.sh) — NEGATIVE. Equal-budget conn_grade ON vs OFF, both arms finished with the standard finish-time --collapse (94g), 4 workers, canonical homemaker-fitness re-score for the .fails breakdown:

• harbor-house (init.dom, budget 2500, seeds 1–3): byte-identical output in every seed (dom, fail list, fitness all diff-clean ON vs OFF) — the secondary comparator key never fired, i.e. the search trajectory never actually hit a tie at fail-count that the grade could break. This is the programme §18 was motivated on (2 of 15 fails on the best layout are not connected), and the signal moved nothing.
• programme-house (init.dom, budget 3000, seeds 1–5): 3/5 seeds tie exactly (byte-identical .fails); seeds 1 and 2 diverge to a different topology with one fewer total fail (8→7 each) — but the diff is entirely adjacency/crinkliness/width/access/size fails, not connectivity. In all 4 seed-arms across both programmes where a not connected fail was actually present (harbor 1&3, programme 3&4), the fail is unchanged in both arms — zero cases of the grade clearing one.
• Conclusion: the grade does not do what §18 designed it to do. It occasionally perturbs tie-breaking among equal-fail-count neighbours (programme-house seeds 1/2), which can incidentally shift the total fail count, but that perturbation never targets circulation connectivity specifically — consistent with a comparator key that is either too weak relative to the primary (-n_fails, fitness) keys to steer topology choice, or whose grade values are rarely distinct enough between the actual neighbours the search compares to break a tie in the intended direction.

Status / next. Kept default OFF (already was). Mechanism (b) (graded proximity as a tertiary key) is falsified by this A/B, not just unconfirmed — do not re-attempt without a different mechanism. The remaining candidate from the original issue is mechanism (a): an explicit insert/relocate-circulation mutation/repair operator, which does not depend on the search stumbling onto a fail-count tie to act. Not started; low priority per DISCOVERED-FROM epic homemaker-py-94g's framing (fitness fidelity, not search capability).

19. Geometry/topology repair for shape-intrinsic fails (homemaker-py-7fm) — DONE (negative)

Motivation. §17 established that ~12 of the harbor-house best layout's 15 residual fails survive the label-only collapse — long-thin cells (width/proportion/crinkliness) whose geometry, not room assignment, is wrong. bd memory collapse-global-94g-and-any-label-usage- optimisation spun this out as its own problem: a mechanism that moves geometry, evaluated for net fail-count effect on the same 6-layout sweep §17 used.

Diagnosis (rules out mechanism (a)). Re-ran the full-fitness ratio inner loop (innerloop.optimise, Nelder-Mead, 1500 evals, warm-started from the evolved ratios — far above the ~80-200/child budget search actually spends) on the 12-fail collapsed best layout: zero change, byte-identical fail lines. These are not local optima of the ratio search reachable with more budget. Tracing two representative fails back through the tree found two distinct structural causes, neither fixable by re-solving ratios on the existing cuts: (1) area starvation — a leaf's defining branch (several levels up) was allocated too little total area for what it has to share with its siblings (a storage leaf wanting 18m² sat in a 6.4m² branch whose sibling got 52.8m² of outside space); (2) orientation mismatch — a leaf is the correctly-area-sized-but-thin remainder of a cut whose rotation runs parallel to its parent rectangle's long axis, so no ratio value on that axis avoids a sliver.

Mechanism (operators.mutate_shape_rotate, operators.mutate_deslim). Two targeted repair operators addressing each cause, in the mutate_level_fix style (structural, not blind-random): _shape_failing(leaf, fit) identifies a named-room leaf whose width or proportion factor actually fails (< FAIL_THRESHOLD under Fitness.quality_width/quality_proportion — not a geometric proxy, which over-flags leaves the Gaussian tail still passes). mutate_shape_rotate re-orients the live cut that produced a failing leaf (targets cause 2); mutate_deslim merges a failing leaf into its sibling, undoing the division that starved it (targets cause 1), leaving the displaced room for mutate_place_missing (already in MUTATIONS) to re-insert elsewhere. Both are registered in MUTATIONS/mutate(), gated on a fit argument (a new fit_ops class alongside the existing reqs_ops) so they no-op — and are excluded from the outer search's weights — wherever a Fitness instance isn't threaded through, exactly as place_missing etc. gate on reqs. driver.search/evolve.py do not yet pass fit through (see Status below), so the operators exist but are currently unreachable from the GA — they were evaluated instead as a finish-time greedy hill-climb (below).

Verification (measured, negative). A finish-time hill-climb applied both operators exhaustively — for every live cut driving a shape fail, all 3 alternate rotations were tried (not just mutate_shape_rotate's single random draw) alongside a deslim + place_missing + ratio-resolve, keeping the best only if it did not increase the fail count — on the same 6 harbor-house evolved layouts as §17 (total fails 187): 0 improving moves found on any layout, on any candidate cut, under any of the 4 tried variants. Manually inspecting the rejected candidates for the representative case (harbor-house evolved-3M-nols-3, leaf 0/rlrlr "la1", the proportion fail traced above) shows why: every one of the 3 rotations and the deslim+ reinsert produced a worse layout — new no outside public access, not adjacent to c, access, or edge too long fails, in every trial. This is §4.2's core lesson (proxy/partial- objective repair of a co-evolved local optimum "is structurally unable to win" — every cut position is simultaneously a size/shape knob and an adjacency/access/circulation knob) now confirmed for structural topology repair, not just ratio-solving: on a tightly co-evolved layout, the cut that makes a leaf thin is also the cut providing some other leaf's public- access or adjacency, so straightening it elsewhere is not free. The residual geometry-intrinsic fails on the harbor-house best layout appear to be close to a genuine Pareto floor for this topology, not a repairable inefficiency — consistent with §17's own framing ("geometry-/ building-bound").

Status / next. mutate_shape_rotate/mutate_deslim land in operators.py, default-excluded from mutate() (no fit threaded through the outer search yet), with dedicated tests (tests/test_operators.py: fail detection, noop-without-fit, targeted-cut selection, merge + place_missing repairability) plus automatic coverage via the existing test_mutations_yield_canonical_genomes parametrisation. 282 tests pass. The finish-time hill-climb script is not productionised (unlike §17's collapse_cmd.py) because it never found an improving move to apply — there is nothing to wire up. Not tested: whether these operators help as in-search GA moves (mechanism (c)) — a full multi-generation run gives selection pressure and population diversity a chance to accept a locally-worse move that a later step or recombination completes, a fundamentally different regime from single-step greedy hill-climbing on an already-finished layout. That A/B (thread fit through driver.search, gate with an enable_shape_repair-style flag as §12.3 did for reassociate, run full-budget with/without) is the remaining open question and would need to be its own measured experiment before further code changes — this session's finding is that the finish-time half of the issue's candidate mechanisms is a dead end, not that geometry repair is impossible in general.

In-search follow-up (measured, 2026-07-22, homemaker-py-161) — also negative. driver.search/search_staged gained enable_shape_repair: bool = False, threading a cached Fitness instance into operators.mutate() only when set (mirrors enable_reassociate's clean- toggle pattern; default off reproduces prior runs byte-for-byte). Full A/B on harbor-house init.dom cold-start, budget=1,000,000, pop=16, child_budget=80, workers=4, seeds 0–3: fails [14,15,12,17] mean 14.50 (off) vs [17,14,16,12] mean 14.75 (on) — no improvement, and the 0.25 delta is far inside the 12–17 seed-to-seed spread in both arms. A smaller pilot (budget=20000, 3 seeds) matched: off mean 31.33, on mean 32.00. In-search selection pressure and population diversity do not rescue shape_rotate/deslim on harbor-house-scale programmes either — the residual fails look like a genuine floor for this representation on this programme, not an inefficiency reachable by richer local operators, in either regime. Code kept (not reverted) for reuse/reproducibility per the enable_reassociate precedent; test test_enable_shape_repair_threads_fit_into_mutate in tests/test_driver.py. Both halves of §19's candidate mechanism space (finish-time and in-search) are now closed negative.

20. In-search global collapse (homemaker-py-qpk) — DONE (positive, size-dependent)

Motivation. §17 (94g) landed the FINISH-TIME global cell↔room collapse — a one-shot label search over the already-searched geometry, applied once to the best layout at the end (harbor-house best 15→12). The original 94g thrust was the PER-EVAL version: run the same collapse inside every fitness eval during search, so the outer GA optimises the collapsed (relabelled) objective directly instead of discovering it only at the end. Deferred behind its own A/B because 9o5 (§13/xi7) found the analogous per-class collapse-as-relaxation NULL/NEGATIVE (OFF beat ON on both example programmes) — the risk carried forward here, AMPLIFIED to global scope, is that max-over-labellings flattens the fitness landscape (many topologies collapse to similar scores) and removes the gradient the outer search climbs.

Mechanism (build). Fitness.collapse_global (§17) is called inside _evaluate_full, gated by a new collapse_insearch conf flag (default OFF, bit-identical when off — same contract as superpose/ conn_grade), at the same point collapse_superposition (9o5) already runs: before any Phase-1 check, on the unmerged tree, so check_space_counts/adjacency/quality downstream see the collapsed labels. Two knobs, both conf-driven: collapse_insearch_adjacency (default True — the fixpoint Jacobi relaxation §17 describes) and collapse_insearch_iters (default 3, vs finish-time's 6 — a per-eval cost, not a one-shot polish; lower until profiling says otherwise). preserve_public_access is always on (never safe to drop silently mid-search). Plumbed through the same minimal path as conn_grade (driver._overrides_for/_fitness_for/_evaluate/search, evolve.py --collapse-insearch / HOMEMAKER_COLLAPSE_INSEARCH) — not threaded into search_staged/ search_annealed/polish_finish, matching conn_grade's existing footprint.

Verified (build-time). On evolved-3M-nols-3.dom (harbor-house, the §17 15→12 fixture), collapse_insearch reaches the byte-identical 12-fail collapsed state as the finish-time pass — expected, since it is the same collapse_global call moved earlier in the same pipeline on a fixed geometry. Flag off reproduces baseline score/fails exactly. tests/test_collapse_insearch.py (8): defaults, conf knobs, _evaluate_full wiring (mocked call-site assertion: fires with the right kwargs when on, never when off), and the end-to-end 15→12 cross-check. 290 tests pass. A 60-eval CLI smoke run (--collapse-insearch, programme-house) confirms the plumbing only, no crash — not a result (mirrors qi6's smoke-only checkpoint).

Cost (measured, evolved-3M-nols-3.dom, 20-eval average). Baseline eval 106 ms; with collapse_insearch + adjacency 205 ms (1.9×); adjacency off 157 ms (1.5×). Per-eval cost is therefore real but not prohibitive at this building size — no incremental/cached variant was needed to make the experiment affordable, contrary to the issue's worst-case worry. A full-budget run will cost roughly 2× the wall-clock of an equal-budget baseline run.

A/B verdict (measured, 2026-07-19, xi7 protocol) — POSITIVE, and the OPPOSITE of the 9o5/xi7 prior. Equal-budget collapse_insearch ON vs OFF, both arms finished with the standard finish-time --collapse (94g) so the comparison is apples-to-apples on the final COLLAPSED score, 4 workers:

• harbor-house (init.dom, budget 2500, seeds 1–3): ON wins 3/3, mean fails 80.3 → 72.0 (s1 85→74, s2 76→65, s3 80→77) — a consistent ~10% fail reduction, no losses.
• programme-house (init.dom, budget 3000, seeds 1–5): ON wins 3/5, mean fails 8.4 → 7.8 (s1 8→5, s2 8→7, s4 10→9 win; s3 8→9, s5 8→9 loss by one fail) — a weaker, noisier signal on this much smaller building, already closer to its geometry floor (§13/§19).
• Combined head-to-head: ON 6, OFF 2.

Unlike 9o5 (a per-CLASS relaxation over interchangeable-but-not-identical codes, where max-over- labellings blurred which topology was actually good), the global WFC-style matching here is the same mechanism §17 already proved monotone/positive at finish time — running it every eval just lets the outer search see the condensed objective instead of discovering it only once, and evidently that gradient is real, not flattening, at least at the scale tested. The effect scales WITH building size (more leaves → more relabelling headroom per eval), the opposite of what the 9o5 fear predicted.

Cost (wall-clock, matches the profiled 1.5–1.9× per-eval figure above). harbor-house mean 102.6s (OFF) → 177.8s (ON), ~1.73×. programme-house mean 39.0s (OFF) → 43.7s (ON), ~1.12× (smaller building → collapse is a smaller fraction of total eval cost).

Status (2026-07-19). Kept default OFF — the programme-house result is too mixed (2 losses in 5 seeds) to flip the default on a small sample, and 9o5/xi7 is a fresh enough scar to want a second, larger-budget confirmation before doing so. But this is a genuine, working, opt-in improvement for larger buildings: --collapse-insearch is documented and ready to use on harbor-house-scale (or bigger) programmes today. A natural follow-up (not filed, low priority) would be a larger-N seed sweep on programme-house alone to see whether the mixed result is just small-sample noise around a true small positive, or a genuine size threshold below which in-search collapse doesn't pay for its ~1.1–1.9× cost.

Larger-N confirmation (homemaker-py-1ph, 2026-07-24) — DEFAULT FLIPPED TO ON. Re-ran the programme-house arm alone at 4× the sample: same protocol (init.dom, budget=3000, 4 workers, both arms finished with the standard finish-time --collapse), 20 fresh seeds (1–20) instead of 5, on the current codebase (post-qpk commits through 161, none of which touch the default-off code path):

• Mean fails: 7.95 (OFF) → 7.10 (ON), a ~10.7% reduction — consistent in direction and magnitude with the original 5-seed sample (8.4 → 7.8) and with harbor-house.
• Head-to-head (excluding 3 ties): 11 wins / 6 losses for ON (was 3/2 at N=5).
• Paired t-test on the 20 per-seed diffs: mean diff 0.85 fails, t=2.38, df=19, two-tailed p ≈ 0.028 — the mixed 3/5 result was small-sample noise around a true small positive, not a genuine programme-house-scale exception.
• Cost: ON still ~1.2–1.3× OFF wall-clock at this size (20.9s mean OFF → 26.4s mean ON), same order as the original measurement.

Confirms the qpk verdict holds at both example scales tested. collapse_insearch default flipped OFF → ON in evolve.py (--collapse-insearch/--no-collapse-insearch, HOMEMAKER_COLLAPSE_INSEARCH) and driver.py (_overrides_for, _fitness_for, _evaluate, search, polish_finish) — opt out per-run with --no-collapse-insearch if a specific programme needs the cheaper finish-time-only path. fitness.Fitness itself is unchanged (still defaults off when collapse_insearch is absent from conf — the default lives in the driver/CLI override layer, same contract as leaf_sharing).

21. Insert/relocate-circulation repair operator (homemaker-py-8sh) — DONE (mixed, kept off)

Motivation. qi6's remaining candidate (§18): mechanism (a), an explicit search-time mutation/repair operator that inserts or relocates a circulation cell to bridge a disconnected circulation component directly, rather than relying on the outer GA to discover connectivity via a comparator-key gradient (mechanism (b)/(c), measured NEGATIVE — the grade never fired on harbor-house and never cleared a genuine not connected fail on programme-house).

Mechanism (build). operators.mutate_bridge_circulation: for each storey, builds the leaf adjacency graph (geometry.leaf_graph) and the circulation sub-components (dom.is_circulation nodes only, mirroring graph.connected_circulation's subgraph). When a storey has more than one component, finds the cheapest path between any pair via a weighted Dijkstra search — edge weight is the average of its endpoints' conversion cost (0 for an already-circulation node or a generic outside O leaf — nothing displaced, same rationale as place_missing's host ranking; 1 for any other non-required leaf; 5 for a leaf typed as a required programme room, crossed only if no cheaper route exists) — and retypes every intermediate leaf on the cheapest cross-component path to C. A displaced required room becomes a missing-space fail for the existing place_missing operator to re-insert elsewhere on a later step, the same division of labour mutate_deslim (§19) uses. Registered in operators.MUTATIONS as a "reqs-optional" op — unlike level_fix/ place_missing it is never zero-weighted for lacking reqs (it needs only the tree's own adjacency graph), so gating is done the reassociate way instead: driver.search's new enable_bridge_circulation flag (default OFF) zeroes its mutation_weights entry rather than relying on an argument being None. Threaded through search_staged and exposed as evolve.py --bridge-circulation / HOMEMAKER_BRIDGE_CIRCULATION. 6 unit tests (tests/test_operators.py): noop when already connected, bridges a synthetic 3-leaf fragmented fixture via the free leaf, falls back to bridging through a required room when it is the only route, and prefers a free O leaf over a required room when both routes tie in hop length. 296 tests pass.

A/B verdict (measured, 2026-07-24, qi6/qpk protocol, experiments/run_8sh_ab.sh). Equal-budget enable_bridge_circulation ON vs OFF, both arms finished with the standard finish-time --collapse (94g), 4 workers, canonical homemaker-fitness re-score for the .fails breakdown — harbor-house (init.dom, budget 2500, seeds 1–3), programme-house (init.dom, budget 3000, seeds 1–5):

programme	seed	fails OFF→ON	not-connected OFF→ON
harbor-house	1	74→67	0→2
harbor-house	2	65→65 (byte-identical)	1→1
harbor-house	3	77→77 (byte-identical)	1→1
programme-house	1	5→5 (tie, fitness differs)	0→0
programme-house	2	7→7 (tie, fitness differs)	0→0
programme-house	3	9→7	1→1
programme-house	4	9→8	1→0
programme-house	5	9→7	1→0

Total fails: harbor-house mean 72.0→69.7, programme-house mean 7.8→6.8 — never worse on any seed (4 wins, 4 ties, 0 losses on total fail count across both programmes). Of the 5 seed-arms whose OFF baseline actually had a not connected fail, 2/5 cleared it (programme-house seeds 4 and 5) — a genuine improvement over qi6 mechanism (b)'s 0/4. But harbor-house seed 1 shows the flip side: its OFF baseline had no not connected fail (0), and ON introduces two — while simultaneously landing the sweep's single largest fail-count win (74→67, fitness 3.2e-26→5.0e-24, almost two orders of magnitude apart) via a visibly different topology, not a locally-adjusted one. mutate_bridge_circulation only ever converts a leaf to circulation, never away from it, so it cannot mechanically increase fragmentation itself — the regression is trajectory-divergence noise (adding any nonzero-weight entry to operators.mutate's weighted draw perturbs the RNG mapping for every subsequent draw, not just the ones that select the new operator, exactly as observed for enable_reassociate/enable_shape_repair/homemaker-py-161 — the same-seed off/on comparison is two genuinely different searches from the same seed, not a controlled single-variable diff). Two of the three harbor-house seeds never diverged at all (byte-identical fitness to 6 significant figures) — at _MUTATION_WEIGHTS' default uniform weighting the operator is drawn roughly 1-in-17 times a mutation fires, and evidently often never lands on a fragmented storey within a 2500-budget run.

Status. Directionally positive and clearly better-targeted than qi6's graded signal (which cleared zero not connected fails in its own measured protocol), but the N=3/N=5 sample is too small and too trajectory-noisy to separate a true small positive from chance, per the same caution collapse_insearch was held to before its homemaker-py-1ph larger-N confirmation. Kept default OFF (enable_bridge_circulation=False in driver.search/search_staged, --no-bridge-circulation in evolve.py). Candidate follow-ups, not yet filed: (a) a larger-N seed sweep (the 1ph protocol) to resolve whether the mean improvement is real; (b) raising bridge_circulation's _MUTATION_WEIGHTS entry above the uniform default (mirroring place_missing's 2.0) so it fires more often per budget, since a not connected fail is exactly as fatal to fitness as a missing space and the operator is currently drawn no more eagerly than cosmetic ops like rotate. See §22 for the larger-N confirmation of both follow-ups — result: null, weight change reverted, default stays OFF.

22. bridge_circulation larger-N + weight confirmation (homemaker-py-lj3/homemaker-py-qjg) — DONE (null)

Motivation. §21's two identified follow-ups — (a) a 1ph-style larger-N seed sweep to resolve whether 8sh's small positive mean-fail improvement was real or small-sample noise, and (b) raising bridge_circulation's _MUTATION_WEIGHTS entry to 2.0 (matching place_missing) so it fires more often — were filed as separate beads (lj3 for the weight, qjg for the sample size) but lj3's own description flagged them as confounded if tested separately: weight and sample-size are different variables, and a real effect from raising the weight could get masked or amplified by the same small-N noise that made §21 inconclusive in the first place. Tested together in one sweep instead of two.

Protocol. driver._MUTATION_WEIGHTS["bridge_circulation"] = 2.0 (matching place_missing, still zeroed via mutation_weights unless enable_bridge_circulation is set — no behaviour change for the default-off path). Same qpk/1ph protocol as §20/§21: equal-budget ON vs OFF, both arms finished with the standard finish-time --collapse (94g), 4 workers, canonical homemaker-fitness re-score for the .fails breakdown, experiments/run_lj3_qjg_ab.sh. Matching 1ph's own 4× scale-up: programme-house (init.dom, budget 3000) 20 seeds (1–20, vs §21's 5), harbor-house (init.dom, budget 2500) 12 seeds (1–12, vs §21's 3).

A/B verdict (measured, 2026-07-25) — NULL, opposite of §21's directional signal.

• programme-house (N=20): mean fails 7.10 (OFF) → 6.95 (ON), mean per-seed diff 0.15 fails. 6 wins / 5 losses / 9 ties for ON. Paired t-test on the 20 diffs: t=0.38, df=19, two-tailed p≈0.71 — indistinguishable from zero.
• harbor-house (N=12): mean fails 71.8 (OFF) → 72.3 (ON), mean per-seed diff −0.5 fails (ON slightly worse on average). 4 wins / 4 losses / 4 ties. Paired t-test: t=−0.41, df=11, two-tailed p≈0.69 — also indistinguishable from zero.
• Connectivity-specific effect, and the concerning part: of the 10 programme-house seeds whose OFF baseline had a genuine not connected fail, 4 cleared it on ON (seeds 3, 4, 5, 6) — but 3 new not connected fails appeared on seeds whose OFF baseline had none (seeds 1, 9, 13), a higher new-fail rate than §21's original uniform-weight sweep saw (0/5 programme-house seeds introduced a new not-connected fail at N=5; here 3/20 = 15% did at the raised weight). harbor-house cleared 1/10 and introduced 0 new, but its total-fail mean still went the wrong way. mutate_bridge_circulation still only ever converts a leaf to circulation, never away — the new fails are §21's trajectory-divergence mechanism (a nonzero-weight operator entry perturbs the RNG draw sequence for every subsequent mutation, not just its own draws), and raising the weight increases how often that perturbation-inducing draw happens, which plausibly explains why the new-fail rate went up rather than down.
• Cost: essentially unchanged from §21 — programme-house 28.1s (OFF) → 28.1s (ON, 1.00×), harbor-house 130.0s (OFF) → 132.6s (ON, 1.02×).

Interpretation. §21's 4-win/4-tie/0-loss, 2/5-not-connected-cleared result at N=3/N=5 was small- sample noise around a true near-zero effect, not a genuine small positive — the same question 1ph asked of collapse_insearch (§20), but here the larger-N answer goes the other way: not confirmed. Raising the mutation weight did not help and, if anything, correlates with a worse trajectory-noise profile (more new not-connected fails per seed) than leaving it at the uniform default, consistent with the weight bump increasing how often the RNG-perturbing draw fires.

Status. _MUTATION_WEIGHTS["bridge_circulation"] = 2.0 reverted — back to the implicit uniform weight (not present in _MUTATION_WEIGHTS), matching pre-lj3 behaviour exactly. enable_bridge_circulation stays default OFF. No further weight/sample-size follow-up planned; operators.mutate_bridge_circulation remains available opt-in (--bridge-circulation/HOMEMAKER_BRIDGE_CIRCULATION) for anyone who wants the connectivity- targeting behaviour despite the neutral aggregate measurement, but is not a candidate for a default flip on the current evidence.

23. Ruin-and-recreate LNS: rebuild a wing with the adjacency-aware constructor (homemaker-py-f1d) — DONE (positive, size-dependent)

Motivation. DESIGN.md's own experiment log by this point is one-sided: every "search machinery" change tried (§11.5 niching+restarts, §11.4 graded objective, §12.3 Wong-Liu reassociation + shape-feasibility, §12.4 granularity, §14 island model, §16 grain annealing, §18 graded connectivity, §19 shape repair, §21/§22 circulation-repair ops) has come back null-to-negative, while construction/seeding QUALITY (§11.6/§11.7 adjacency-aware seeding, §12.2 proportion-aware seeding) is the only lever that has ever moved the fail count. operators._assign_adjacency_aware — the constructor behind both constructive_topology and lift_base_to_storeys — currently only ever runs once, at seeding. The proposal: reuse it repeatedly DURING search as a large-neighbourhood- search (LNS) ruin-and-recreate move, betting that the one technique with a real track record generalises better than another new comparator-key or population-management idea.

Mechanism (build). operators.mutate_ruin_recreate: pick a divided, live-cut subtree ("wing") of one storey holding a genuine partial neighbourhood of that storey's leaves (>=2, <= half — not a single-leaf relabel already covered by retype/swap, not a whole-floor rebuild already covered by the initial seed), un-divide it back to one leaf, then regrow and retype it with _assign_adjacency_aware, seeded (fixed_circ) from whichever already-typed circulation leaves border the wing — the same mechanism lift_base_to_storeys uses to grow an upper storey off an inherited core (§11.7), so the rebuilt interior spine reconnects to the surviving one instead of growing a disconnected island. The wing's required-space room-code budget is preserved exactly (same multiset); only its internal circulation/outside counts and split are rebuilt, at the same circ_divisor=3/outside_divisor=3 ratio the constructive seeders default to (not threaded from the run config — kept parameter-light, like bridge_circulation).

_assign_adjacency_aware gained a new scope parameter (leaves eligible for retyping; fixed_circ may then name border leaves OUTSIDE scope as dominating-set seeds only, never retyped) so the wing rebuild can share the exact constructor code without touching the rest of the storey. scope=None (every existing caller) reproduces the prior unrestricted behaviour exactly — verified no other caller's output changed. Gated like reassociate/bridge_circulation: zero mutation weight unless enable_ruin_recreate=True (driver.search/search_staged, evolve.py --ruin-recreate/HOMEMAKER_RUIN_RECREATE, default off).

Verified (build-time). 200 applications of mutate_ruin_recreate chained onto fresh constructive_topology harbor-house seeds (40 seeds × 5 steps): zero missing-space regressions (graph.check_space_counts), every child a canonical genome (encode(decode(encode(x))) == encode(x)). 297 existing tests pass unchanged (the new op is exercised by the existing test_mutations_yield_canonical_genomes parametrization, which calls it with reqs=None and gets the documented noop). A child_probe-instrumented driver.search run confirmed the operator is actually selected by mutate() at its configured weight (not dead code).

Initial A/B (measured, 2026-07-25/26, qpk protocol) — NULL, but underpowered. Equal-budget enable_ruin_recreate ON (implicit uniform mutation weight, ~7.5% draw probability among ~13 active ops) vs OFF, both arms finished with the standard finish-time --collapse (94g), 4 workers:

• harbor-house (budget 2500, seeds 1–3): 1 loss (74→81), 2 ties.
• programme-house (budget 3000, seeds 1–5): 4 ties, 1 win (9→8).
• Combined: 1 win / 1 loss / 6 ties out of 8, mean fails 31.9 (OFF) → 32.6 (ON) — indistinguishable from zero, in the same direction as most of this log's other null results.
• A direct child_probe instrumentation of one of the tied harbor-house runs found ruin_recreate fired once in 32 children — the initial sample is dominated by trajectories where the operator simply never got a turn, not by turns it lost. Six of the eight exact ties (fitness scalar identical to 6 significant figures, not just fail count) are consistent with this: the op's rare draws mostly didn't survive tournament selection into the recorded lineage.

Weight follow-up (measured, 2026-07-26) — reran the ON arm only with _MUTATION_WEIGHTS["ruin_recreate"] = 3.0 (matching place_missing, mirroring the lj3 weight-bump precedent) at the same seeds/budgets, directly comparable to the existing OFF baseline:

• programme-house (seeds 1–5): 4 wins, 1 tie, 0 losses — 7→1, 9→7, 9→8, 9→7, 5→5. A striking, one-sided result, including one seed dropping from 7 fails to 1 (verified deterministic on rerun).
• harbor-house (seeds 1–3): 1 win (77→73), 1 loss (74→82), 1 tie — still mixed.

Larger-N confirmation (measured, 2026-07-26) — extended both arms to 10 fresh programme-house seeds (6–15) and 5 fresh harbor-house seeds (4–8) at the same weight=3.0, same protocol:

• programme-house, all 15 seeds combined: 8 wins / 1 loss / 6 ties. Mean fails 7.07 (OFF) → 6.00 (ON), a ~15% reduction. Wilcoxon signed-rank p≈0.041; sign-test p≈0.020 (one-sided) — holds up at conventional significance, not small-sample noise around zero (the 8sh/1ph/qi6/lj3 pattern this log warns about).
• harbor-house, all 8 seeds combined: 3 wins / 2 losses / 3 ties. Mean fails 73.0 (OFF) → 74.5 (ON) — no consistent effect, if anything a very slight negative lean, echoing §20's (collapse_insearch) opposite-direction size split but with the SMALLER building this time as the one that benefits.

Interpretation. A rare case in this log where a search-machinery idea shows a real, statistically-supported effect — but only on the smaller/simpler example programme. Plausible reading: programme-house's smaller room count means a wing rebuild samples a much larger fraction of the whole floor's topology per move (higher effective locality-vs-scope ratio), so the constructor's proven adjacency-aware placement quality dominates; harbor-house's much larger room count means the same wing size is a small, noisier perturbation relative to the whole building, and correlates with the ~2× per-op cost of _assign_adjacency_aware (leaf-graph rebuild + dominating-set search) not translating into more useful search steps within the same eval budget on that scale.

Status (2026-07-26). enable_ruin_recreate stays default OFF — harbor-house shows no benefit and the two example programmes disagree on direction, so flipping the global default is not supported by this evidence (same conservative bar §20 applied before its own larger-N confirmation). _MUTATION_WEIGHTS["ruin_recreate"] = 3.0 is kept in the source (only takes effect when the flag is on) since it is the validated-effective setting. --ruin-recreate/HOMEMAKER_RUIN_RECREATE is documented and ready to use today on programme-house-scale (smaller/simpler) programmes; a natural follow-up (not filed, low priority) would be a third or fourth example programme at a size between the two tested here, to locate the size threshold this result implies rather than inferring it from just two data points.

24. Ruin-and-recreate size-threshold sweep (homemaker-py-y51) — INCONCLUSIVE, no clean threshold

Motivation. §23's follow-up: locate the room-count threshold where ruin_recreate (weight=3.0) stops helping, rather than inferring it from programme-house (6 rooms, win) vs harbor-house (37 room instances, null/slight-negative) alone.

No natural third example exists. programme-house2 is the same 6-room size as programme-house (a geometry-fix variant, not a size variant); maple-court (26 space types, many with count, ~more room instances than harbor-house) is bigger than harbor-house, not between the two. So this used option (b) from §23: a synthetic room-count sweep on programme-house's own patterns.config, scaling the b1/t1/b2/t2 bedroom+ensuite module count by an integer factor (k=2..5 → 10/14/18/22 room instances), holding room-type mix, storey limits, ratios and adjacency constant, with the footprint (init.dom) scaled in area to match (examples/y51-sweep-{10,14,18,22}). budget=3000 calibrated so every size leaves nontrivial residual fails at seed 1 (14/29/27/43), not saturated to 0.

Measured (2026-07-26, experiments/run_y51_sweep.sh + run_y51_sweep_ln.sh) — paired seeds, --ruin-recreate (weight=3.0) ON vs OFF, both arms finish with the default --collapse (94g), 4 workers. Initial pass: 5 seeds at every size. Larger-N confirmation: 5 more seeds (N=10 total) at the two sizes whose initial 5-seed read was most striking (n=14, the only size that initially lost; n=18, the strongest initial win) — mirroring this log's own larger-N-confirmation pattern.

n_rooms	N	W/L/T	mean fails OFF→ON	Δ%	Wilcoxon p
10	5	3W/1L/1T	17.80 → 16.40	+7.9%	0.625
14	10	3W/5L/2T	28.90 → 28.70	+0.7%	0.945
18	10	7W/2L/1T	36.50 → 33.10	+9.3%	0.098
22	5	2W/3L/0T	41.20 → 40.80	+1.0%	0.875

Interpretation. This does not reproduce a clean monotonic decay of the effect as room count rises from programme-house's 6 to harbor-house's 37. n=14 came back a clean null after larger-N confirmation (the initial 5-seed 0W/4L read did not hold — noise, exactly the pattern this log repeatedly warns about). n=18 shows the strongest trend of the four synthetic sizes (a plausible-but- not-quite-significant ~9% mean improvement, p≈0.10) despite sitting between two much weaker/null sizes (14 and 22) — a non-monotonic bounce inconsistent with a simple "smaller wing-rebuild-to-floor ratio → bigger effect" threshold as a function of room count alone.

Two readings, not mutually exclusive:

1. Still underpowered. §23's own programme-house confirmation needed N=15 seeds to reach p=0.041 for a similar-magnitude effect (~15% reduction); N=5/N=10 here is likely too little to resolve an effect this size cleanly at any of these sizes, so the bounce may just be sampling noise on top of a real but weak trend across 10-22 rooms.
2. Methodological caveat: this sweep is not a clean proxy for "room count." It scales room count by duplicating already-anonymous, already-interchangeable room codes (count: on b1/t1/b2/t2) — the same mechanism harbor-house itself uses "to reduce complexity" (its own patterns.config comment). Duplicating interchangeable codes may make placement systematically easier for _assign_adjacency_aware than harbor-house's mix of many genuinely distinct room types at the same instance count would be, so this sweep's room-count axis may not isolate the same "topology fraction sampled per wing move" variable that §23 hypothesised drives the effect.

Status (2026-07-26). enable_ruin_recreate stays default OFF; no per-size default flip is supported by this evidence — the sweep did not locate a clean threshold. §23's practical guidance (safe to opt in on programme-house-scale, ~6-room programmes; not validated at harbor-house scale) stands unchanged. A real follow-up, if pursued, needs either (a) much larger N (~15+ seeds) at a smaller set of sizes to resolve whether the n=18 trend is real, or (b) a genuinely distinct third example programme (real room-type diversity at an intermediate room count, not a duplicated-code sweep on programme-house) to avoid the interchangeable-room confound above.

25. 2-opt local search past the collapse_global Jacobi plateau (homemaker-py-9wi) — DONE (positive, opt-in)

Motivation. §17's collapse_global adjacency relaxation is a Jacobi/WFC-style loop: each round re-solves a linear assignment (_best_assignment) using an adjacency bonus computed from the previous round's neighbour labels. That is exact per round, but the true objective is quadratic — a satisfied adjacency depends on a pair of labels, not one — so synchronous Jacobi can plateau short of the joint optimum. Worked example (test_two_opt_polish_escapes_jacobi_plateau): a 4-cell chain p1─q1─p2─q2 with two disjoint adjacency requirements (p1<->p2, q1<->q2) has a fully-satisfying relabelling (p1─p2─q1─q2 or similar), but starting from the interleaved layout the Jacobi loop 2-cycles between two labellings that each satisfy zero of the four requirements, and never escapes within iters.

Mechanism (Fitness._two_opt_adjacency_polish). Runs once, after the Jacobi loop reaches its fixpoint (or exhausts iters). For every same-level pair of supply leaves, try swapping their current labels; keep the swap only if it strictly increases the total reward (own quality/threshold value + fail_w per satisfied adjacency) summed over the two leaves and every leaf adjacent to either — the only cells a label swap between i and j can change. Repeats to a fixpoint (or local_search_passes, default 20). Same-level-only pairing keeps the hard level constraint for free (both codes already matched their own leaf's level pre-swap, and the two leaves share a level, so the swap is valid on both sides). A swap is applied only on strict improvement, so this is monotone by construction — it can only reduce, never increase, the objective's implied fail count, same guarantee as the Hungarian solve it refines. Fitness._collapse_value factors the shared (leaf, code) → base-value computation out of the collapse_global assignment-matrix build so both the matrix and the polish score a pair identically.

Why 2-opt over CP-SAT/OR-Tools. The issue proposed either a 2-opt local search or a CP-SAT (OR-Tools) encoding of the labelling QAP. Went with 2-opt: no new dependency (the project has no ortools), and it extends the existing Jacobi machinery directly rather than replacing it with a separate solver. QAP is NP-hard in general, so this is a local search, not an exact solve — but it strictly dominates the Jacobi-only result by construction, and collapse_finish's keep-better wrapper is an additional safety net regardless.

Wiring. collapse_global(local_search=False, local_search_passes=20) — the method-level default stays off (see §28: it's also called every fitness eval via collapse_insearch/qpk, a hot path this polish was never measured against). homemaker-collapse --local-search/--no-local-search and evolve.py's --collapse now default it on at the one-shot finish-time call sites — see §28 (homemaker-py-cdl) for the broader sweep and wiring that flipped those defaults.

Verification. Swept all 11 harbor-house evolved-*/3m/materialised-3M .dom files, comparing collapse_global(local_search=False) against local_search=True: 10/11 matched exactly (Jacobi was already at the 2-opt-local optimum on those layouts), 0 regressed, 1 improved (evolved-anneal-3M.dom 21→19 fails — resolved a genuine mutual da1<->k1 adjacency miss the Jacobi loop couldn't reach). Runtime <1s even on the largest file (evolved-3M.dom, 90 base fails). §28 extends this to a 46-file sweep and turns the finish-time default on. Tests: tests/test_collapse_global.py gains test_two_opt_polish_escapes_jacobi_plateau (7 total in that file); 298/298 pass project-wide.

26. Multi-use leaves / type superposition (homemaker-py-9o5/xi7/b3v) — DONE (negative), backfilled

Closed 2026-06-30 (9o5, xi7) / 2026-07-17 (b3v); written up retroactively — this section was missing when §17/§20 above were written, even though both reference its verdict directly ("mirrors 9o5", "the opposite of the 9o5/xi7 verdict"). Numbered at the end of the log rather than renumbering §14-§25 to preserve every existing cross-reference.

Motivation. A leaf that legitimately serves several DIFFERENT compatible programme codes at once (study+guest bedroom, kitchen+dining — Stewart Brand's "loose-fit" long-life rooms), distinct from §13.3 leaf-sharing which aggregates k instances of the same code. Two readings were scoped: (a) superposition as a SEARCH RELAXATION — carry an uncommitted set of candidate types per leaf during search, collapse (argmax re-type) to specific usages only at scoring time, for a smoother landscape; (b) multi-use as the permanent DESIGN GOAL, surviving into the output with no collapse. Path (a) was built and validated (below); path (b) was never started.

Mechanism (path a, built). programme.derive_interchange_classes: codes form an equivalence class (connected component, size ≥ 2) under a symmetric interchangeable() relation — S1 both sized and non-generic (no c/o/s), S2 size/width/proportion targets within LOCKED ratio bounds (R_SIZE=1.5, R_WIDTH=1.3, R_PROP=1.5), S3 compatible level and service stack, S4 no direct required-adjacency edge between the two codes (adjacency pairs are coexisting rooms, not one substitutable leaf). Pure function of the parsed programme — classes are auto-derived, no hand-authored list needed on the happy path. Fitness.collapse_superposition re-types every superposed leaf to its best in-class usage each eval, before any check: per class, an optimal supply (leaves currently in the class) → demand (class codes × required count) matching, area-weighted usage quality as the objective (brute-force ≤CLASS_CAP = 4! permutations, else scipy Hungarian — the same _best_assignment §17/§25 later reuse at global scope). Runs on the UNMERGED tree, so counts/adjacency/quality downstream see the condensed types with no changes needed to graph.py/dom.py/operators.py — the key design realisation was that because collapse re-types at eval time, Node never needs a persisted class/serves field and no mutation operator needs a "retype within class" move; the genome can carry any in-class type and collapse fixes it. Gated behind superpose (default OFF, bit-identical when off — verified against 233 pre-existing tests). tests/test_superposition.py (20): derivation (service/adjacency/level guards, the real programme-house programme), assignment (brute force + Hungarian + surplus supply/demand), end-to-end collapse re-typing, veto-hatch behaviour.

A/B verdict (xi7, measured 2026-06-30) — NULL/NEGATIVE. Equal-budget --superpose ON vs OFF, measuring the COLLAPSED (final) score:

• programme-house (init.dom, budget 3000, 4 workers, seeds 1–5): OFF wins 4/5 (s1 8f>10f, s2 11f>12f, s3 10f>12f, s4 10f>10f-tied-fitness — all OFF strictly better or equal fails), ON wins only s5 (10f→8f).
• harbor-house (init.dom, budget 2500, seeds 1–3): OFF wins 2/3 (s2 33f<38f, s3 43f<48f); ON wins s1 alone (50f<51f).
• Superposition does not reach better layouts; in most seeds ON has ≥ OFF fails — the per-eval collapse re-typing perturbs counts/adjacency rather than smoothing the search, the same failure mode later sections would call "landscape flattening."

Relaxation-gap instrumentation (xi7 §7.4) — ruled OUT as the cause. Logged relaxed (unconstrained best-case usage-quality) vs collapsed value on the same matched leaves across the 5 programme-house ON runs: total gap_ratio 1.01–1.23 (per-class peaks up to 1.52) — small-to-moderate, not the large gap the original risk note feared. Because collapse is per-eval, there is no separate relaxed phase to diverge from — search already optimises the collapsed objective by construction. Conclusion: path (a) underperforms not from a relaxation gap but because the geometry floor (§11–§13) dominates — type labels are not the binding constraint on these programmes, so easing them buys nothing while the re-typing adds feasibility noise. This is the diagnosis §20 (qpk) later cites when arguing its own in-search collapse is a different mechanism (a hard-constraint-respecting global relabel, not a per-class relaxation over interchangeable-but-not-identical codes) and so isn't pre-falsified by this verdict.

Veto hatch (b3v, closed 2026-07-17) — the one real false-positive found. Harbor-house's programme auto-derives a transitive 8-code chain {da1,ef1,k1,la1,m,me1,n,ws1} spanning a 6× size range (Meeting 10 m² .. Dining/Neighbourhood 60 m²) — semantically nonsensical (Meeting↔Dining↔Kitchen↔ Mechanical are not interchangeable) but sanctioned by the S1–S4 relation as written (each adjacent pair in the chain individually satisfies the ratio bounds; connectivity is transitive). xi7's harbor-house losses show ON adding fails in both loss seeds (38→33 became 38 vs 33; 48→43 became 48 vs 43) — consistent with this misgroup actively hurting. Fix: SpaceReq.interchange (default True), settable interchange: false per code in patterns.config, honoured by interchangeable()'s S0 check — an architect veto for one code without disabling superposition globally. superpose itself stays default OFF regardless (the xi7 verdict was null/negative overall), so the hatch only matters if/when superposition is deliberately enabled on a real config.

Status. --superpose stays default OFF; path (a) is not recommended without a fundamentally different mechanism (the geometry floor, not the labelling relaxation, is what needs to move — the same conclusion §11–§13's construction-quality work and §19's negative geometry-repair result both reach from other directions). Path (b) (multi-use as a permanent design goal, no collapse) was never attempted — remains open if revisited, but low priority given (a)'s outcome and the project's broader 0-for-several record on search-machinery/fitness-shaping bets vs construction-quality bets (see mi7, §27, for the same pattern one experiment later).

27. 3D bubble-diagram adjacency fitness signal (homemaker-py-mi7) — DONE (negative)

Closed 2026-07-25, the session immediately before §25's 9wi. bubble.py was left in the repo uncommitted as a documented reference per the original close note; committed alongside this write-up so the reference this section makes to it is actually resolvable.

Motivation. graph.py's adjacency checks are binary (is X adjacent to Y, yes/no) and, like §18's connectivity fail, give the search no gradient toward a better overall spatial arrangement — only toward satisfying each declared pair. Idea: build the programme's required-space adjacency as a graph, relax it into a 3D "bubble diagram" (a spring/repulsion physics simulation, architecture's traditional adjacency-diagramming technique), then score a candidate layout by how well its real room-to-room distances correlate with a relaxed target's distances — an additional graded fitness term / search- guidance signal, in the spirit of §18's graded connectivity but for general spatial layout rather than circulation topology specifically.

Mechanism (bubble.py, prototype only, never wired into fitness.py). requirement_graph: one node per required room instance (code, or code#i for count>1), generic c/o/s adjacency targets collapsed to one shared hub node per code (per whole building, not per storey — a known simplification), edges to a multi-count code fan out to all its instances at reduced weight (satisfying adjacency needs only one matching neighbour). generate_targets: relax the requirement graph from n_restarts random 3D starts with a spring force (ideal edge length = sum of target-area-equivalent circle radii) plus overlap-only repulsion plus a level-height pull on the z axis; because relaxation is non-convex and multi-modal (different starts settle on e.g. opposite-handed but equally valid arrangements), keep up to keep distinct low-energy solutions (pairwise-distance-vector correlation ≥ dedup_corr = duplicate) rather than one canonical target. similarity: weighted Pearson correlation between an actual Dom layout's real weighted shortest-path distances and a target bubble's Euclidean distances, over matched non-generic room instances, weighted 1/hop_distance in the requirement graph so the many hub-mediated "just wants to be near circulation" pairs (weak positional evidence) don't drown out the few directly-declared adjacencies (strong evidence). best_similarity takes the max across the kept alternative targets. matched_leaves maps anonymous multi-count codes to actual leaves by a fixed centroid-order rule — flagged in the module docstring as a known simplification, not a real assignment solver. topological_similarity is a cheaper no-embedding alternative: hop-distance correlation directly on graph topology (real multi-cell circulation network on both sides), skipping the physics simulation and multi-restart dedup entirely.

Validation (measured 2026-07-25) — NULL on both formulations, both programmes. Correlated each similarity metric against real evolved trajectories (not static examples) via driver.search:

• programme-house (n=100 recorded individuals): embedding ρ≈0.05, topological ρ≈−0.06 — flat. This is the cleanest data point: programme-house has zero multi-count anonymous codes, so matched_leaves' fixed centroid-order heuristic cannot be confounding the result, and it's still flat.
• harbor-house (budget 6000, n=75, fitness 3e-28→3.9e-17, fails 83→51 over the trajectory): similarity() (embedding) spearman=0.164, p=0.16 (n.s.); topological_similarity() spearman=−0.160, p=0.17 (n.s.) — noisier than programme-house (heavy anonymous-count codes: n×5, m×3, t×6, r×10, of×2, a real uncontrolled confound for the centroid-order matching there) but tells the same story.
• No statistically significant correlation anywhere, across 2 independent formulations (spatial embedding vs pure topology) × 2 programmes, with real search trajectories rather than canned batches.

Status. Do not pursue graph-relaxation-derived or pure-topological adjacency-matching as a fitness signal for this project without a fundamentally different formulation. If revisited, the harbor-house anonymous-code confound would need a real assignment solver (Hungarian/brute-force, mirroring 9o5's CLASS_CAP pattern) before drawing any programme-specific conclusion there — but programme-house's clean, confound-free null already argues against the core idea regardless. bubble.py stays in the repo as a working, documented reference, not wired into fitness.py. Consistent with the project's broader pattern (§11.4/11.5, §12.3/12.4, §14, §16, §21, §22, §26 above): search-machinery / fitness-shaping changes have been null-to-negative essentially every time they've been tried; only construction/seeding quality and representation-relaxation changes (leaf-sharing §13.3, global collapse §17/§25) have moved the needle. This is another data point for that pattern, not an exception.

28. Default the 9wi 2-opt polish on for finish-time collapse (homemaker-py-cdl) — DONE (positive)

Motivation. §25 (homemaker-py-9wi) validated the 2-opt adjacency polish on harbor-house alone (11 files, 1 improvement, 0 regressions) and left it opt-in pending a broader, non-synthetic sweep and the evolve.py/driver.collapse_best wiring to expose it outside the standalone homemaker-collapse CLI. This closes that follow-up.

Broader sweep. Extended the harbor-house comparison to programme-house's 34 .dom files (real evolved candidates, not synthetic), 46 files total across both example sets. Compared collapse_finish(local_search=False) against local_search=True (both keep-better against the uncollapsed base, per §17): 0 regressions, 2 improvements — the known harbor-house evolved-anneal-3M.dom (21→19 fails) plus a new one on programme-house, a82f07068e4408fdd0d5e3dc469a8dee.dom (3→2 fails); every other file matched exactly. Confirms the finding generalises past the single synthetic dataset §25 was validated on.

Where the default did NOT change. collapse_global's own local_search=False default (§25) was left untouched. collapse_global runs twice in this codebase: once as a one-shot finish-time pass (collapse_finish, homemaker-collapse, driver.collapse_best) and once per fitness eval inside _evaluate_full when collapse_insearch/qpk (§20) is on — the latter is the hot path of the entire evolutionary search, run thousands of times per run, and the 46-file sweep only measured the one-shot cost (<1s even on the largest file). Flipping the method-level default would have silently turned the 2-opt pass on inside that hot loop too, an untested and likely-costly change out of scope for this issue. So the default stays False at the method level, and each one-shot call site turns it on explicitly instead.

Wiring. homemaker-collapse --local-search/--no-local-search (collapse_cmd.py) now defaults True (was False). Added homemaker-evolve --collapse-local-search/--no-collapse-local-search (evolve.py), default True, passed through to driver.collapse_best(..., local_search=...) — which already forwarded arbitrary **collapse_kw to fit.collapse_finish, so no signature change was needed there. The new flag is a no-op under --no-collapse (nothing to polish if the finish-time collapse itself is skipped).

Verification. 298/298 tests pass (no test changes needed — test_collapse_global.py's explicit local_search=True/False cases already covered both method-level defaults). Re-ran homemaker-collapse standalone on evolved-anneal-3M.dom with no flags to confirm the new CLI default reproduces the 19-fail result end-to-end.

29. Beam/best-first search over adjacency-aware room placement (homemaker-py-c94) — DONE (inconclusive, mixed on harbor-house, null on programme-house)

Motivation. Construction/seeding quality is the one lever with a consistent positive track record (§11.6/§11.7 adjacency-aware seeding, §12.2 proportion-aware seeding, §23 f1d's reuse of the same constructor mid-search). operators._assign_adjacency_aware places rooms onto the circulation-dominated leaf set with a single greedy pass: hardest-constrained code first, each dropped onto whichever open slot currently satisfies the most of its declared secondary adjacency (beyond c) against already-placed neighbours. Because the pass never revisits a placement, an early code with no already-typed neighbours to match against (every mutual pair's first-placed half, e.g. harbor-house's k1↔da1) picks blind — any open slot scores identically at that step — and an unlucky tie-break could strand it from a partner that would only be placed several steps later. The proposal: explore the same per-room slot decisions with a width-K beam/best-first search instead of one irrevocable pass, scored by a cheap proxy (no geometry/fitness calls), and measure whether it ever finds a genuinely better seed before considering investing further (e.g. wiring it into the outer search config).

Mechanism (build). _assign_adjacency_aware gained a beam_width: int = 1 parameter (operators.py); beam_width<=1 (default) is byte-identical to the prior greedy code path — verified by test_construction_beam_width_default_matches_greedy and by the full 298-test suite passing unchanged before any beam-specific test was added (302/302 after adding four new beam-specific tests). beam_width>1 instead routes room placement through the new _beam_place_rooms: keeps up to beam_width partial placements alive, each step branching every surviving state into its top-beam_width candidate slots for the current code (same ranking greedy uses), scored by the running total of secondary-adjacency matches satisfied so far. This is genuinely cheap — no geometry or fitness calls, since circulation/outside are already fixed before room placement starts and the leaf-adjacency graph (_nbrs, deg, idx, dominated) is shared read-only across every branch — then prunes back to beam_width states before the next code, returning the highest-scoring complete placement. Threaded through as construction_beam_width in constructive_topology, lift_base_to_storeys, driver.search, and driver.search_staged (all default 1, matching the project's existing knob-threading convention for circ_divisor/depth_balanced/etc. — no CLI flag added, consistent with those other construction-only knobs). Not threaded into mutate_ruin_recreate (kept parameter-light, like bridge_circulation/ruin_recreate's own circ/outside ratios, §23).

Verified functioning (synthetic, not a no-op). A hand-built adversarial 4-slot graph (two disjoint adjacent pairs, codes a↔b mutually required plus a filler x placed between them) confirms the mechanism is real: beam_width=1 places a by an arbitrary tie-break, x then greedily grabs a's only neighbour before b gets a turn, stranding the pair (a-b adjacent=False); beam_width>=2 recovers the correct joint placement (a-b adjacent=True) by keeping a's alternate slot choice alive long enough for b's later score to reward it. This is exactly the "no lookahead" failure mode _assign_adjacency_aware's one-shot pass is structurally prone to, and confirms the beam can and does out-score greedy when the graph offers a genuine trade-off.

Raw-seed check (2026-07-27) — misleadingly byte-identical, later shown insufficient. Before running any search, a cheap diagnostic scored constructive_topology's raw output directly (no GA, one score_with_fails call per seed): beam_width 1/4/8, 15 rng trials each, on programme-house and harbor-house — fail counts and adjacency/access fail counts identical to the last digit across all three widths, every trial. Extended to lift_base_to_storeys (the Stage-2 seeder) at widths 1/4/8/20, 10 trials: again byte-identical at every width, including beam_width=20 (near-exhaustive for the ~15-17 codes per storey these programmes carry). A step-by-step trace of a real harbor-house construction (da1→k1→ws1) confirmed the beam does explore physically distinct slot branches, but every branch reached the same cumulative score every time — harbor-house's circulation-spine geometry usually offers several equally-good neighbours per code, so a lone raw-seed sample rarely hits a real trade-off. This was wrongly taken as proof an end-to-end run would also be byte-identical (a single root's construction never diverging was treated as sufficient to conclude the full bootstrap population never would either) — see the correction below.

End-to-end correction (2026-07-28, prompted by user question "should the default be 1? can we find out by running the two example programmes from a clean start?") — the raw-seed argument was wrong. Ran driver.search from a clean bootstrap (init.dom, n_workers=1 for reproducibility, budget 1500) at construction_beam_width 1 vs 4, same seed both arms, 5 seeds each programme:

programme	seed	bw=1 fails	bw=4 fails	result
harbor-house	1	60	58	bw4 win
harbor-house	2	67	52	bw4 win (large)
harbor-house	3	53	53	tie
harbor-house	4	52	52	tie
harbor-house	5	52	62	bw4 loss
programme-house	1–5	(9,6,11,11,9)	identical	tie, all 5

harbor-house: 2 wins / 1 loss / 2 ties, mean fails 56.8 (bw1) → 55.4 (bw4) — a small mean improvement pulled mostly by seed 2's outlier, with a real loss on seed 5. programme-house: 5/5 ties, matching the raw-seed prediction exactly. The harbor-house divergence itself confirms the raw-seed reasoning's flaw: driver.search's bootstrap builds pop_size individuals, each consuming a different slice of the RNG stream (unlike the single-root raw-seed check), and once even one population member's construction hits a genuine beam-vs-greedy tie-break divergence, the GA's subsequent structure-dependent choices (which subtree a mutation targets, crossover points) cascade into a different trajectory from there — even though the raw RNG numbers drawn are bit-identical between arms. "The one seed I checked never diverged" does not imply "no seed in a population of many ever will."

Interpretation. The mechanism works (§ above, verified on a synthetic graph built to need it), and does occasionally get real traction on harbor-house's larger, more room-dense programme — but the 5-seed result is the same small-N, mixed-direction shape this log has repeatedly warned produces false signal (§23 f1d's initial 8-run sweep, explicitly flagged there as "the 8sh/1ph/qi6/lj3 pattern"): a genuine loss (seed 5) sits alongside the two wins, and N=5 is far short of what f1d's own larger-N confirmation needed (N=15/8) to separate a real effect from noise. programme-house shows no effect at any N tested, consistent with both the raw-seed check and its smaller, simpler room graph.

Status. construction_beam_width stays default 1 — the direct answer to "should the default be 1": yes, current evidence does not clear this project's bar for flipping a default (cf. §20/§23's own "only after larger-N confirmation" standard), though harbor-house's mixed result (unlike programme-house's clean tie) means this is genuinely unresolved rather than a confident null. The code and tests stay in the tree as a working, verified-functioning building block (operators._beam_place_rooms), consistent with keeping validated-but-inconclusive mechanisms available rather than reverting them (cf. bubble.py, §27). A natural follow-up — not filed, low priority, matching f1d's own unfiled size-threshold follow-up (§23/§24) — would be a larger-N harbor-house-only sweep (N=15+, matching f1d's and y51's bar) to determine whether the mean-improvement lean is real or an artefact of seed 2's outlier.