bruno/homemaker-layout
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

qpk: in-search global collapse — run collapse_global per-eval during search

Browse source 
Runs the 94g finish-time cell↔room collapse inside every fitness eval
(collapse_insearch conf flag, default off, bit-identical when off) instead
of once at the end, so search optimises the collapsed objective directly.
Plumbed through fitness.py/driver.py/evolve.py the same way superpose/
conn_grade are; --collapse-insearch CLI flag.

A/B validated against the xi7 protocol (equal budget, both arms finished
with standard finish-time --collapse): POSITIVE, opposite of the 9o5/xi7
prior. harbor-house ON wins 3/3 (mean fails 80.3->72.0); programme-house
mixed 3/5 (mean fails 8.4->7.8). Kept default off pending a larger
programme-house sample; documented as a working opt-in for harbor-house-
scale-or-larger programmes. Full writeup in DESIGN.md §20.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
Bruno Postle 2026-07-19 20:35:18 +01:00
parent 07a4739576
commit 2b7a7d2926
6 changed files with 281 additions and 28 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

32
.beads/issues.jsonl
View file

@ -56,7 +56,7 @@
{"id":"homemaker-py-k2g","title":"Topology genome: base-floor tree + per-floor deltas + type assignment","description":"DESIGN.md §5.2, §7 Phase 2. Genome = base-floor slicing topology (primary) + per-leaf type assignment + per-floor divide/undivide deltas (Below-inheritance as regulariser; cut owned by lowest storey where its path is divided — §10). Must round-trip to/from dom.py Node trees so the oracle and inner loop consume it directly. Includes storey count and per-floor type overrides.","acceptance_criteria":"Genome \u003c-\u003e .dom round-trip on all 35 corpus files preserves fitness; multi-storey wall stacking preserved","status":"closed","priority":2,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-06-11T23:37:26Z","created_by":"Bruno Postle","updated_at":"2026-06-12T12:52:34Z","started_at":"2026-06-12T10:55:21Z","closed_at":"2026-06-12T12:52:34Z","close_reason":"genome.py encode/decode lands. 35/35 oracle fitness parity after round-trip (flag-on); genome fixed-point + owned-projection tests. Dead-field discovery: corpus upper storeys carry drifted dead divisions (97) and rotations (187) — canonicalised by decode, validated fitness-neutral.","dependency_count":0,"dependent_count":1,"comment_count":0} {"id":"homemaker-py-k2g","title":"Topology genome: base-floor tree + per-floor deltas + type assignment","description":"DESIGN.md §5.2, §7 Phase 2. Genome = base-floor slicing topology (primary) + per-leaf type assignment + per-floor divide/undivide deltas (Below-inheritance as regulariser; cut owned by lowest storey where its path is divided — §10). Must round-trip to/from dom.py Node trees so the oracle and inner loop consume it directly. Includes storey count and per-floor type overrides.","acceptance_criteria":"Genome \u003c-\u003e .dom round-trip on all 35 corpus files preserves fitness; multi-storey wall stacking preserved","status":"closed","priority":2,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-06-11T23:37:26Z","created_by":"Bruno Postle","updated_at":"2026-06-12T12:52:34Z","started_at":"2026-06-12T10:55:21Z","closed_at":"2026-06-12T12:52:34Z","close_reason":"genome.py encode/decode lands. 35/35 oracle fitness parity after round-trip (flag-on); genome fixed-point + owned-projection tests. Dead-field discovery: corpus upper storeys carry drifted dead divisions (97) and rotations (187) — canonicalised by decode, validated fitness-neutral.","dependency_count":0,"dependent_count":1,"comment_count":0}
{"id":"homemaker-py-d0s","title":"Experiment: inner-loop optimiser bake-off at equal oracle budgets","description":"DESIGN.md §7 Phase 1, §8.3. DOF is only ~rooms-1 (6–7 on corpus). Compare Nelder-Mead vs CMA-ES vs batched multi-start pattern search at equal oracle-call budgets, measuring fitness gained per oracle call and wall-clock (batch-friendliness matters — §4.6). Measure, don't commit blind.","acceptance_criteria":"Table of fitness-per-budget across \u003e=3 candidates; one optimiser chosen and recorded in DESIGN.md","status":"closed","priority":2,"issue_type":"task","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-06-11T23:36:59Z","created_by":"Bruno Postle","updated_at":"2026-06-13T08:48:13Z","started_at":"2026-06-12T21:22:15Z","closed_at":"2026-06-13T08:48:13Z","close_reason":"Bake-off complete: CMA-ES confirmed as Phase 1/2 optimiser. NM wins quality per eval but sequential architecture incompatible with batching (§4.6). Compass stalls on narrow valleys. Results in DESIGN.md §8.3 and experiments/bakeoff_innerloop.*","dependencies":[{"issue_id":"homemaker-py-d0s","depends_on_id":"homemaker-py-1p0","type":"blocks","created_at":"2026-06-12T00:39:35Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":0,"comment_count":0} {"id":"homemaker-py-d0s","title":"Experiment: inner-loop optimiser bake-off at equal oracle budgets","description":"DESIGN.md §7 Phase 1, §8.3. DOF is only ~rooms-1 (6–7 on corpus). Compare Nelder-Mead vs CMA-ES vs batched multi-start pattern search at equal oracle-call budgets, measuring fitness gained per oracle call and wall-clock (batch-friendliness matters — §4.6). Measure, don't commit blind.","acceptance_criteria":"Table of fitness-per-budget across \u003e=3 candidates; one optimiser chosen and recorded in DESIGN.md","status":"closed","priority":2,"issue_type":"task","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-06-11T23:36:59Z","created_by":"Bruno Postle","updated_at":"2026-06-13T08:48:13Z","started_at":"2026-06-12T21:22:15Z","closed_at":"2026-06-13T08:48:13Z","close_reason":"Bake-off complete: CMA-ES confirmed as Phase 1/2 optimiser. NM wins quality per eval but sequential architecture incompatible with batching (§4.6). Compass stalls on narrow valleys. Results in DESIGN.md §8.3 and experiments/bakeoff_innerloop.*","dependencies":[{"issue_id":"homemaker-py-d0s","depends_on_id":"homemaker-py-1p0","type":"blocks","created_at":"2026-06-12T00:39:35Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":0,"comment_count":0}
{"id":"homemaker-py-161","title":"In-search evaluation of shape_rotate/deslim GA operators (7fm follow-up)","description":"homemaker-py-7fm's finish-time hill-climb found zero improving moves for\nmutate_shape_rotate/mutate_deslim (operators.py) on the 6-layout harbor-house\nsweep — every candidate move traded a shape fail for a new adjacency/access\nfail on the already co-evolved layout. That's a different regime from\nin-search use: a full multi-generation GA run gives selection pressure and\npopulation diversity a chance to accept a locally-worse move that a later\nstep or recombination completes.\n\nTo test: thread `fit` through driver.search (currently only reqs is passed to\noperators.mutate; shape_rotate/deslim need `fit` and currently no-op inside\nthe GA). Gate with an enable_shape_repair-style flag, mirroring how\nenable_reassociate (§12.3) let 9gp.2 do a clean A/B. Run full-budget\nharbor-house search with/without across seeds, compare final fail counts.\n\nIf negative again, the geometry-intrinsic residual on harbor-house-scale\nprogrammes may be a genuine floor for this representation, not a repairable\ninefficiency (consistent with §17/§19's framing). See DESIGN.md §19 and bd\nmemory collapse-global-94g-and-any-label-usage-optimisation for full context.","status":"open","priority":3,"issue_type":"task","owner":"bruno@postle.net","created_at":"2026-07-19T10:05:51Z","created_by":"Bruno Postle","updated_at":"2026-07-19T10:05:51Z","dependency_count":0,"dependent_count":0,"comment_count":0} {"id":"homemaker-py-161","title":"In-search evaluation of shape_rotate/deslim GA operators (7fm follow-up)","description":"homemaker-py-7fm's finish-time hill-climb found zero improving moves for\nmutate_shape_rotate/mutate_deslim (operators.py) on the 6-layout harbor-house\nsweep — every candidate move traded a shape fail for a new adjacency/access\nfail on the already co-evolved layout. That's a different regime from\nin-search use: a full multi-generation GA run gives selection pressure and\npopulation diversity a chance to accept a locally-worse move that a later\nstep or recombination completes.\n\nTo test: thread `fit` through driver.search (currently only reqs is passed to\noperators.mutate; shape_rotate/deslim need `fit` and currently no-op inside\nthe GA). Gate with an enable_shape_repair-style flag, mirroring how\nenable_reassociate (§12.3) let 9gp.2 do a clean A/B. Run full-budget\nharbor-house search with/without across seeds, compare final fail counts.\n\nIf negative again, the geometry-intrinsic residual on harbor-house-scale\nprogrammes may be a genuine floor for this representation, not a repairable\ninefficiency (consistent with §17/§19's framing). See DESIGN.md §19 and bd\nmemory collapse-global-94g-and-any-label-usage-optimisation for full context.","status":"open","priority":3,"issue_type":"task","owner":"bruno@postle.net","created_at":"2026-07-19T10:05:51Z","created_by":"Bruno Postle","updated_at":"2026-07-19T10:05:51Z","dependency_count":0,"dependent_count":0,"comment_count":0}
{"id":"homemaker-py-qpk","title":"In-search WFC collapse: run collapse_global per-eval during search (A/B vs finish-time)","description":"94g landed and validated the FINISH-TIME global cell-\u003eroom collapse (label search\nover a fixed geometry, monotone, best layout 15-\u003e12). The original 94g thrust was\na per-eval IN-SEARCH collapse: evolution searches unlabelled floorplans and the\nfitness collapses (optimally labels) each candidate before scoring, so search\noptimises the condensed objective directly. This issue is that step.\n\nMECHANISM: call collapse_global (or a cheaper incremental variant) inside\n_evaluate_full before the checks, same as the 9o5 collapse_superposition hook\n(fitness.py:_evaluate_full gates on self._superpose). Reuse the 94g substrate:\nc/o/s partition, level hard constraint, adjacency relaxation, public-access pin,\nthreshold objective.\n\nRISK (carried from homemaker-py-xi7, why 9o5 went negative): fitness =\nmax-over-labellings flattens/roughens the landscape — many topologies collapse to\nsimilar best scores, removing the gradient evolution climbs. AMPLIFIED at full\n(global) scope. The finish-time result does NOT de-risk this: finish-time is\nstrictly cannot-worsen by construction; in-search changes the objective every\neval. MUST A/B superpose-global ON vs OFF with a relaxation-gap log, exactly like\nxi7 did for 9o5, before adopting.\n\nCOST: collapse_global builds graphs + an assignment relaxation per eval — far more\nthan 9o5's per-class collapse. Needs an incremental/cheap variant or caching to be\naffordable in the inner loop; profile first.\n\nPrereq ordering: circulation placement (homemaker-py-qi6) and shape repair change\nthe skeleton/geometry the collapse labels over, so ideally sequence those first.\nRelated: 94g (finish-time, done), xi7 (9o5 A/B + relaxation-gap log), 9o5.","status":"open","priority":3,"issue_type":"feature","owner":"bruno@postle.net","created_at":"2026-07-18T10:12:54Z","created_by":"Bruno Postle","updated_at":"2026-07-18T10:12:54Z","dependencies":[{"issue_id":"homemaker-py-qpk","depends_on_id":"homemaker-py-94g","type":"discovered-from","created_at":"2026-07-18T11:12:54Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":0,"dependent_count":0,"comment_count":0} {"id":"homemaker-py-qpk","title":"In-search WFC collapse: run collapse_global per-eval during search (A/B vs finish-time)","description":"94g landed and validated the FINISH-TIME global cell-\u003eroom collapse (label search\nover a fixed geometry, monotone, best layout 15-\u003e12). The original 94g thrust was\na per-eval IN-SEARCH collapse: evolution searches unlabelled floorplans and the\nfitness collapses (optimally labels) each candidate before scoring, so search\noptimises the condensed objective directly. This issue is that step.\n\nMECHANISM: call collapse_global (or a cheaper incremental variant) inside\n_evaluate_full before the checks, same as the 9o5 collapse_superposition hook\n(fitness.py:_evaluate_full gates on self._superpose). Reuse the 94g substrate:\nc/o/s partition, level hard constraint, adjacency relaxation, public-access pin,\nthreshold objective.\n\nRISK (carried from homemaker-py-xi7, why 9o5 went negative): fitness =\nmax-over-labellings flattens/roughens the landscape — many topologies collapse to\nsimilar best scores, removing the gradient evolution climbs. AMPLIFIED at full\n(global) scope. The finish-time result does NOT de-risk this: finish-time is\nstrictly cannot-worsen by construction; in-search changes the objective every\neval. MUST A/B superpose-global ON vs OFF with a relaxation-gap log, exactly like\nxi7 did for 9o5, before adopting.\n\nCOST: collapse_global builds graphs + an assignment relaxation per eval — far more\nthan 9o5's per-class collapse. Needs an incremental/cheap variant or caching to be\naffordable in the inner loop; profile first.\n\nPrereq ordering: circulation placement (homemaker-py-qi6) and shape repair change\nthe skeleton/geometry the collapse labels over, so ideally sequence those first.\nRelated: 94g (finish-time, done), xi7 (9o5 A/B + relaxation-gap log), 9o5.","notes":"A/B VALIDATION COMPLETE (2026-07-19), xi7 protocol, 4 workers, equal eval\nbudget, both arms finished with standard --collapse (94g finish-time) so\ncomparison is on the final COLLAPSED score:\n\nharbor-house (init.dom, budget=2500, seeds 1-3): ON WINS 3/3.\n mean fails 80.3 -\u003e 72.0 (s1 85-\u003e74, s2 76-\u003e65, s3 80-\u003e77). No losses.\nprogramme-house (init.dom, budget=3000, seeds 1-5): ON wins 3/5.\n mean fails 8.4 -\u003e 7.8 (s1 8-\u003e5, s2 8-\u003e7, s4 10-\u003e9 win; s3 8-\u003e9, s5 8-\u003e9\n loss by 1 fail). Weaker/noisier on this much smaller building (already\n near its geometry floor, see section13/section19).\nCOMBINED head-to-head: ON 6, OFF 2.\n\nVERDICT: POSITIVE, and the OPPOSITE of the 9o5/xi7 prior (which was\nNULL/NEGATIVE for the per-class interchange relaxation). Unlike 9o5, this is\nthe SAME global WFC-style matching section17/94g already proved\nmonotone/positive at finish time -- running it every eval lets the outer\nsearch see the condensed objective instead of discovering it only once, and\nthe gradient survives rather than flattening. Effect scales WITH building\nsize (harbor-house clean 3/3 vs programme-house mixed), opposite of the 9o5\nlandscape-flattening fear.\n\nCOST: harbor-house wall-clock 102.6s(OFF)-\u003e177.8s(ON) ~1.73x;\nprogramme-house 39.0s-\u003e43.7s ~1.12x. Matches the profiled 1.5-1.9x/eval\nfigure.\n\nDECISION: kept default OFF (programme-house sample too mixed/small to flip\ndefault; 9o5/xi7 scar warrants a second larger-budget confirmation first),\nbut --collapse-insearch is a genuine, tested, working opt-in for\nharbor-house-scale-or-larger programmes. DESIGN.md section20 has full\nwriteup + per-seed numbers. Not filing a follow-up issue -- a larger-N\nprogramme-house seed sweep would be the natural next step if this gets\nrevisited, noted in DESIGN.md as a low-priority idea, not a blocker.\n\nRaw run logs/doms: /tmp scratchpad qpk_ab/ (not committed, ephemeral).","status":"in_progress","priority":3,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-07-18T10:12:54Z","created_by":"Bruno Postle","updated_at":"2026-07-19T19:18:21Z","started_at":"2026-07-19T10:11:11Z","dependencies":[{"issue_id":"homemaker-py-qpk","depends_on_id":"homemaker-py-94g","type":"discovered-from","created_at":"2026-07-18T11:12:54Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":0,"dependent_count":0,"comment_count":0}
{"id":"homemaker-py-kpu","title":"Schedule B: in-run leaf-sharing annealing (ramp grain down, unfold at each step)","description":"Spun out of homemaker-py-yaa, whose investigation is complete. yaa proved Schedule A (two-phase warm-start) works ONLY when shared leaves are unfolded at the sharing-\u003eno-sharing transition: naive warm-start stalls at 8.66e-08/70 fails, but unfold-then-de-share reaches 4.19e-06/15 fails — matching the direct --no-leaf-sharing baseline. operators.unfold_shared_leaves() is built, tested, and proven.\n\nSchedule B is the in-run variant: instead of a manual two-phase chain, anneal leaf_share_factor down within a single driver run (e.g. 4-\u003e3-\u003e2-\u003eoff) at eval thresholds. At each grain transition: (1) rebuild the cached (dir,sharing) evaluator at the new grain, (2) UNFOLD shared leaves that drop below the new grain so the population stays materialised (reuse operators.unfold_shared_leaves), (3) re-evaluate the whole population under the new evaluator, (4) resume local search. Gradual grain ramp = graduated non-convexity: avoids a single fitness cliff, keeps gross topology fixed on the smaller effective problem early, polishes per-room size/proportion/width late.\n\nDriver hooks needed (driver.py): the evaluator is cached per (dir, sharing) at fitness.py:415 and driver caches one per worker; the ramp must rebuild it and re-score the pop at each threshold. Modest change. Compare head-to-head vs (a) direct baseline 5.14e-06 and (b) the manual unfold warm-chain 4.19e-06 from yaa — does a graduated ramp beat a single hard unfold transition?\n\nWants the circulation-aware unfold from homemaker-py-8iv once available.","notes":"A/B DONE — NEGATIVE (2026-07-17). harbor-house 3M (500k/grain x3 + 1.5M polish, workers 4, ~22h): Schedule B = 1.26e-08 / 23 fails (canonical byte-for-byte). Loses decisively to both targets: direct baseline 5.14e-06/15 and warm-chain 4.19e-06/15 (~400x worse, +8 fails). Graduated ramp FALSIFIED: each grain step spikes fails (phase-end 19-\u003e21-\u003e27, final unfold 27-\u003e36); per-phase budget re-polishes partially-materialised states that the next step materialises further, so coarse-grain gains don't carry forward; polish started from a deeper hole (36) than the warm chain's single clean transition and only reached 23. The sharing-phase topology skeleton (yaa) is best cashed in ONCE at full grain, not annealed. Machinery retained (search_annealed, --anneal-grain, unfold above=, seed_pop, max_share override) — correct/tested/honest — but §15 single-transition finish stays the default. DESIGN §16 updated. Closing.","status":"closed","priority":3,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-07-15T06:48:11Z","created_by":"Bruno Postle","updated_at":"2026-07-17T15:33:01Z","started_at":"2026-07-16T06:47:57Z","closed_at":"2026-07-17T15:33:01Z","close_reason":"Closed","dependencies":[{"issue_id":"homemaker-py-kpu","depends_on_id":"homemaker-py-8iv","type":"blocks","created_at":"2026-07-15T07:48:30Z","created_by":"Bruno Postle","metadata":"{}"},{"issue_id":"homemaker-py-kpu","depends_on_id":"homemaker-py-yaa","type":"blocks","created_at":"2026-07-15T07:48:28Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":2,"dependent_count":0,"comment_count":0} {"id":"homemaker-py-kpu","title":"Schedule B: in-run leaf-sharing annealing (ramp grain down, unfold at each step)","description":"Spun out of homemaker-py-yaa, whose investigation is complete. yaa proved Schedule A (two-phase warm-start) works ONLY when shared leaves are unfolded at the sharing-\u003eno-sharing transition: naive warm-start stalls at 8.66e-08/70 fails, but unfold-then-de-share reaches 4.19e-06/15 fails — matching the direct --no-leaf-sharing baseline. operators.unfold_shared_leaves() is built, tested, and proven.\n\nSchedule B is the in-run variant: instead of a manual two-phase chain, anneal leaf_share_factor down within a single driver run (e.g. 4-\u003e3-\u003e2-\u003eoff) at eval thresholds. At each grain transition: (1) rebuild the cached (dir,sharing) evaluator at the new grain, (2) UNFOLD shared leaves that drop below the new grain so the population stays materialised (reuse operators.unfold_shared_leaves), (3) re-evaluate the whole population under the new evaluator, (4) resume local search. Gradual grain ramp = graduated non-convexity: avoids a single fitness cliff, keeps gross topology fixed on the smaller effective problem early, polishes per-room size/proportion/width late.\n\nDriver hooks needed (driver.py): the evaluator is cached per (dir, sharing) at fitness.py:415 and driver caches one per worker; the ramp must rebuild it and re-score the pop at each threshold. Modest change. Compare head-to-head vs (a) direct baseline 5.14e-06 and (b) the manual unfold warm-chain 4.19e-06 from yaa — does a graduated ramp beat a single hard unfold transition?\n\nWants the circulation-aware unfold from homemaker-py-8iv once available.","notes":"A/B DONE — NEGATIVE (2026-07-17). harbor-house 3M (500k/grain x3 + 1.5M polish, workers 4, ~22h): Schedule B = 1.26e-08 / 23 fails (canonical byte-for-byte). Loses decisively to both targets: direct baseline 5.14e-06/15 and warm-chain 4.19e-06/15 (~400x worse, +8 fails). Graduated ramp FALSIFIED: each grain step spikes fails (phase-end 19-\u003e21-\u003e27, final unfold 27-\u003e36); per-phase budget re-polishes partially-materialised states that the next step materialises further, so coarse-grain gains don't carry forward; polish started from a deeper hole (36) than the warm chain's single clean transition and only reached 23. The sharing-phase topology skeleton (yaa) is best cashed in ONCE at full grain, not annealed. Machinery retained (search_annealed, --anneal-grain, unfold above=, seed_pop, max_share override) — correct/tested/honest — but §15 single-transition finish stays the default. DESIGN §16 updated. Closing.","status":"closed","priority":3,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-07-15T06:48:11Z","created_by":"Bruno Postle","updated_at":"2026-07-17T15:33:01Z","started_at":"2026-07-16T06:47:57Z","closed_at":"2026-07-17T15:33:01Z","close_reason":"Closed","dependencies":[{"issue_id":"homemaker-py-kpu","depends_on_id":"homemaker-py-8iv","type":"blocks","created_at":"2026-07-15T07:48:30Z","created_by":"Bruno Postle","metadata":"{}"},{"issue_id":"homemaker-py-kpu","depends_on_id":"homemaker-py-yaa","type":"blocks","created_at":"2026-07-15T07:48:28Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":2,"dependent_count":0,"comment_count":0}
{"id":"homemaker-py-8iv","title":"unfold: route circulation to interior children (access/adjacency fails)","description":"Follow-up to homemaker-py-yaa. operators.unfold_shared_leaves() materialises a share=k leaf into a BALANCED binary subtree of k equal-target children. This closes the count deficit (all critical missing-room fails) but the balanced split creates interior children with no direct edge onto a corridor, so it introduces access/adjacency polish fails. Measured on harbor-house evolved-3M.dom: after unfold, 0 critical but 59 total fails, of which ~14 access + ~12 adjacency are attributable to the unrouted interior rooms (18 size / 2 width / 2 proportion are the k*target-\u003eper-leaf sizing mismatch, a separate concern).\n\nIdea: make the unfold subdivision circulation-aware instead of purely balanced — bias each cut so every new child retains an edge onto the shared leaf's original access boundary (or onto a sibling circulation leaf), mirroring the adjacency-aware constructive seeder (§11.6). Options: (a) orient/order the k-leaf subtree so children fan off the corridor side rather than nesting inward; (b) reserve a thin circulation spine within the unfolded block; (c) let a few post-unfold local-search evals fix it (cheaper, but that is exactly what the warm-start already does). Compare final endpoint with/without circulation-aware unfold against the 3M direct baseline (5.14e-06).\n\nRelates to the in-run annealing driver change (Schedule B, option B in yaa): if annealing rebuilds+re-evaluates the population at each grain transition, the unfold used there wants the same circulation-aware subdivision.","notes":"A/B VERDICT (seed 0, budget 150k, 4 workers, warm-start no-sharing polish from\nevolved-3M.dom): GRID WINS DECISIVELY. Circulation-aware slice LOSES.\n slice: 41 fails, fitness 3.52e-14\n grid : 25 fails, fitness 2.36e-09 (~5 orders better, 16 fewer fails)\nGrid led at EVERY milestone and the gap widened, not a near-tie:\n ~12k evals slice 67 / grid 49; ~36k slice 56 / grid 39;\n ~85k slice 46 / grid 30; ~130k slice 41 / grid 25.\nSlice never crossed. The thin-slab geometric debt (proportion/long/width) from\nforcing all k rooms onto one corridor wall costs MORE than the access routing\nsaves: local search re-routes access via topology moves (level_retype,\nplace_missing, level_fix) faster than it can widen thin slices (which it can't,\nwithout topology change — k equal slices of a compact leaf are intrinsically\nthin). Grid's squarer children are the better warm-start; yaa already showed grid\nreaches 4.19e-06.\n\nCONCLUSION: circulation-aware slicing is the WRONG trade. Retain the grid unfold.\nThe 8iv hypothesis (route access at unfold time) is falsified for the warm-start\nregime — access is better left to local search on a squarer seed. n=1 but the\ngap is large and monotone across the whole 150k-eval trajectory.","status":"closed","priority":3,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-07-12T15:38:34Z","created_by":"Bruno Postle","updated_at":"2026-07-16T06:35:46Z","started_at":"2026-07-15T13:46:49Z","closed_at":"2026-07-16T06:35:46Z","close_reason":"Investigated and falsified. Circulation-aware unfold (slice shared leaves perpendicular to their access edge so every child touches the corridor) was implemented + unit-tested, but the warm-start A/B (evolved-3M seed, 150k-eval no-sharing polish) shows it LOSES decisively to the existing balanced grid: slice 41 fails/3.5e-14 vs grid 25 fails/2.4e-09, grid leading monotonically at every milestone. Forcing k rooms onto one wall makes intrinsically thin slices whose geometric debt (proportion/long/width) local search cannot pay down without topology change, whereas grid's squarer children let local search re-route access cheaply via level_retype/place_missing/level_fix. Conclusion: retain grid unfold; access is better left to local search on a squarer seed. Code reverted (operators.py, test_operators.py back to grid). Findings in issue notes; A/B traces in examples/harbor-house/ab-8iv-*.","dependencies":[{"issue_id":"homemaker-py-8iv","depends_on_id":"homemaker-py-yaa","type":"blocks","created_at":"2026-07-12T16:48:16Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":1,"comment_count":0} {"id":"homemaker-py-8iv","title":"unfold: route circulation to interior children (access/adjacency fails)","description":"Follow-up to homemaker-py-yaa. operators.unfold_shared_leaves() materialises a share=k leaf into a BALANCED binary subtree of k equal-target children. This closes the count deficit (all critical missing-room fails) but the balanced split creates interior children with no direct edge onto a corridor, so it introduces access/adjacency polish fails. Measured on harbor-house evolved-3M.dom: after unfold, 0 critical but 59 total fails, of which ~14 access + ~12 adjacency are attributable to the unrouted interior rooms (18 size / 2 width / 2 proportion are the k*target-\u003eper-leaf sizing mismatch, a separate concern).\n\nIdea: make the unfold subdivision circulation-aware instead of purely balanced — bias each cut so every new child retains an edge onto the shared leaf's original access boundary (or onto a sibling circulation leaf), mirroring the adjacency-aware constructive seeder (§11.6). Options: (a) orient/order the k-leaf subtree so children fan off the corridor side rather than nesting inward; (b) reserve a thin circulation spine within the unfolded block; (c) let a few post-unfold local-search evals fix it (cheaper, but that is exactly what the warm-start already does). Compare final endpoint with/without circulation-aware unfold against the 3M direct baseline (5.14e-06).\n\nRelates to the in-run annealing driver change (Schedule B, option B in yaa): if annealing rebuilds+re-evaluates the population at each grain transition, the unfold used there wants the same circulation-aware subdivision.","notes":"A/B VERDICT (seed 0, budget 150k, 4 workers, warm-start no-sharing polish from\nevolved-3M.dom): GRID WINS DECISIVELY. Circulation-aware slice LOSES.\n slice: 41 fails, fitness 3.52e-14\n grid : 25 fails, fitness 2.36e-09 (~5 orders better, 16 fewer fails)\nGrid led at EVERY milestone and the gap widened, not a near-tie:\n ~12k evals slice 67 / grid 49; ~36k slice 56 / grid 39;\n ~85k slice 46 / grid 30; ~130k slice 41 / grid 25.\nSlice never crossed. The thin-slab geometric debt (proportion/long/width) from\nforcing all k rooms onto one corridor wall costs MORE than the access routing\nsaves: local search re-routes access via topology moves (level_retype,\nplace_missing, level_fix) faster than it can widen thin slices (which it can't,\nwithout topology change — k equal slices of a compact leaf are intrinsically\nthin). Grid's squarer children are the better warm-start; yaa already showed grid\nreaches 4.19e-06.\n\nCONCLUSION: circulation-aware slicing is the WRONG trade. Retain the grid unfold.\nThe 8iv hypothesis (route access at unfold time) is falsified for the warm-start\nregime — access is better left to local search on a squarer seed. n=1 but the\ngap is large and monotone across the whole 150k-eval trajectory.","status":"closed","priority":3,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-07-12T15:38:34Z","created_by":"Bruno Postle","updated_at":"2026-07-16T06:35:46Z","started_at":"2026-07-15T13:46:49Z","closed_at":"2026-07-16T06:35:46Z","close_reason":"Investigated and falsified. Circulation-aware unfold (slice shared leaves perpendicular to their access edge so every child touches the corridor) was implemented + unit-tested, but the warm-start A/B (evolved-3M seed, 150k-eval no-sharing polish) shows it LOSES decisively to the existing balanced grid: slice 41 fails/3.5e-14 vs grid 25 fails/2.4e-09, grid leading monotonically at every milestone. Forcing k rooms onto one wall makes intrinsically thin slices whose geometric debt (proportion/long/width) local search cannot pay down without topology change, whereas grid's squarer children let local search re-route access cheaply via level_retype/place_missing/level_fix. Conclusion: retain grid unfold; access is better left to local search on a squarer seed. Code reverted (operators.py, test_operators.py back to grid). Findings in issue notes; A/B traces in examples/harbor-house/ab-8iv-*.","dependencies":[{"issue_id":"homemaker-py-8iv","depends_on_id":"homemaker-py-yaa","type":"blocks","created_at":"2026-07-12T16:48:16Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":1,"comment_count":0}
{"id":"homemaker-py-yaa","title":"Investigate leaf-sharing annealing: shared early, materialise-and-de-share later","description":"Follow-up to homemaker-py-3l6. Leaf sharing is a fitness-evaluation knob over an identical genome representation (fitness.py:415; driver caches one evaluator per (dir, sharing)), and leaf_share_factor is a *grain* (0/1=off, N\u003e=2=share at grain N). That makes a coarse-to-fine / continuation schedule feasible: keep sharing on early to fix gross topology (level connectivity, adjacencies, massing) on a smaller effective problem, then reduce sharing to polish per-room size/proportion/width.\n\nTwo schedules to evaluate:\n A. Two-phase warm-start (no code): run sharing to convergence, feed the output .dom as seed to a --no-leaf-sharing run.\n B. In-run annealing (modest driver change): ramp leaf_share_factor down (e.g. 4-\u003e3-\u003e2-\u003eoff) at eval thresholds, rebuilding the cached evaluator and re-evaluating the population at each transition. Gradual grain ramp avoids a single fitness cliff (graduated non-convexity).\n\nKEY IDEA (Bruno): at the sharing-\u003eno-sharing transition, do not rely on place_missing/divide to rediscover the missing rooms. Instead PROGRAMMATICALLY SUBDIVIDE each shared leaf into the correct number of distinct spaces as an explicit 'unfold' operation. This directly pays down the materialisation deficit that otherwise makes a sharing-run seed start deep in the fail hole (evolved-3M.dom was missing ~12 rooms). The unfold turns a shared leaf of code X (share=k) into k sibling leaves of code X splitting its footprint, so the de-shared genome already satisfies the per-room count before local search resumes. This is also option 3 in 3l6 (materialise shared leaves on write) but applied mid-search at the phase change.\n\nRisk to characterise: sharing re-centres size targets on k*target, so the sharing-optimal massing (fewer, larger rooms) is geometrically different from the per-leaf optimum; the transferable value may be the adjacency/topology skeleton, not the sizing. The unfold subdivision needs to produce children with sensible individual proportions/widths, not just area.\n\nBaseline for comparison (harbor-house, init.dom, 3M, niced 1-worker warm chain): direct --no-leaf-sharing reached canonical 4.19e-06, 15 fails, 0 critical, still climbing. A head-to-head warm-start-from-sharing run (seed evolved-3M.dom) is running now (evolved-warmshare.dom) to measure whether the sharing topology, once forced honest, catches the direct route.","notes":"CONCLUSIVE (warm chain complete, ~2.67M total evals): evolved-unfold.dom = 4.19e-06, 15 fails, 0 critical (breakdown: 6 level, 5 size, 1 width). This MATCHES the direct --no-leaf-sharing baseline (evolved-3M-nols-2 4.19e-06/15 fails; -nols-3 5.14e-06/15 fails).\n\nVERDICT for yaa:\n- Schedule A NAIVE (warm-start from raw sharing seed): FAILS — stalls at 8.66e-08, 70 fails; place_missing/divide cannot dig out the ~15-room count deficit.\n- Schedule A + UNFOLD (operators.unfold_shared_leaves at the transition): CATCHES the direct route (4.19e-06, 15 fails). Bruno's key idea validated: the sharing-phase adjacency/topology skeleton is transferable; the materialisation (count) deficit — not the k*target sizing mismatch — was the sole blocker. Unfold pays it down so de-share local search resumes from a competitive genome.\n\nREMAINING: Schedule B (in-run annealing: ramp leaf_share_factor 4-\u003e3-\u003e2-\u003eoff mid-run, rebuilding+re-evaluating the population and unfolding at each grain step) is the still-unimplemented driver change. Now well-motivated: the unfold primitive it needs is built and proven. Recommend spinning Schedule B into its own implementation issue and closing yaa as the investigation it was scoped as. Circulation-routing refinement to unfold tracked in 8iv.","status":"closed","priority":3,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-07-05T16:33:35Z","created_by":"Bruno Postle","updated_at":"2026-07-15T13:37:07Z","started_at":"2026-07-12T14:52:21Z","closed_at":"2026-07-15T13:37:07Z","close_reason":"Closed","dependencies":[{"issue_id":"homemaker-py-yaa","depends_on_id":"homemaker-py-3l6","type":"blocks","created_at":"2026-07-05T17:33:56Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":2,"comment_count":0} {"id":"homemaker-py-yaa","title":"Investigate leaf-sharing annealing: shared early, materialise-and-de-share later","description":"Follow-up to homemaker-py-3l6. Leaf sharing is a fitness-evaluation knob over an identical genome representation (fitness.py:415; driver caches one evaluator per (dir, sharing)), and leaf_share_factor is a *grain* (0/1=off, N\u003e=2=share at grain N). That makes a coarse-to-fine / continuation schedule feasible: keep sharing on early to fix gross topology (level connectivity, adjacencies, massing) on a smaller effective problem, then reduce sharing to polish per-room size/proportion/width.\n\nTwo schedules to evaluate:\n A. Two-phase warm-start (no code): run sharing to convergence, feed the output .dom as seed to a --no-leaf-sharing run.\n B. In-run annealing (modest driver change): ramp leaf_share_factor down (e.g. 4-\u003e3-\u003e2-\u003eoff) at eval thresholds, rebuilding the cached evaluator and re-evaluating the population at each transition. Gradual grain ramp avoids a single fitness cliff (graduated non-convexity).\n\nKEY IDEA (Bruno): at the sharing-\u003eno-sharing transition, do not rely on place_missing/divide to rediscover the missing rooms. Instead PROGRAMMATICALLY SUBDIVIDE each shared leaf into the correct number of distinct spaces as an explicit 'unfold' operation. This directly pays down the materialisation deficit that otherwise makes a sharing-run seed start deep in the fail hole (evolved-3M.dom was missing ~12 rooms). The unfold turns a shared leaf of code X (share=k) into k sibling leaves of code X splitting its footprint, so the de-shared genome already satisfies the per-room count before local search resumes. This is also option 3 in 3l6 (materialise shared leaves on write) but applied mid-search at the phase change.\n\nRisk to characterise: sharing re-centres size targets on k*target, so the sharing-optimal massing (fewer, larger rooms) is geometrically different from the per-leaf optimum; the transferable value may be the adjacency/topology skeleton, not the sizing. The unfold subdivision needs to produce children with sensible individual proportions/widths, not just area.\n\nBaseline for comparison (harbor-house, init.dom, 3M, niced 1-worker warm chain): direct --no-leaf-sharing reached canonical 4.19e-06, 15 fails, 0 critical, still climbing. A head-to-head warm-start-from-sharing run (seed evolved-3M.dom) is running now (evolved-warmshare.dom) to measure whether the sharing topology, once forced honest, catches the direct route.","notes":"CONCLUSIVE (warm chain complete, ~2.67M total evals): evolved-unfold.dom = 4.19e-06, 15 fails, 0 critical (breakdown: 6 level, 5 size, 1 width). This MATCHES the direct --no-leaf-sharing baseline (evolved-3M-nols-2 4.19e-06/15 fails; -nols-3 5.14e-06/15 fails).\n\nVERDICT for yaa:\n- Schedule A NAIVE (warm-start from raw sharing seed): FAILS — stalls at 8.66e-08, 70 fails; place_missing/divide cannot dig out the ~15-room count deficit.\n- Schedule A + UNFOLD (operators.unfold_shared_leaves at the transition): CATCHES the direct route (4.19e-06, 15 fails). Bruno's key idea validated: the sharing-phase adjacency/topology skeleton is transferable; the materialisation (count) deficit — not the k*target sizing mismatch — was the sole blocker. Unfold pays it down so de-share local search resumes from a competitive genome.\n\nREMAINING: Schedule B (in-run annealing: ramp leaf_share_factor 4-\u003e3-\u003e2-\u003eoff mid-run, rebuilding+re-evaluating the population and unfolding at each grain step) is the still-unimplemented driver change. Now well-motivated: the unfold primitive it needs is built and proven. Recommend spinning Schedule B into its own implementation issue and closing yaa as the investigation it was scoped as. Circulation-routing refinement to unfold tracked in 8iv.","status":"closed","priority":3,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-07-05T16:33:35Z","created_by":"Bruno Postle","updated_at":"2026-07-15T13:37:07Z","started_at":"2026-07-12T14:52:21Z","closed_at":"2026-07-15T13:37:07Z","close_reason":"Closed","dependencies":[{"issue_id":"homemaker-py-yaa","depends_on_id":"homemaker-py-3l6","type":"blocks","created_at":"2026-07-05T17:33:56Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":2,"comment_count":0}
@ -81,26 +81,26 @@
{"id":"homemaker-py-erc.6","title":"Experiment: inner-loop slack-expansion objective term","description":"Inner-loop counterpart to plot-fill construction. If Diagnostic B shows the inner loop has room to expand leaves into slack but no objective gradient to do so (the scalar rewards hitting target area but not exceeding it where slack exists), add a term/incentive so the ratio optimiser pushes leaf boundaries out to consume neighbouring slack and satisfy size, rather than parking at target.\n\nCONDITIONAL on Diagnostic B: build this only if B localizes the gap to the inner loop (room to expand, no gradient); if B shows construction targets too-small dims, prefer the plot-fill construction sibling. Must preserve the §5.4 inner-loop cliff / §4.9 lexicographic protection — the term sits where it cannot displace the fail-count ordering. A/B vs §12.2 baseline, seeds 0/1/2, 20000 evals, staged, default-OFF. Record DESIGN.md §13.6.","notes":"DEPRIORITISED by Diagnostic B (§13.2). B shows the inner loop CANNOT repair undersize: the slack is depth-driven maldistribution baked into the frozen topology, and the equal-offset ratio DOF cannot shrink a 14x leaf to feed a starved one without trading into shape fails (0.5^n cliff). Wrong DOF and wrong direction — the blocker is slicing POSITION, not a missing expansion reward. Fix belongs upstream in construction/topology (erc.4 re-scoped, erc.3). Keep as a low-priority follow-up only if a depth-balanced construction still leaves a residual size gradient the inner loop could pick up.","status":"closed","priority":4,"issue_type":"feature","owner":"bruno@postle.net","created_at":"2026-06-22T23:16:24Z","created_by":"Bruno Postle","updated_at":"2026-06-28T13:22:22Z","closed_at":"2026-06-28T13:22:22Z","close_reason":"wont-fix (DESIGN §13.7): Diag B (§13.2) showed the inner loop cannot repair undersize (wrong DOF — slicing position, frozen-topology ratios). Superseded by depth-balanced construction (erc.4). Condition unmet.","dependencies":[{"issue_id":"homemaker-py-erc.6","depends_on_id":"homemaker-py-erc","type":"parent-child","created_at":"2026-06-23T00:16:23Z","created_by":"Bruno Postle","metadata":"{}"},{"issue_id":"homemaker-py-erc.6","depends_on_id":"homemaker-py-erc.2","type":"blocks","created_at":"2026-06-23T00:16:47Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":0,"comment_count":0} {"id":"homemaker-py-erc.6","title":"Experiment: inner-loop slack-expansion objective term","description":"Inner-loop counterpart to plot-fill construction. If Diagnostic B shows the inner loop has room to expand leaves into slack but no objective gradient to do so (the scalar rewards hitting target area but not exceeding it where slack exists), add a term/incentive so the ratio optimiser pushes leaf boundaries out to consume neighbouring slack and satisfy size, rather than parking at target.\n\nCONDITIONAL on Diagnostic B: build this only if B localizes the gap to the inner loop (room to expand, no gradient); if B shows construction targets too-small dims, prefer the plot-fill construction sibling. Must preserve the §5.4 inner-loop cliff / §4.9 lexicographic protection — the term sits where it cannot displace the fail-count ordering. A/B vs §12.2 baseline, seeds 0/1/2, 20000 evals, staged, default-OFF. Record DESIGN.md §13.6.","notes":"DEPRIORITISED by Diagnostic B (§13.2). B shows the inner loop CANNOT repair undersize: the slack is depth-driven maldistribution baked into the frozen topology, and the equal-offset ratio DOF cannot shrink a 14x leaf to feed a starved one without trading into shape fails (0.5^n cliff). Wrong DOF and wrong direction — the blocker is slicing POSITION, not a missing expansion reward. Fix belongs upstream in construction/topology (erc.4 re-scoped, erc.3). Keep as a low-priority follow-up only if a depth-balanced construction still leaves a residual size gradient the inner loop could pick up.","status":"closed","priority":4,"issue_type":"feature","owner":"bruno@postle.net","created_at":"2026-06-22T23:16:24Z","created_by":"Bruno Postle","updated_at":"2026-06-28T13:22:22Z","closed_at":"2026-06-28T13:22:22Z","close_reason":"wont-fix (DESIGN §13.7): Diag B (§13.2) showed the inner loop cannot repair undersize (wrong DOF — slicing position, frozen-topology ratios). Superseded by depth-balanced construction (erc.4). Condition unmet.","dependencies":[{"issue_id":"homemaker-py-erc.6","depends_on_id":"homemaker-py-erc","type":"parent-child","created_at":"2026-06-23T00:16:23Z","created_by":"Bruno Postle","metadata":"{}"},{"issue_id":"homemaker-py-erc.6","depends_on_id":"homemaker-py-erc.2","type":"blocks","created_at":"2026-06-23T00:16:47Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":0,"comment_count":0}
{"id":"homemaker-py-erc.5","title":"Experiment: compactness-aware cuts (minimize leaf perimeter/area)","description":"Attacks the #1 factor, crinkliness (346) — a per-leaf perimeter/area property DISTINCT from proportion (aspect ratio). Proportion-aware seeding (leu.2) sizes splits but does not bias toward balanced, square-ish subdivision. Add a KD-tree-style 'keep both children compact' cut rule (prefer the cut orientation/position that minimises summed child perimeter/area) in construction.\n\nCONDITIONAL on Diagnostic A: if A shows per-leaf shape-fail is FLAT across densities (floor intrinsic to slicing density), better cuts at the same leaf count will not pay → this should be closed wont-fix in favour of leaf-sharing. Only build if A shows shape-fail RISES with density. A/B vs §12.2 baseline, seeds 0/1/2, 20000 evals, staged, default-OFF. Record DESIGN.md §13.5.","notes":"DEPRIORITISED by erc.1 verdict (§13.1): per-leaf shape-fail flat vs slicing density and cuts already squarest (_size_divisions_from_targets picks squarest rotation) yet still ~1.8 fails/leaf =\u003e little compactness headroom at fixed leaf count. Floor is intrinsic to leaf COUNT, not cut quality. Revisit only if leaf-sharing (erc.3) underdelivers.","status":"closed","priority":4,"issue_type":"feature","owner":"bruno@postle.net","created_at":"2026-06-22T23:16:21Z","created_by":"Bruno Postle","updated_at":"2026-06-28T13:22:17Z","closed_at":"2026-06-28T13:22:17Z","close_reason":"wont-fix (DESIGN §13.7): Diag A (§13.1) showed the floor is intrinsic to leaf COUNT not cut quality; revisit condition was 'only if leaf-sharing underdelivers' but leaf-sharing OVER-delivered (−32…−39%, §13.3). Condition unmet.","dependencies":[{"issue_id":"homemaker-py-erc.5","depends_on_id":"homemaker-py-erc","type":"parent-child","created_at":"2026-06-23T00:16:21Z","created_by":"Bruno Postle","metadata":"{}"},{"issue_id":"homemaker-py-erc.5","depends_on_id":"homemaker-py-erc.1","type":"blocks","created_at":"2026-06-23T00:16:43Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":0,"comment_count":0} {"id":"homemaker-py-erc.5","title":"Experiment: compactness-aware cuts (minimize leaf perimeter/area)","description":"Attacks the #1 factor, crinkliness (346) — a per-leaf perimeter/area property DISTINCT from proportion (aspect ratio). Proportion-aware seeding (leu.2) sizes splits but does not bias toward balanced, square-ish subdivision. Add a KD-tree-style 'keep both children compact' cut rule (prefer the cut orientation/position that minimises summed child perimeter/area) in construction.\n\nCONDITIONAL on Diagnostic A: if A shows per-leaf shape-fail is FLAT across densities (floor intrinsic to slicing density), better cuts at the same leaf count will not pay → this should be closed wont-fix in favour of leaf-sharing. Only build if A shows shape-fail RISES with density. A/B vs §12.2 baseline, seeds 0/1/2, 20000 evals, staged, default-OFF. Record DESIGN.md §13.5.","notes":"DEPRIORITISED by erc.1 verdict (§13.1): per-leaf shape-fail flat vs slicing density and cuts already squarest (_size_divisions_from_targets picks squarest rotation) yet still ~1.8 fails/leaf =\u003e little compactness headroom at fixed leaf count. Floor is intrinsic to leaf COUNT, not cut quality. Revisit only if leaf-sharing (erc.3) underdelivers.","status":"closed","priority":4,"issue_type":"feature","owner":"bruno@postle.net","created_at":"2026-06-22T23:16:21Z","created_by":"Bruno Postle","updated_at":"2026-06-28T13:22:17Z","closed_at":"2026-06-28T13:22:17Z","close_reason":"wont-fix (DESIGN §13.7): Diag A (§13.1) showed the floor is intrinsic to leaf COUNT not cut quality; revisit condition was 'only if leaf-sharing underdelivers' but leaf-sharing OVER-delivered (−32…−39%, §13.3). Condition unmet.","dependencies":[{"issue_id":"homemaker-py-erc.5","depends_on_id":"homemaker-py-erc","type":"parent-child","created_at":"2026-06-23T00:16:21Z","created_by":"Bruno Postle","metadata":"{}"},{"issue_id":"homemaker-py-erc.5","depends_on_id":"homemaker-py-erc.1","type":"blocks","created_at":"2026-06-23T00:16:43Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":1,"dependent_count":0,"comment_count":0}
{"id":"homemaker-py-2g5","title":"Rebuild occlusion/daylight/sun subsystem in Python (post-Phase-5, after optimisation fully native)","description":"DESIGN.md §6 port scope — a whole subsystem, not a term. quality_daylight (Leaf.pm:281-296) needs Urb::Misc::Sun + Urb::Field::Occlusion (+CIESky); quality_uncrinkliness also takes the occlusion object. Indoor spaces return 1 for daylight; cost is outdoor spaces + crinkliness. Port Sun_horizontal (262980-minute normalisation) and the occlusion wall set from Dom-\u003eWalls.","acceptance_criteria":"Daylight and crinkliness factors match Perl (float tolerance) across the corpus, including multi-storey cases","notes":"Re-scoped 2026-06-12: occlusion disabled in the Urb oracle instead of ported (see homemaker-py-gp2). Native fitness ships with simple crinkliness (illumination factor = 1, in homemaker-py-gnw). This issue is now the eventual Python occlusion rebuild, only after optimisation works entirely in Python. Restores outdoor-daylight and shaded-wall selection pressure.\nReframed 2026-06-17: orthogonal to epic homemaker-py-c4c. This is fitness FIDELITY (restoring daylight + shaded-wall selection pressure to match Perl), not search CAPABILITY — it changes what 'good' means, not the search's ability to find good. It will NOT improve final designs in the sense currently sought. Stays P4, deferred until the topology-search-quality epic lands and optimisation is fully native.","status":"open","priority":4,"issue_type":"feature","owner":"bruno@postle.net","created_at":"2026-06-11T23:38:25Z","created_by":"Bruno Postle","updated_at":"2026-06-17T19:14:48Z","dependency_count":0,"dependent_count":0,"comment_count":0} {"id":"homemaker-py-2g5","title":"Rebuild occlusion/daylight/sun subsystem in Python (post-Phase-5, after optimisation fully native)","description":"DESIGN.md §6 port scope — a whole subsystem, not a term. quality_daylight (Leaf.pm:281-296) needs Urb::Misc::Sun + Urb::Field::Occlusion (+CIESky); quality_uncrinkliness also takes the occlusion object. Indoor spaces return 1 for daylight; cost is outdoor spaces + crinkliness. Port Sun_horizontal (262980-minute normalisation) and the occlusion wall set from Dom-\u003eWalls.","acceptance_criteria":"Daylight and crinkliness factors match Perl (float tolerance) across the corpus, including multi-storey cases","notes":"Re-scoped 2026-06-12: occlusion disabled in the Urb oracle instead of ported (see homemaker-py-gp2). Native fitness ships with simple crinkliness (illumination factor = 1, in homemaker-py-gnw). This issue is now the eventual Python occlusion rebuild, only after optimisation works entirely in Python. Restores outdoor-daylight and shaded-wall selection pressure.\nReframed 2026-06-17: orthogonal to epic homemaker-py-c4c. This is fitness FIDELITY (restoring daylight + shaded-wall selection pressure to match Perl), not search CAPABILITY — it changes what 'good' means, not the search's ability to find good. It will NOT improve final designs in the sense currently sought. Stays P4, deferred until the topology-search-quality epic lands and optimisation is fully native.","status":"open","priority":4,"issue_type":"feature","owner":"bruno@postle.net","created_at":"2026-06-11T23:38:25Z","created_by":"Bruno Postle","updated_at":"2026-06-17T19:14:48Z","dependency_count":0,"dependent_count":0,"comment_count":0}
{"_type":"memory","key":"island-model-psk-14-is-a-null-priming","value":"Island model (psk, §14) is a NULL: priming a population from N converged independent elites + crossover-heavy migration does not beat best-of-N at equal total budget (maple island 124 vs control 116). The child_probe instrument shows WHY: area-matched crossover across independently-converged elites almost never synthesizes (1-3 of ~64 children beat the better parent, max drop 2-5) because the slicing encoding is non-canonical (9gp), so splices are disruptive not combinatorial. Search-machinery null #3 after graded-objective and niching/restarts; residual stays geometry/shape-bound."}
{"_type":"memory","key":"proportion-aware-constructive-seeding-leu-2-12-2","value":"Proportion-aware constructive seeding (leu.2/§12.2): sizing seed cuts from target AREAS only regresses (thin slivers wreck aspect); you must ALSO pick each cut's rotation for child squareness. It is a convergence ACCELERATOR via a deeper local optimum around the constructed topology: wins where that topology is roughly right and budget is scarce (harbor -13%, maple -10% at 20k evals) but DELAYS small programmes where the seed must be restructured by undivide (programme-house regresses at fixed budget, yet reaches the floor given budget - speed, not asymptote). Default-on. Also: n_storeys must honour storey_minimum, not just level: keys (programme-house storey_minimum:2, all rooms level:0 - was seeded 1 storey short; cq1)."} {"_type":"memory","key":"proportion-aware-constructive-seeding-leu-2-12-2","value":"Proportion-aware constructive seeding (leu.2/§12.2): sizing seed cuts from target AREAS only regresses (thin slivers wreck aspect); you must ALSO pick each cut's rotation for child squareness. It is a convergence ACCELERATOR via a deeper local optimum around the constructed topology: wins where that topology is roughly right and budget is scarce (harbor -13%, maple -10% at 20k evals) but DELAYS small programmes where the seed must be restructured by undivide (programme-house regresses at fixed budget, yet reaches the floor given budget - speed, not asymptote). Default-on. Also: n_storeys must honour storey_minimum, not just level: keys (programme-house storey_minimum:2, all rooms level:0 - was seeded 1 storey short; cq1)."}
{"_type":"memory","key":"user-preference-bruno-this-is-a-fedora-system","value":"User preference (Bruno): this is a Fedora system — NEVER install Python packages via pip without asking first; always ask whether to install the rpm via dnf (e.g. python3-cma) before considering pip. Applies to any dependency additions."}
{"_type":"memory","key":"warm-x0-initialization-bug-pattern-when-a-topology","value":"warm_x0 initialization bug pattern: when a topology operator explicitly sets division ratios on a newly-created node (e.g. compound_fix sets node.division=[0.25,0.25] for t3), parent.ratios has no entry for that node (it was a leaf). warm_x0 defaults it to 0.5, corrupting the inner loop's starting point and making the operator invisible to lex comparison. Fix: only propagate child ratios for nodes where the parent node was NOT already divided; stale hidden nodes revealed by structural mutations (swap flipping b.below) must NOT contribute their pre-writeback values. See driver.py lines 259-267 (fixed 2026-06-14)."}
{"_type":"memory","key":"9o5-multi-use-leaves-is-path-a-superposition","value":"9o5 multi-use leaves is path (a) — superposition as SEARCH RELAXATION that COLLAPSES to specific usage at the end, NOT path (b) loose-fit/no-collapse. Bruno's intent: codes with SIMILAR leaf requirements form an interchangeable equivalence class; during evolution the solver doesn't commit which leaf serves which specific usage (smoother landscape, no fighting over exact leaf usage); at the end the layout is CONDENSED to specific usages by brute-forcing the in-class assignment (3 interchangeable usages over 3 leaves = 3! = 6 combinations to check, pick best). 'Derive automatically' compatibility = requirement-similarity grouping. This reverses the issue's stated 'path b preferred' note."} {"_type":"memory","key":"9o5-multi-use-leaves-is-path-a-superposition","value":"9o5 multi-use leaves is path (a) — superposition as SEARCH RELAXATION that COLLAPSES to specific usage at the end, NOT path (b) loose-fit/no-collapse. Bruno's intent: codes with SIMILAR leaf requirements form an interchangeable equivalence class; during evolution the solver doesn't commit which leaf serves which specific usage (smoother landscape, no fighting over exact leaf usage); at the end the layout is CONDENSED to specific usages by brute-forcing the in-class assignment (3 interchangeable usages over 3 leaves = 3! = 6 combinations to check, pick best). 'Derive automatically' compatibility = requirement-similarity grouping. This reverses the issue's stated 'path b preferred' note."}
{"_type":"memory","key":"collapse-global-94g-and-any-label-usage-optimisation","value":"collapse_global (94g) and any label/usage optimisation CANNOT fix geometry-intrinsic fails. The harbor-house 15-fail best layout contains long-thin cells that are useless whatever room usage is assigned — their width/proportion/crinkliness fails are shape-bound, not label slack. Two consequences: (1) do not over-claim collapse gains — only ~2-3 of that layout's fails are reclaimable relabel slack, the rest are geometry- or building-level bound; (2) the threshold objective must not be tuned to 'pass' a degenerate cell via a permissive room type — a metric-pass on a physically useless space is gaming, not a fix. Real remedies for these are geometry/topology search (cell shape) and circulation placement, filed separately, not the collapse."}
{"_type":"memory","key":"correction-to-urb-fitness-bug-memory-bruno-2026","value":"CORRECTION to urb-fitness-bug memory (Bruno, 2026-06-12): 'C' is NOT a 'covered' type — Is_Covered is a geometric predicate (indoor space above). Urb's generic types are canonically UPPERCASE: C=circulation, O=outside, S=sahn (get_space_types qw/C O S/; corpus is 100% uppercase, never 'c'/'o' leaves). The mixed-case designs that fired the latent ratio_type first-match bug were created by homemaker's own operator type pool emitting lowercase 'c'/'o' — fixed: driver/operators now emit uppercase generics only, and class checks use t[0].lower() in 'cos'. The Urb class-sum patch stays as defensive hardening (zero impact on canonical designs). Native port (3y7/gnw): treat type classes case-insensitively, generics canonically uppercase."}
{"_type":"memory","key":"deceptive-valleys-in-topology-search-when-every-single","value":"Deceptive valleys in topology search: when every single-step mutation from a target state passes through a high-fail intermediary (e.g. level_fix displaces a room into 5+ new fails), a compound operator that atomically applies two coordinated changes can escape. Design compound operators to land on the low-fail state directly, bypassing the deceptive gradient. Programme-house example: level_compound_fix atomically moves the level-constrained room AND re-inserts the displaced room adjacent to C in one step (operators.py, 2026-06-14)."}
{"_type":"memory","key":"homemaker-py-3l6-fix-leaf-sharing-evolve-runs","value":"homemaker-py-3l6 fix: leaf-sharing evolve runs now auto-finish before write via driver.polish_finish — unfold_shared_leaves() then a warm-started leaf_sharing=False polish search (--polish-budget, default budget//2). Makes the written .dom honest under canonical homemaker-fitness (internal==canonical when leaf_sharing off). Interrupt path forces polish_budget=0 (unfold+rescore only). This is yaa's unfold-then-polish, made automatic; Schedule B annealing is still kpu."} {"_type":"memory","key":"homemaker-py-3l6-fix-leaf-sharing-evolve-runs","value":"homemaker-py-3l6 fix: leaf-sharing evolve runs now auto-finish before write via driver.polish_finish — unfold_shared_leaves() then a warm-started leaf_sharing=False polish search (--polish-budget, default budget//2). Makes the written .dom honest under canonical homemaker-fitness (internal==canonical when leaf_sharing off). Interrupt path forces polish_budget=0 (unfold+rescore only). This is yaa's unfold-then-polish, made automatic; Schedule B annealing is still kpu."}
{"_type":"memory","key":"strategy-decision-2026-06-12-bruno-occlusion-daylight","value":"Strategy decision 2026-06-12 (Bruno): occlusion/daylight is ORTHOGONAL to building a scalable optimiser. Disable it in Urb (env flag, homemaker-py-gp2) rather than port it; native fitness uses simple crinkliness (illumination factor = 1); rebuild occlusion in Python only after optimisation is fully native (homemaker-py-2g5, now P4). Consequence: all scores change when the flag flips — re-baseline corpus/.score, DESIGN \\$4.5 gains, gate bars at one clean boundary AFTER homemaker-py-1p0 closes; Phase-2 urb-evolve benchmark must run with the same flag."}
{"_type":"memory","key":"island-model-psk-14-is-a-null-priming","value":"Island model (psk, §14) is a NULL: priming a population from N converged independent elites + crossover-heavy migration does not beat best-of-N at equal total budget (maple island 124 vs control 116). The child_probe instrument shows WHY: area-matched crossover across independently-converged elites almost never synthesizes (1-3 of ~64 children beat the better parent, max drop 2-5) because the slicing encoding is non-canonical (9gp), so splices are disruptive not combinatorial. Search-machinery null #3 after graded-objective and niching/restarts; residual stays geometry/shape-bound."}
{"_type":"memory","key":"never-use-corpus-filenames-candidate-001-dom-candidate","value":"Never use corpus filenames (candidate-001.dom, candidate-002.dom, generated.dom, init.dom, etc.) as --output targets when running experiments. These are test fixtures. Always write experimental outputs to scratch/ or a timestamped path. Lesson from 2026-06-14: warm-start runs overwrote candidate-001/002.dom and broke graph tests."} {"_type":"memory","key":"never-use-corpus-filenames-candidate-001-dom-candidate","value":"Never use corpus filenames (candidate-001.dom, candidate-002.dom, generated.dom, init.dom, etc.) as --output targets when running experiments. These are test fixtures. Always write experimental outputs to scratch/ or a timestamped path. Lesson from 2026-06-14: warm-start runs overwrote candidate-001/002.dom and broke graph tests."}
{"_type":"memory","key":"programme-house-optimisation-result-2026-06-14-15","value":"Programme-house optimisation result (2026-06-14/15): best achievable is 1 fail (l1 wrong level, score ~0.005). 0 fails is geometrically impossible: l1 (min 27m²) must occupy ll (~23m²) at level 0, which eliminates the t3-adj-C provider; dividing ll into lll(l1)+llr(C) gives llr proportion ~6:1 (fails). Python memetic optimizer achieves 1 fail in 50k evals vs Perl optimiser's 2-3 fails. Winning topology: TWO C nodes at level 0 — ll(C) for t3-adj-C via geometric contact, rl(C) for staircase via tree-sibling adjacency to rrr(O). Best .dom: scratch/from-warmstart-fixed.dom and scratch/from-compound3-fixed.dom."} {"_type":"memory","key":"urb-oracle-nondeterminism-urb-fitness-pl-output-varies","value":"Urb oracle nondeterminism: urb-fitness.pl output varies run-to-run from Perl hash-order randomisation — .fails line ORDER shuffles (compare sorted, use oracle.Score.fail_lines) and the score float can flip by ~1 ULP (compare with math.isclose rel_tol=1e-12, never ==). Not a batching artifact; affects single runs too. Matters for the Phase 3 native-fitness parity gate (homemaker-py-uxz)."}
{"_type":"memory","key":"correction-to-urb-fitness-bug-memory-bruno-2026","value":"CORRECTION to urb-fitness-bug memory (Bruno, 2026-06-12): 'C' is NOT a 'covered' type — Is_Covered is a geometric predicate (indoor space above). Urb's generic types are canonically UPPERCASE: C=circulation, O=outside, S=sahn (get_space_types qw/C O S/; corpus is 100% uppercase, never 'c'/'o' leaves). The mixed-case designs that fired the latent ratio_type first-match bug were created by homemaker's own operator type pool emitting lowercase 'c'/'o' — fixed: driver/operators now emit uppercase generics only, and class checks use t[0].lower() in 'cos'. The Urb class-sum patch stays as defensive hardening (zero impact on canonical designs). Native port (3y7/gnw): treat type classes case-insensitively, generics canonically uppercase."}
{"_type":"memory","key":"experiment-harness-gotcha-the-leaf-sharing-relaxed-objective","value":"Experiment harness gotcha: the leaf-sharing RELAXED objective (§13.3) is injected ONLY by monkeypatching fitness.load_config in the parent process (run_staged_search.py / probe scripts). This is parent-process-only and does NOT propagate into ProcessPoolExecutor workers (n_workers\u003e1), which re-import fitness fresh and score under the STRICT on-disk patterns.config -\u003e r.n_fails MISMATCH (worker strict vs parent relaxed re-score). ALL §13.x floor runs were therefore SERIAL. Any future PARALLEL leaf-sharing experiment will silently mis-score until leaf_sharing lives on disk/CLI (tracked: homemaker-py-x3b). The parallel driver itself is correct; both paths score via load_config(programme_dir)."}
{"_type":"memory","key":"multi-storey-staircase-consistency-when-dividing-or-retyping","value":"Multi-storey staircase consistency: when dividing or retyping a circulation (C) leaf at one level, the same structural change should be propagated to the matching leaf on ALL other storeys so the stair core path is maintained. The optimizer cannot fix staircase disruptions through trial-and-error geometry alone — it requires a synchronized multi-level operator that applies the same topology change to every storey simultaneously."}
{"_type":"memory","key":"unfold-strategy-for-shared-leaves-homemaker-py-8iv","value":"Unfold strategy for shared leaves (homemaker-py-8iv, resolved 2026-07-16): use the BALANCED GRID (operators._grow_balanced/_size_subtree_equal), NOT circulation-aware slicing. Slicing a shared leaf perpendicular to its access edge so every child touches the corridor was implemented + A/B-tested and LOST decisively (150k-eval warm-start polish from evolved-3M: slice 41 fails/3.5e-14 vs grid 25 fails/2.4e-09, grid ahead at every milestone). Reason: k rooms all touching one wall are intrinsically thin slices; that geometric debt (proportion/long/width) is unfixable without topology change, while the grid's squarer children let local search re-route access cheaply via level_retype/place_missing/level_fix. Lesson: at the sharing-\u003eno-sharing transition, prioritise squarer children and leave access to local search; do not reintroduce slicing in Schedule B (kpu)."} {"_type":"memory","key":"unfold-strategy-for-shared-leaves-homemaker-py-8iv","value":"Unfold strategy for shared leaves (homemaker-py-8iv, resolved 2026-07-16): use the BALANCED GRID (operators._grow_balanced/_size_subtree_equal), NOT circulation-aware slicing. Slicing a shared leaf perpendicular to its access edge so every child touches the corridor was implemented + A/B-tested and LOST decisively (150k-eval warm-start polish from evolved-3M: slice 41 fails/3.5e-14 vs grid 25 fails/2.4e-09, grid ahead at every milestone). Reason: k rooms all touching one wall are intrinsically thin slices; that geometric debt (proportion/long/width) is unfixable without topology change, while the grid's squarer children let local search re-route access cheaply via level_retype/place_missing/level_fix. Lesson: at the sharing-\u003eno-sharing transition, prioritise squarer children and leave access to local search; do not reintroduce slicing in Schedule B (kpu)."}
{"_type":"memory","key":"cli-tool-style-prefer-python-m-homemaker-module","value":"CLI tool style: prefer python -m homemaker.module --parameters pattern, installable via pip install -e . with pyproject.toml entry_points. Not standalone bin/ scripts."} {"_type":"memory","key":"cli-tool-style-prefer-python-m-homemaker-module","value":"CLI tool style: prefer python -m homemaker.module --parameters pattern, installable via pip install -e . with pyproject.toml entry_points. Not standalone bin/ scripts."}
{"_type":"memory","key":"run-to-run-reproducibility-in-homemaker-layout-serial","value":"Run-to-run reproducibility in homemaker-layout: serial search (workers=1) is byte-for-byte deterministic; parallel (workers\u003e1) is now deterministic too AFTER fixing driver._run_batch to admit futures in submission order (was as_completed/completion order, bug xcy). Reproducibility holds only for a FIXED worker count — serial vs parallel differ because children-per-iteration is 1 vs n_workers (different batch granularity), which is expected, not a bug. The constructive seeder was NEVER nondeterministic: _assign_adjacency_aware has unique idx tiebreaks; comparing topologies with Python builtin hash() of the signature STRING is invalid (PYTHONHASHSEED salts str hashing per process) — use a stable hash (sha1) or genome.signature equality."} {"_type":"memory","key":"collapse-global-94g-and-any-label-usage-optimisation","value":"collapse_global (94g) and any label/usage optimisation CANNOT fix geometry-intrinsic fails. The harbor-house 15-fail best layout contains long-thin cells that are useless whatever room usage is assigned — their width/proportion/crinkliness fails are shape-bound, not label slack. Two consequences: (1) do not over-claim collapse gains — only ~2-3 of that layout's fails are reclaimable relabel slack, the rest are geometry- or building-level bound; (2) the threshold objective must not be tuned to 'pass' a degenerate cell via a permissive room type — a metric-pass on a physically useless space is gaming, not a fix. Real remedies for these are geometry/topology search (cell shape) and circulation placement, filed separately, not the collapse."}
{"_type":"memory","key":"urb-fitness-bug-found-fixed-2026-06-12","value":"Urb fitness bug found+fixed 2026-06-12 (patch in /home/bruno/src/urb, uncommitted): ProgrammeDriven.pm ratio_o/ratio_type grepped case-insensitively over the ratios hash and took the FIRST key — nondeterministic (x4.5 score swings) for designs with mixed-case type classes (both 'c' circulation and 'C' covered). Fixed to SUM the class (matches Is_Circulation//Is_Outside semantics); 35/35 corpus scores unchanged. CRITICAL for homemaker-py-3y7/gnw: the native port must implement class-SUM ratios. Building.pm has the same unpatched pattern (site-driven path, not used by our oracle). Also: the memetic search reward-hacked this bug before the fix — search results predating it are noise artifacts."} {"_type":"memory","key":"deceptive-valleys-in-topology-search-when-every-single","value":"Deceptive valleys in topology search: when every single-step mutation from a target state passes through a high-fail intermediary (e.g. level_fix displaces a room into 5+ new fails), a compound operator that atomically applies two coordinated changes can escape. Design compound operators to land on the low-fail state directly, bypassing the deceptive gradient. Programme-house example: level_compound_fix atomically moves the level-constrained room AND re-inserts the displaced room adjacent to C in one step (operators.py, 2026-06-14)."}
{"_type":"memory","key":"urb-oracle-nondeterminism-urb-fitness-pl-output-varies","value":"Urb oracle nondeterminism: urb-fitness.pl output varies run-to-run from Perl hash-order randomisation — .fails line ORDER shuffles (compare sorted, use oracle.Score.fail_lines) and the score float can flip by ~1 ULP (compare with math.isclose rel_tol=1e-12, never ==). Not a batching artifact; affects single runs too. Matters for the Phase 3 native-fitness parity gate (homemaker-py-uxz)."} {"_type":"memory","key":"programme-house-optimisation-result-2026-06-14-15","value":"Programme-house optimisation result (2026-06-14/15): best achievable is 1 fail (l1 wrong level, score ~0.005). 0 fails is geometrically impossible: l1 (min 27m²) must occupy ll (~23m²) at level 0, which eliminates the t3-adj-C provider; dividing ll into lll(l1)+llr(C) gives llr proportion ~6:1 (fails). Python memetic optimizer achieves 1 fail in 50k evals vs Perl optimiser's 2-3 fails. Winning topology: TWO C nodes at level 0 — ll(C) for t3-adj-C via geometric contact, rl(C) for staircase via tree-sibling adjacency to rrr(O). Best .dom: scratch/from-warmstart-fixed.dom and scratch/from-compound3-fixed.dom."}
{"_type":"memory","key":"experiment-seeding-pitfall-run-search-scaled-py-s","value":"Experiment seeding pitfall: run_search_scaled.py's default PH_SEED (c964…dom) is a FINISHED programme-house design — passing it warm-starts and floors at ~3 fails, NOT a blank-slate topology search. For blank-slate runs comparable to §11.5/§11.6 baselines, seed from examples/programme-house/init.dom (a bare undivided plot; driver bootstrap auto-triggers only on bare plots). Bit the 6zy sweep — first pass used c964 and falsely showed 3-fail floor across the whole grid."}
{"_type":"memory","key":"user-preference-bruno-this-is-a-fedora-system","value":"User preference (Bruno): this is a Fedora system — NEVER install Python packages via pip without asking first; always ask whether to install the rpm via dnf (e.g. python3-cma) before considering pip. Applies to any dependency additions."}
{"_type":"memory","key":"adjacency-in-binary-slicing-tree-is-structural-not","value":"Adjacency in binary slicing tree is structural, not geometric: the inner-loop NM cannot fix topological adjacency failures. Two paths exist: (1) tree-sibling adjacency — a node is adjacent to its sibling in the tree; (2) cross-zone geometric adjacency — leaves from different subtrees that happen to share a boundary. Staircase/adjacency fails require a topology mutation that changes which nodes are siblings or which zones touch. This was proved empirically on programme-house: staircase fail from rot=0 layout could not be fixed by NM but was fixed by level_retype creating a two-C topology (2026-06-14/15)."}
{"_type":"memory","key":"ld2-13-6-interior-o-seed-diagnostic-all","value":"ld2/§13.6 interior-O seed diagnostic: ALL crinkliness fails in the constructed bal+share seed are UNDER-exposed (crink\u003c0.62, landlocked rooms with no facade + no uncovered-O neighbour) — zero over-exposed sliver fails. So the erc crinkliness residual is genuine under-daylighting, validating the interior light-well premise. Default outside_divisor=6 was too sparse (null: harbor 147-\u003e142, crinkliness even rose). odiv=3 is the seed-optimal joint setting: harbor seed fails 147-\u003e129 (-18), maple 219-\u003e206 (-14), landlocked fails drop, at cost of more leaves (harbor +4, maple +8). Because it ADDS leaves it carries the §13.4 wash-out risk; A/B to convergence pending."} {"_type":"memory","key":"ld2-13-6-interior-o-seed-diagnostic-all","value":"ld2/§13.6 interior-O seed diagnostic: ALL crinkliness fails in the constructed bal+share seed are UNDER-exposed (crink\u003c0.62, landlocked rooms with no facade + no uncovered-O neighbour) — zero over-exposed sliver fails. So the erc crinkliness residual is genuine under-daylighting, validating the interior light-well premise. Default outside_divisor=6 was too sparse (null: harbor 147-\u003e142, crinkliness even rose). odiv=3 is the seed-optimal joint setting: harbor seed fails 147-\u003e129 (-18), maple 219-\u003e206 (-14), landlocked fails drop, at cost of more leaves (harbor +4, maple +8). Because it ADDS leaves it carries the §13.4 wash-out risk; A/B to convergence pending."}
{"_type":"memory","key":"experiment-harness-gotcha-the-leaf-sharing-relaxed-objective","value":"Experiment harness gotcha: the leaf-sharing RELAXED objective (§13.3) is injected ONLY by monkeypatching fitness.load_config in the parent process (run_staged_search.py / probe scripts). This is parent-process-only and does NOT propagate into ProcessPoolExecutor workers (n_workers\u003e1), which re-import fitness fresh and score under the STRICT on-disk patterns.config -\u003e r.n_fails MISMATCH (worker strict vs parent relaxed re-score). ALL §13.x floor runs were therefore SERIAL. Any future PARALLEL leaf-sharing experiment will silently mis-score until leaf_sharing lives on disk/CLI (tracked: homemaker-py-x3b). The parallel driver itself is correct; both paths score via load_config(programme_dir)."} {"_type":"memory","key":"run-to-run-reproducibility-in-homemaker-layout-serial","value":"Run-to-run reproducibility in homemaker-layout: serial search (workers=1) is byte-for-byte deterministic; parallel (workers\u003e1) is now deterministic too AFTER fixing driver._run_batch to admit futures in submission order (was as_completed/completion order, bug xcy). Reproducibility holds only for a FIXED worker count — serial vs parallel differ because children-per-iteration is 1 vs n_workers (different batch granularity), which is expected, not a bug. The constructive seeder was NEVER nondeterministic: _assign_adjacency_aware has unique idx tiebreaks; comparing topologies with Python builtin hash() of the signature STRING is invalid (PYTHONHASHSEED salts str hashing per process) — use a stable hash (sha1) or genome.signature equality."}
{"_type":"memory","key":"warm-x0-initialization-bug-pattern-when-a-topology","value":"warm_x0 initialization bug pattern: when a topology operator explicitly sets division ratios on a newly-created node (e.g. compound_fix sets node.division=[0.25,0.25] for t3), parent.ratios has no entry for that node (it was a leaf). warm_x0 defaults it to 0.5, corrupting the inner loop's starting point and making the operator invisible to lex comparison. Fix: only propagate child ratios for nodes where the parent node was NOT already divided; stale hidden nodes revealed by structural mutations (swap flipping b.below) must NOT contribute their pre-writeback values. See driver.py lines 259-267 (fixed 2026-06-14)."}
{"_type":"memory","key":"homemaker-py-pythonpath-set-pythonpath-home-bruno-src","value":"homemaker-layout PYTHONPATH: package installed as 'homemaker-layout' via pip install -e . so 'import homemaker_layout' works from anywhere without PYTHONPATH. For running tests use 'python -m pytest' from project root /home/bruno/src/homemaker-layout (pyproject.toml adds src/ automatically). Never try pip show homemaker — that's the old homemaker-addon conflict."} {"_type":"memory","key":"homemaker-py-pythonpath-set-pythonpath-home-bruno-src","value":"homemaker-layout PYTHONPATH: package installed as 'homemaker-layout' via pip install -e . so 'import homemaker_layout' works from anywhere without PYTHONPATH. For running tests use 'python -m pytest' from project root /home/bruno/src/homemaker-layout (pyproject.toml adds src/ automatically). Never try pip show homemaker — that's the old homemaker-addon conflict."}
{"_type":"memory","key":"multi-storey-staircase-consistency-when-dividing-or-retyping","value":"Multi-storey staircase consistency: when dividing or retyping a circulation (C) leaf at one level, the same structural change should be propagated to the matching leaf on ALL other storeys so the stair core path is maintained. The optimizer cannot fix staircase disruptions through trial-and-error geometry alone — it requires a synchronized multi-level operator that applies the same topology change to every storey simultaneously."} {"_type":"memory","key":"adjacency-in-binary-slicing-tree-is-structural-not","value":"Adjacency in binary slicing tree is structural, not geometric: the inner-loop NM cannot fix topological adjacency failures. Two paths exist: (1) tree-sibling adjacency — a node is adjacent to its sibling in the tree; (2) cross-zone geometric adjacency — leaves from different subtrees that happen to share a boundary. Staircase/adjacency fails require a topology mutation that changes which nodes are siblings or which zones touch. This was proved empirically on programme-house: staircase fail from rot=0 layout could not be fixed by NM but was fixed by level_retype creating a two-C topology (2026-06-14/15)."}
{"_type":"memory","key":"strategy-decision-2026-06-12-bruno-occlusion-daylight","value":"Strategy decision 2026-06-12 (Bruno): occlusion/daylight is ORTHOGONAL to building a scalable optimiser. Disable it in Urb (env flag, homemaker-py-gp2) rather than port it; native fitness uses simple crinkliness (illumination factor = 1); rebuild occlusion in Python only after optimisation is fully native (homemaker-py-2g5, now P4). Consequence: all scores change when the flag flips — re-baseline corpus/.score, DESIGN \\$4.5 gains, gate bars at one clean boundary AFTER homemaker-py-1p0 closes; Phase-2 urb-evolve benchmark must run with the same flag."} {"_type":"memory","key":"experiment-seeding-pitfall-run-search-scaled-py-s","value":"Experiment seeding pitfall: run_search_scaled.py's default PH_SEED (c964…dom) is a FINISHED programme-house design — passing it warm-starts and floors at ~3 fails, NOT a blank-slate topology search. For blank-slate runs comparable to §11.5/§11.6 baselines, seed from examples/programme-house/init.dom (a bare undivided plot; driver bootstrap auto-triggers only on bare plots). Bit the 6zy sweep — first pass used c964 and falsely showed 3-fail floor across the whole grid."}
{"_type":"memory","key":"urb-fitness-bug-found-fixed-2026-06-12","value":"Urb fitness bug found+fixed 2026-06-12 (patch in /home/bruno/src/urb, uncommitted): ProgrammeDriven.pm ratio_o/ratio_type grepped case-insensitively over the ratios hash and took the FIRST key — nondeterministic (x4.5 score swings) for designs with mixed-case type classes (both 'c' circulation and 'C' covered). Fixed to SUM the class (matches Is_Circulation//Is_Outside semantics); 35/35 corpus scores unchanged. CRITICAL for homemaker-py-3y7/gnw: the native port must implement class-SUM ratios. Building.pm has the same unpatched pattern (site-driven path, not used by our oracle). Also: the memetic search reward-hacked this bug before the fix — search results predating it are noise artifacts."}

71
DESIGN.md
View file

@ -2535,3 +2535,74 @@ gate with an `enable_shape_repair`-style flag as §12.3 did for `reassociate`, r
with/without) is the remaining open question and would need to be its own measured experiment 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 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. issue's candidate mechanisms is a dead end, not that geometry repair is impossible in general.
## 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 / next.** 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.

44
src/homemaker_layout/driver.py
View file

@ -41,13 +41,16 @@ _CHILD_INNER_KW: dict = {}
def _overrides_for(leaf_sharing: bool, superpose: bool, def _overrides_for(leaf_sharing: bool, superpose: bool,
max_share: int | None = None, max_share: int | None = None,
conn_grade: bool = False) -> dict | None: conn_grade: bool = False,
collapse_insearch: bool = False) -> dict | None:
"""Run-level conf overrides for the native evaluator (None when all off). """Run-level conf overrides for the native evaluator (None when all off).
``max_share`` (homemaker-py-kpu) overrides the evaluator's ``leaf_share_max`` ``max_share`` (homemaker-py-kpu) overrides the evaluator's ``leaf_share_max``
grain cap for the in-run annealing ramp; ``None`` leaves the config default. grain cap for the in-run annealing ramp; ``None`` leaves the config default.
``conn_grade`` (homemaker-py-qi6) turns the graded proximity scalar into the ``conn_grade`` (homemaker-py-qi6) turns the graded proximity scalar into the
circulation-connectivity signal (§18). circulation-connectivity signal (§18). ``collapse_insearch`` (homemaker-py-
qpk) runs the 94g global cell<->room collapse inside every fitness eval
instead of once at finish time.
""" """
ov: dict = {} ov: dict = {}
if leaf_sharing: if leaf_sharing:
@ -58,6 +61,8 @@ def _overrides_for(leaf_sharing: bool, superpose: bool,
ov["leaf_share_max"] = int(max_share) ov["leaf_share_max"] = int(max_share)
if conn_grade: if conn_grade:
ov["conn_grade"] = True ov["conn_grade"] = True
if collapse_insearch:
ov["collapse_insearch"] = True
return ov or None return ov or None
@ -65,7 +70,8 @@ def _overrides_for(leaf_sharing: bool, superpose: bool,
def _fitness_for(programme_dir: str, leaf_sharing: bool = False, def _fitness_for(programme_dir: str, leaf_sharing: bool = False,
superpose: bool = False, superpose: bool = False,
max_share: int | None = None, max_share: int | None = None,
conn_grade: bool = False) -> "fitness.Fitness": conn_grade: bool = False,
collapse_insearch: bool = False) -> "fitness.Fitness":
"""Cached Fitness evaluator per (programme dir, leaf_sharing) (config load is """Cached Fitness evaluator per (programme dir, leaf_sharing) (config load is
the cost). the cost).
@ -76,7 +82,8 @@ def _fitness_for(programme_dir: str, leaf_sharing: bool = False,
inner loop instead of reading the on-disk (sharing-free) patterns.config. inner loop instead of reading the on-disk (sharing-free) patterns.config.
Cached per process — workers fork their own copy. Cached per process — workers fork their own copy.
""" """
overrides = _overrides_for(leaf_sharing, superpose, max_share, conn_grade) overrides = _overrides_for(leaf_sharing, superpose, max_share, conn_grade,
collapse_insearch)
conf, cost = fitness.load_config(programme_dir, overrides=overrides) conf, cost = fitness.load_config(programme_dir, overrides=overrides)
return fitness.Fitness(conf, cost) return fitness.Fitness(conf, cost)
@ -153,7 +160,8 @@ def _evaluate(root: dom.Node, programme_dir, urb_root, x0, budget, inner_kw,
leaf_sharing: bool = False, leaf_sharing: bool = False,
superpose: bool = False, superpose: bool = False,
max_share: int | None = None, max_share: int | None = None,
conn_grade: bool = False) -> tuple[Individual, int]: conn_grade: bool = False,
collapse_insearch: bool = False) -> tuple[Individual, int]:
# §12.3 shape-feasibility pre-filter (homemaker-py-9gp.1): if even the best # §12.3 shape-feasibility pre-filter (homemaker-py-9gp.1): if even the best
# achievable (proportion-aware) geometry of this topology already has at least # achievable (proportion-aware) geometry of this topology already has at least
# as many shape fails as the incumbent's TOTAL fails — and exceeds the tunable # as many shape fails as the incumbent's TOTAL fails — and exceeds the tunable
@ -161,12 +169,13 @@ def _evaluate(root: dom.Node, programme_dir, urb_root, x0, budget, inner_kw,
# eval instead of spending the full inner-loop budget. The best_n_fails guard # eval instead of spending the full inner-loop budget. The best_n_fails guard
# makes the proxy safe: a topology whose shape-fail floor is still below the # makes the proxy safe: a topology whose shape-fail floor is still below the
# incumbent is never discarded. Pruned individuals are tagged and never admitted. # incumbent is never discarded. Pruned individuals are tagged and never admitted.
overrides = _overrides_for(leaf_sharing, superpose, max_share, conn_grade) overrides = _overrides_for(leaf_sharing, superpose, max_share, conn_grade,
collapse_insearch)
if (feasibility_max_shape_fails is not None and best_n_fails is not None): if (feasibility_max_shape_fails is not None and best_n_fails is not None):
pred = operators.predicted_shape_fails( pred = operators.predicted_shape_fails(
root, _reqs_for(str(programme_dir)), root, _reqs_for(str(programme_dir)),
_fitness_for(str(programme_dir), leaf_sharing, superpose, max_share, _fitness_for(str(programme_dir), leaf_sharing, superpose, max_share,
conn_grade)) conn_grade, collapse_insearch))
if pred > feasibility_max_shape_fails and pred >= best_n_fails: if pred > feasibility_max_shape_fails and pred >= best_n_fails:
ind = Individual(root=root, fitness=0.0, n_fails=pred, ratios={}, ind = Individual(root=root, fitness=0.0, n_fails=pred, ratios={},
lineage=f"pruned/{lineage}", grade=0.0, lineage=f"pruned/{lineage}", grade=0.0,
@ -182,7 +191,7 @@ def _evaluate(root: dom.Node, programme_dir, urb_root, x0, budget, inner_kw,
if want_grade: if want_grade:
_, _, grade = _fitness_for( _, _, grade = _fitness_for(
str(programme_dir), leaf_sharing, superpose, max_share, str(programme_dir), leaf_sharing, superpose, max_share,
conn_grade).score_with_grade( conn_grade, collapse_insearch).score_with_grade(
copy.deepcopy(root)) copy.deepcopy(root))
ind = Individual(root=root, fitness=r.fitness, n_fails=r.n_fails, ind = Individual(root=root, fitness=r.fitness, n_fails=r.n_fails,
ratios=innerloop.ratio_map(root), lineage=lineage, ratios=innerloop.ratio_map(root), lineage=lineage,
@ -237,6 +246,7 @@ def search(
outside_divisor: int = 3, outside_divisor: int = 3,
max_share: int | None = None, max_share: int | None = None,
seed_pop: list[dom.Node] | None = None, seed_pop: list[dom.Node] | None = None,
collapse_insearch: bool = False,
) -> SearchResult: ) -> SearchResult:
"""Run the memetic loop from ``seed_root`` until ``budget`` oracle """Run the memetic loop from ``seed_root`` until ``budget`` oracle
evaluations are consumed. Returns the best individual found; its ``root`` evaluations are consumed. Returns the best individual found; its ``root``
@ -278,6 +288,12 @@ def search(
kpu) supplies an explicit initial population of decoded roots — evaluated kpu) supplies an explicit initial population of decoded roots — evaluated
under this phase's evaluator instead of bootstrapping or single-seeding — so a under this phase's evaluator instead of bootstrapping or single-seeding — so a
grain-anneal ramp can hand a whole population from one phase to the next. grain-anneal ramp can hand a whole population from one phase to the next.
``collapse_insearch`` (homemaker-py-qpk, EXPERIMENTAL, default off) runs the
94g global cell<->room collapse inside every fitness eval instead of once at
finish time, so search optimises the collapsed objective directly. Carries
the 9o5/xi7 landscape-flattening risk at global scope — do not flip default
on without a positive A/B (DESIGN.md §17 follow-on).
""" """
from .oracle import DEFAULT_URB_ROOT from .oracle import DEFAULT_URB_ROOT
@ -415,7 +431,8 @@ def search(
best_nf = result.best.n_fails if result.best is not None else None best_nf = result.best.n_fails if result.best is not None else None
full = [ full = [
(root, programme_dir, urb_root, x0, budget_, kw_, lin, use_grade, (root, programme_dir, urb_root, x0, budget_, kw_, lin, use_grade,
mx, best_nf, leaf_sharing, superpose, max_share, conn_grade) mx, best_nf, leaf_sharing, superpose, max_share, conn_grade,
collapse_insearch)
for root, x0, budget_, kw_, lin in tasks for root, x0, budget_, kw_, lin in tasks
] ]
if _pool is not None: if _pool is not None:
@ -497,7 +514,8 @@ def search(
leaf_sharing=leaf_sharing, leaf_sharing=leaf_sharing,
superpose=superpose, superpose=superpose,
max_share=max_share, max_share=max_share,
conn_grade=conn_grade) conn_grade=conn_grade,
collapse_insearch=collapse_insearch)
n_evals += used n_evals += used
admit(seed_ind, pop) admit(seed_ind, pop)
@ -588,6 +606,7 @@ def polish_finish(
seed: int = 0, seed: int = 0,
n_workers: int = 1, n_workers: int = 1,
superpose: bool = False, superpose: bool = False,
collapse_insearch: bool = False,
rescore_budget: int = 200, rescore_budget: int = 200,
log=None, log=None,
) -> SearchResult: ) -> SearchResult:
@ -632,14 +651,15 @@ def polish_finish(
unfolded, programme_dir, budget=polish_budget, pop_size=pop_size, unfolded, programme_dir, budget=polish_budget, pop_size=pop_size,
child_budget=child_budget, p_crossover=p_crossover, seed=seed, child_budget=child_budget, p_crossover=p_crossover, seed=seed,
n_workers=n_workers, bootstrap=False, leaf_sharing=False, n_workers=n_workers, bootstrap=False, leaf_sharing=False,
superpose=superpose, log=log, superpose=superpose, collapse_insearch=collapse_insearch, log=log,
) )
else: else:
# No polish: re-optimise the unfolded genome's ratios once and score it # No polish: re-optimise the unfolded genome's ratios once and score it
# canonically so the written .dom and reported fitness are honest. # canonically so the written .dom and reported fitness are honest.
ind, used = _evaluate( ind, used = _evaluate(
unfolded, programme_dir, None, x0=None, budget=rescore_budget, unfolded, programme_dir, None, x0=None, budget=rescore_budget,
inner_kw={}, lineage="unfold", leaf_sharing=False, superpose=superpose) inner_kw={}, lineage="unfold", leaf_sharing=False, superpose=superpose,
collapse_insearch=collapse_insearch)
r2 = SearchResult(best=ind, population=[ind], n_evals=used, n_topologies=1) r2 = SearchResult(best=ind, population=[ind], n_evals=used, n_topologies=1)
r2.n_distinct_signatures = 1 r2.n_distinct_signatures = 1
r2.history = [(0, ind.fitness, ind.lineage)] r2.history = [(0, ind.fitness, ind.lineage)]

12
src/homemaker_layout/evolve.py
View file

@ -105,6 +105,15 @@ def _parse_args(argv=None) -> argparse.Namespace:
"circulation that the binary 'not connected' fail lacks. " "circulation that the binary 'not connected' fail lacks. "
"Does not change the scalar fitness or fail count " "Does not change the scalar fitness or fail count "
"(default: off)") "(default: off)")
p.add_argument("--collapse-insearch", dest="collapse_insearch",
action=argparse.BooleanOptionalAction,
default=_env_bool("HOMEMAKER_COLLAPSE_INSEARCH", False),
help="homemaker-py-qpk (EXPERIMENTAL, §17 follow-on): run the "
"94g global cell→room collapse inside every fitness eval "
"instead of once at finish time, so search optimises the "
"collapsed objective directly. Carries the 9o5/xi7 "
"landscape-flattening risk at global scope — validate "
"with an A/B before relying on it (default: off)")
p.add_argument("--anneal-grain", type=str, p.add_argument("--anneal-grain", type=str,
default=os.environ.get("HOMEMAKER_ANNEAL_GRAIN"), default=os.environ.get("HOMEMAKER_ANNEAL_GRAIN"),
metavar="LADDER", metavar="LADDER",
@ -172,6 +181,7 @@ def main(argv=None) -> int:
file=sys.stderr) file=sys.stderr)
print(f"superpose : {args.superpose}", file=sys.stderr) print(f"superpose : {args.superpose}", file=sys.stderr)
print(f"conn grade : {args.conn_grade}", file=sys.stderr) print(f"conn grade : {args.conn_grade}", file=sys.stderr)
print(f"collapse in-search : {args.collapse_insearch}", file=sys.stderr)
print(f"output : {out or 'stdout'}", file=sys.stderr, flush=True) print(f"output : {out or 'stdout'}", file=sys.stderr, flush=True)
anneal_ladder = None anneal_ladder = None
@ -221,6 +231,7 @@ def main(argv=None) -> int:
leaf_share_factor=args.leaf_share_factor, leaf_share_factor=args.leaf_share_factor,
superpose=args.superpose, superpose=args.superpose,
conn_grade=args.conn_grade, conn_grade=args.conn_grade,
collapse_insearch=args.collapse_insearch,
log=lambda m: print(m, file=sys.stderr, flush=True), log=lambda m: print(m, file=sys.stderr, flush=True),
) )
_finish_sharing = args.leaf_sharing _finish_sharing = args.leaf_sharing
@ -250,6 +261,7 @@ def main(argv=None) -> int:
seed=args.seed, seed=args.seed,
n_workers=args.workers, n_workers=args.workers,
superpose=args.superpose, superpose=args.superpose,
collapse_insearch=args.collapse_insearch,
log=lambda m: print(m, file=sys.stderr, flush=True), log=lambda m: print(m, file=sys.stderr, flush=True),
) )

28
src/homemaker_layout/fitness.py
View file

@ -224,6 +224,20 @@ class Fitness:
from .programme import CLASS_CAP as _CLASS_CAP from .programme import CLASS_CAP as _CLASS_CAP
self._class_cap = int(self.conf("superpose_class_cap") or _CLASS_CAP) self._class_cap = int(self.conf("superpose_class_cap") or _CLASS_CAP)
self._interchange_classes: list | None = None # lazily derived self._interchange_classes: list | None = None # lazily derived
# homemaker-py-qpk: IN-SEARCH global collapse. Default OFF. When on,
# collapse_global (the 94g finish-time cell<->room relabel) runs INSIDE
# _evaluate_full every eval instead of once at the end, so search
# optimises the collapsed objective directly (mirrors the 9o5 per-eval
# collapse above, at GLOBAL scope). Carries the 9o5/xi7 landscape-
# flattening risk amplified to the whole building; gated behind its own
# A/B (DESIGN.md §17 follow-on, homemaker-py-qpk) — do not default on
# without a positive result.
self._collapse_insearch = bool(self.conf("collapse_insearch"))
adj = self.conf("collapse_insearch_adjacency")
self._collapse_insearch_adjacency = True if adj is None else bool(adj)
# Fewer Jacobi passes than the finish-time default (6): a per-eval cost,
# not a one-shot polish — profile before raising.
self._collapse_insearch_iters = int(self.conf("collapse_insearch_iters") or 3)
# ------------------------------------------------------------------ # # ------------------------------------------------------------------ #
# Type superposition + collapse (homemaker-py-9o5) # Type superposition + collapse (homemaker-py-9o5)
@ -1420,6 +1434,20 @@ class Fitness:
if self._superpose: if self._superpose:
self.collapse_superposition(root) self.collapse_superposition(root)
# homemaker-py-qpk: IN-SEARCH global collapse (DESIGN.md §17 follow-on).
# Runs before any check, same as collapse_superposition above, so counts/
# adjacency/quality downstream see the collapsed (relabelled) types. Uses
# its own graph build (fixed geometry, only labels move) — safe to call
# on the unmerged tree, exactly as collapse_global's finish-time use does.
if self._collapse_insearch:
self.collapse_global(
root,
adjacency=self._collapse_insearch_adjacency,
objective="threshold",
preserve_public_access=True,
iters=self._collapse_insearch_iters,
)
# --- Phase 1: UNMERGED tree checks --- # --- Phase 1: UNMERGED tree checks ---
check_fails, missing = graph_mod.check_space_counts( check_fails, missing = graph_mod.check_space_counts(
root, programme, self._leaf_sharing, self._max_share) root, programme, self._leaf_sharing, self._max_share)

122
tests/test_collapse_insearch.py Normal file
View file

@ -0,0 +1,122 @@
"""Tests for the IN-SEARCH global collapse (homemaker-py-qpk, DESIGN.md §17
follow-on): running Fitness.collapse_global inside every fitness eval instead
of once at finish time.
Covers:
- default-OFF guarantee + conf-driven knobs (adjacency, iters)
- _evaluate_full wiring: collapse_global is invoked (with the right kwargs)
when the flag is on, never when it is off
- end-to-end effect on a real evolved layout, cross-checked against the
documented 94g finish-time result (DESIGN.md §17: 15 -> 12 fails on
evolved-3M-nols-3.dom)
"""
import copy
from pathlib import Path
from unittest.mock import patch
import pytest
from homemaker_layout import dom as dom_mod
from homemaker_layout.dom import Node, _link_subtree
from homemaker_layout.fitness import Fitness, load_config
HARBOR = Path(__file__).parent.parent / "examples" / "harbor-house"
def _two_leaf_root(t_left: str, t_right: str, side: float = 6.0, div: float = 0.4):
from homemaker_layout import geometry
geometry.clear_cache()
root = Node(
node=[[0, 0], [side, 0], [side, side], [0, side]],
rotation=0, division=[div, div],
left=Node(type=t_left), right=Node(type=t_right),
)
_link_subtree(root, None, "")
return root
# --------------------------------------------------------------------------- #
# Defaults + conf-driven knobs
# --------------------------------------------------------------------------- #
def test_collapse_insearch_default_off():
fit = Fitness()
assert fit._collapse_insearch is False
assert fit._collapse_insearch_adjacency is True
assert fit._collapse_insearch_iters == 3
def test_collapse_insearch_flag_on():
fit = Fitness(conf={"collapse_insearch": True})
assert fit._collapse_insearch is True
def test_collapse_insearch_adjacency_knob_off():
fit = Fitness(conf={"collapse_insearch": True,
"collapse_insearch_adjacency": False})
assert fit._collapse_insearch_adjacency is False
def test_collapse_insearch_iters_knob():
fit = Fitness(conf={"collapse_insearch": True, "collapse_insearch_iters": 5})
assert fit._collapse_insearch_iters == 5
# --------------------------------------------------------------------------- #
# _evaluate_full wiring
# --------------------------------------------------------------------------- #
def test_evaluate_full_calls_collapse_global_when_on():
fit = Fitness(conf={"collapse_insearch": True, "collapse_insearch_iters": 2,
"spaces": {"b1": {"size": [16.0, 4.0], "count": 2}}})
root = _two_leaf_root("b1", "b1")
with patch.object(Fitness, "collapse_global", wraps=fit.collapse_global) as m:
fit.score_with_fails(root)
m.assert_called_once()
_, kw = m.call_args
assert kw["adjacency"] is True
assert kw["objective"] == "threshold"
assert kw["preserve_public_access"] is True
assert kw["iters"] == 2
def test_evaluate_full_does_not_call_collapse_global_when_off():
fit = Fitness(conf={"spaces": {"b1": {"size": [16.0, 4.0], "count": 2}}})
root = _two_leaf_root("b1", "b1")
with patch.object(Fitness, "collapse_global", wraps=fit.collapse_global) as m:
fit.score_with_fails(root)
m.assert_not_called()
# --------------------------------------------------------------------------- #
# End-to-end: matches the documented 94g finish-time result
# --------------------------------------------------------------------------- #
@pytest.mark.skipif(not HARBOR.is_dir(), reason="harbor-house example absent")
def test_collapse_insearch_reproduces_94g_finish_time_result():
# DESIGN.md §17: the finish-time collapse takes this layout 15 -> 12 fails.
# collapse_insearch runs the SAME collapse_global earlier in the SAME
# pipeline (before Phase-1 checks instead of after the whole search), so it
# must reach the identical fail count on this fixed-geometry layout.
conf, cost = load_config(HARBOR)
conf_ci, _ = load_config(HARBOR, overrides={"collapse_insearch": True})
fit, fit_ci = Fitness(conf, cost), Fitness(conf_ci, cost)
root = dom_mod.load(str(HARBOR / "evolved-3M-nols-3.dom"))
_, f_base = fit.score_with_fails(copy.deepcopy(root))
_, f_ci = fit_ci.score_with_fails(copy.deepcopy(root))
assert len(f_base) == 15
assert len(f_ci) == 12
@pytest.mark.skipif(not HARBOR.is_dir(), reason="harbor-house example absent")
def test_collapse_insearch_off_reproduces_baseline():
conf, cost = load_config(HARBOR)
fit = Fitness(conf, cost)
root = dom_mod.load(str(HARBOR / "evolved-3M-nols-3.dom"))
s1, f1 = fit.score_with_fails(copy.deepcopy(root))
s2, f2 = Fitness(conf, cost).score_with_fails(copy.deepcopy(root))
assert s1 == pytest.approx(s2)
assert f1 == f2