f1d: ruin-and-recreate LNS operator, validated positive on programme-house

Adds operators.mutate_ruin_recreate: un-divides one wing of a storey and rebuilds it with the same adjacency-aware constructor the seeders use (_assign_adjacency_aware, generalised with a new `scope` param), seeded from the surviving circulation bordering the wing. Gated behind enable_ruin_recreate (default off) / --ruin-recreate, same pattern as reassociate/bridge_circulation. A/B (qpk protocol, DESIGN.md §23): initial uniform-weight run was underpowered (fired ~1/32 children), null. A weight=3.0 follow-up (_MUTATION_WEIGHTS["ruin_recreate"]) showed a statistically significant win on programme-house across 15 seeds (8W/1L/6T, mean fails 7.07->6.00, Wilcoxon p=0.041) but no consistent effect on harbor-house across 8 seeds (3W/2L/3T). Kept default off pending a size-threshold follow-up. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-07-26 09:31:42 +01:00 · 2026-07-26 09:31:42 +01:00 · 0d94e58119
commit 0d94e58119
parent bc11394499
8 changed files with 391 additions and 25 deletions
--- a/.beads/issues.jsonl
+++ b/.beads/issues.jsonl
--- a/DESIGN.md
+++ b/DESIGN.md
@ -2801,3 +2801,97 @@ uniform weight (not present in `_MUTATION_WEIGHTS`), matching pre-`lj3` behaviou
 (`--bridge-circulation`/`HOMEMAKER_BRIDGE_CIRCULATION`) for anyone who wants the connectivity-
 targeting behaviour despite the neutral aggregate measurement, but is not a candidate for a default
 flip on the current evidence.
+
+## 23. Ruin-and-recreate LNS: rebuild a wing with the adjacency-aware constructor (`homemaker-py-f1d`) — DONE (positive, size-dependent)
+
+**Motivation.** DESIGN.md's own experiment log by this point is one-sided: every "search machinery"
+change tried (§11.5 niching+restarts, §11.4 graded objective, §12.3 Wong-Liu reassociation +
+shape-feasibility, §12.4 granularity, §14 island model, §16 grain annealing, §18 graded
+connectivity, §19 shape repair, §21/§22 circulation-repair ops) has come back null-to-negative,
+while construction/seeding QUALITY (§11.6/§11.7 adjacency-aware seeding, §12.2 proportion-aware
+seeding) is the only lever that has ever moved the fail count. `operators._assign_adjacency_aware`
+— the constructor behind both `constructive_topology` and `lift_base_to_storeys` — currently only
+ever runs once, at seeding. The proposal: reuse it repeatedly DURING search as a large-neighbourhood-
+search (LNS) ruin-and-recreate move, betting that the one technique with a real track record
+generalises better than another new comparator-key or population-management idea.
+
+**Mechanism (build).** `operators.mutate_ruin_recreate`: pick a divided, live-cut subtree ("wing")
+of one storey holding a genuine partial neighbourhood of that storey's leaves (>=2, <= half — not a
+single-leaf relabel already covered by `retype`/`swap`, not a whole-floor rebuild already covered by
+the initial seed), un-divide it back to one leaf, then regrow and retype it with
+`_assign_adjacency_aware`, seeded (`fixed_circ`) from whichever already-typed circulation leaves
+border the wing — the same mechanism `lift_base_to_storeys` uses to grow an upper storey off an
+inherited core (§11.7), so the rebuilt interior spine reconnects to the surviving one instead of
+growing a disconnected island. The wing's required-space room-code budget is preserved exactly
+(same multiset); only its internal circulation/outside counts and split are rebuilt, at the same
+circ_divisor=3/outside_divisor=3 ratio the constructive seeders default to (not threaded from the
+run config — kept parameter-light, like `bridge_circulation`).
+
+`_assign_adjacency_aware` gained a new `scope` parameter (leaves eligible for retyping; `fixed_circ`
+may then name border leaves OUTSIDE `scope` as dominating-set seeds only, never retyped) so the wing
+rebuild can share the exact constructor code without touching the rest of the storey. `scope=None`
+(every existing caller) reproduces the prior unrestricted behaviour exactly — verified no other
+caller's output changed. Gated like `reassociate`/`bridge_circulation`: zero mutation weight unless
+`enable_ruin_recreate=True` (`driver.search`/`search_staged`, `evolve.py
+--ruin-recreate`/`HOMEMAKER_RUIN_RECREATE`, default off).
+
+**Verified (build-time).** 200 applications of `mutate_ruin_recreate` chained onto fresh
+`constructive_topology` harbor-house seeds (40 seeds × 5 steps): zero missing-space regressions
+(`graph.check_space_counts`), every child a canonical genome (`encode(decode(encode(x))) == encode(x)`).
+297 existing tests pass unchanged (the new op is exercised by the existing
+`test_mutations_yield_canonical_genomes` parametrization, which calls it with `reqs=None` and gets
+the documented noop). A `child_probe`-instrumented `driver.search` run confirmed the operator is
+actually selected by `mutate()` at its configured weight (not dead code).
+
+**Initial A/B (measured, 2026-07-25/26, qpk protocol) — NULL, but underpowered.** Equal-budget
+`enable_ruin_recreate` ON (implicit uniform mutation weight, ~7.5% draw probability among ~13 active
+ops) vs OFF, both arms finished with the standard finish-time `--collapse` (94g), 4 workers:
+
+- **harbor-house** (budget 2500, seeds 1–3): 1 loss (74→81), 2 ties.
+- **programme-house** (budget 3000, seeds 1–5): 4 ties, 1 win (9→8).
+- **Combined: 1 win / 1 loss / 6 ties out of 8**, mean fails 31.9 (OFF) → 32.6 (ON) — indistinguishable
+  from zero, in the same direction as most of this log's other null results.
+- A direct `child_probe` instrumentation of one of the tied harbor-house runs found
+  `ruin_recreate` fired **once in 32 children** — the initial sample is dominated by trajectories
+  where the operator simply never got a turn, not by turns it lost. Six of the eight exact ties
+  (fitness scalar identical to 6 significant figures, not just fail count) are consistent with
+  this: the op's rare draws mostly didn't survive tournament selection into the recorded lineage.
+
+**Weight follow-up (measured, 2026-07-26) — reran the ON arm only** with
+`_MUTATION_WEIGHTS["ruin_recreate"] = 3.0` (matching `place_missing`, mirroring the `lj3` weight-bump
+precedent) at the same seeds/budgets, directly comparable to the existing OFF baseline:
+
+- **programme-house** (seeds 1–5): **4 wins, 1 tie, 0 losses** — 7→1, 9→7, 9→8, 9→7, 5→5. A striking,
+  one-sided result, including one seed dropping from 7 fails to 1 (verified deterministic on rerun).
+- **harbor-house** (seeds 1–3): 1 win (77→73), 1 loss (74→82), 1 tie — still mixed.
+
+**Larger-N confirmation (measured, 2026-07-26)** — extended both arms to 10 fresh programme-house
+seeds (6–15) and 5 fresh harbor-house seeds (4–8) at the same weight=3.0, same protocol:
+
+- **programme-house, all 15 seeds combined: 8 wins / 1 loss / 6 ties.** Mean fails **7.07 (OFF) →
+  6.00 (ON)**, a ~15% reduction. Wilcoxon signed-rank p≈0.041; sign-test p≈0.020 (one-sided) — holds
+  up at conventional significance, not small-sample noise around zero (the 8sh/1ph/qi6/lj3 pattern
+  this log warns about).
+- **harbor-house, all 8 seeds combined: 3 wins / 2 losses / 3 ties.** Mean fails **73.0 (OFF) → 74.5
+  (ON)** — no consistent effect, if anything a very slight negative lean, echoing §20's
+  (`collapse_insearch`) opposite-direction size split but with the SMALLER building this time as
+  the one that benefits.
+
+**Interpretation.** A rare case in this log where a search-machinery idea shows a real,
+statistically-supported effect — but only on the smaller/simpler example programme. Plausible
+reading: programme-house's smaller room count means a wing rebuild samples a much larger fraction of
+the whole floor's topology per move (higher effective locality-vs-scope ratio), so the constructor's
+proven adjacency-aware placement quality dominates; harbor-house's much larger room count means the
+same wing size is a small, noisier perturbation relative to the whole building, and correlates with
+the ~2× per-op cost of `_assign_adjacency_aware` (leaf-graph rebuild + dominating-set search) not
+translating into more useful search steps within the same eval budget on that scale.
+
+**Status (2026-07-26).** `enable_ruin_recreate` stays **default OFF** — harbor-house shows no
+benefit and the two example programmes disagree on direction, so flipping the global default is not
+supported by this evidence (same conservative bar §20 applied before its own larger-N confirmation).
+`_MUTATION_WEIGHTS["ruin_recreate"] = 3.0` is kept in the source (only takes effect when the flag is
+on) since it is the validated-effective setting. `--ruin-recreate`/`HOMEMAKER_RUIN_RECREATE` is
+documented and ready to use today on programme-house-scale (smaller/simpler) programmes; a natural
+follow-up (not filed, low priority) would be a third or fourth example programme at a size between
+the two tested here, to locate the size threshold this result implies rather than inferring it from
+just two data points.
--- a/experiments/run_f1d_ab.sh
+++ b/experiments/run_f1d_ab.sh
@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+# f1d A/B: does the ruin-and-recreate LNS move (un-divide one wing of a
+# storey, rebuild it with the adjacency-aware constructor seeded from the
+# surviving circulation bordering the wing) reduce the fail count relative
+# to the current baseline (small local mutation operators only)? Same qpk
+# protocol as 8sh/qi6: equal-budget 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.
+#
+# Authoritative metrics: total fail count read from the .fails file
+# homemaker-fitness writes. Each run appends one TSV row so partial results
+# survive an interrupt.
+#
+# Usage: experiments/run_f1d_ab.sh
+set -u
+cd "$(dirname "$0")/.."
+
+WORKERS=4
+OUT=scratch/f1d_ab; mkdir -p "$OUT"
+TSV=scratch/f1d_ab_results.tsv
+[ -f "$TSV" ] || printf 'programme\tseed\truin\tbudget\tfails\tfitness\telapsed_s\n' > "$TSV"
+
+run() {  # programme seed ruin(0|1) budget
+  local prog="$1" seed="$2" rr="$3" budget="$4"
+  local tag="rr${rr}"
+  local dom="$OUT/${prog}_${tag}_s${seed}.dom"
+  local log="$OUT/${prog}_${tag}_s${seed}.log"
+  local flag="--no-ruin-recreate"; [ "$rr" = 1 ] && flag="--ruin-recreate"
+  echo ">>> $prog seed=$seed ruin_recreate=$rr budget=$budget"
+  local t0; t0=$(date +%s)
+  homemaker-evolve "examples/$prog/init.dom" \
+    --budget "$budget" --workers "$WORKERS" --seed "$seed" \
+    $flag --output "$dom" > "$log" 2>&1
+  local t1; t1=$(date +%s)
+  local fitness fails
+  fitness=$(sed -n 's/^best *: \([0-9.e+-]*\) .*/\1/p' "$log")
+  fails=$(sed -n 's/^best *: [0-9.e+-]* (\([0-9]*\) fails).*/\1/p' "$log")
+  ( cd "examples/$prog" && homemaker-fitness "$(realpath "../../$dom")" > /dev/null 2>&1 )
+  printf '%s\t%s\t%s\t%s\t%s\t%s\t%s\n' \
+    "$prog" "$seed" "$rr" "$budget" "${fails:-ERR}" "${fitness:-ERR}" "$((t1-t0))" >> "$TSV"
+  echo "    -> ${fails:-ERR} fails, fitness=${fitness:-ERR}, $((t1-t0))s"
+}
+
+# harbor-house: budget 2500, seeds 1-3 (qpk protocol)
+for seed in 1 2 3; do run harbor-house "$seed" 0 2500; done
+for seed in 1 2 3; do run harbor-house "$seed" 1 2500; done
+
+# programme-house: budget 3000, seeds 1-5 (qpk protocol)
+for seed in 1 2 3 4 5; do run programme-house "$seed" 0 3000; done
+for seed in 1 2 3 4 5; do run programme-house "$seed" 1 3000; done
+
+echo "=== f1d ruin_recreate A/B complete ==="
+column -t -s $'\t' "$TSV"
--- a/experiments/run_f1d_larger_n.sh
+++ b/experiments/run_f1d_larger_n.sh
@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# f1d larger-N confirmation: the weight=3.0 follow-up (run_f1d_w3_ab.sh)
+# showed a striking, consistent programme-house improvement (4 wins + 1 tie,
+# 0 losses, mean fails 8.6 -> 5.75 across seeds 1-5) but a mixed harbor-house
+# result (1 win, 1 tie, 1 loss). Extends BOTH arms to seeds 6-15 on
+# programme-house and 4-8 on harbor-house so the N=5/N=3 initial reads are
+# confirmed or falsified on fresh seeds, mirroring the project's own
+# 1ph/qjg larger-N-confirmation pattern. driver._MUTATION_WEIGHTS still
+# carries the temporary "ruin_recreate": 3.0 from the w3 follow-up.
+#
+# Usage: experiments/run_f1d_larger_n.sh
+set -u
+cd "$(dirname "$0")/.."
+
+WORKERS=4
+OUT=scratch/f1d_ln; mkdir -p "$OUT"
+TSV=scratch/f1d_ln_results.tsv
+[ -f "$TSV" ] || printf 'programme\tseed\truin\tbudget\tfails\tfitness\telapsed_s\n' > "$TSV"
+
+run() {  # programme seed ruin(0|1) budget
+  local prog="$1" seed="$2" rr="$3" budget="$4"
+  local tag="rr${rr}"
+  local dom="$OUT/${prog}_${tag}_s${seed}.dom"
+  local log="$OUT/${prog}_${tag}_s${seed}.log"
+  local flag="--no-ruin-recreate"; [ "$rr" = 1 ] && flag="--ruin-recreate"
+  echo ">>> $prog seed=$seed ruin_recreate=$rr budget=$budget"
+  local t0; t0=$(date +%s)
+  homemaker-evolve "examples/$prog/init.dom" \
+    --budget "$budget" --workers "$WORKERS" --seed "$seed" \
+    $flag --output "$dom" > "$log" 2>&1
+  local t1; t1=$(date +%s)
+  local fitness fails
+  fitness=$(sed -n 's/^best *: \([0-9.e+-]*\) .*/\1/p' "$log")
+  fails=$(sed -n 's/^best *: [0-9.e+-]* (\([0-9]*\) fails).*/\1/p' "$log")
+  ( cd "examples/$prog" && homemaker-fitness "$(realpath "../../$dom")" > /dev/null 2>&1 )
+  printf '%s\t%s\t%s\t%s\t%s\t%s\t%s\n' \
+    "$prog" "$seed" "$rr" "$budget" "${fails:-ERR}" "${fitness:-ERR}" "$((t1-t0))" >> "$TSV"
+  echo "    -> ${fails:-ERR} fails, fitness=${fitness:-ERR}, $((t1-t0))s"
+}
+
+# programme-house: seeds 6-15, both arms
+for seed in $(seq 6 15); do run programme-house "$seed" 0 3000; done
+for seed in $(seq 6 15); do run programme-house "$seed" 1 3000; done
+
+# harbor-house: seeds 4-8, both arms
+for seed in $(seq 4 8); do run harbor-house "$seed" 0 2500; done
+for seed in $(seq 4 8); do run harbor-house "$seed" 1 2500; done
+
+echo "=== f1d larger-N confirmation complete ==="
+column -t -s $'\t' "$TSV"
--- a/experiments/run_f1d_w3_ab.sh
+++ b/experiments/run_f1d_w3_ab.sh
@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+# f1d weight follow-up: the initial run_f1d_ab.sh A/B (uniform uneabled weight,
+# ~7.5% activation among active ops) came back essentially null (1 win, 1
+# loss, 6 ties out of 8 paired seeds) with a directly-instrumented run showing
+# ruin_recreate fired only ~1/32 children -- likely underpowered rather than
+# a clean negative. driver._MUTATION_WEIGHTS now carries a temporary
+# "ruin_recreate": 3.0 (matching place_missing's weight, mirroring the lj3
+# weight-bump precedent) to raise the activation rate; this reruns the ON arm
+# only at the SAME seeds/budgets as run_f1d_ab.sh so it is directly
+# comparable against the existing rr=0 baseline rows in
+# scratch/f1d_ab_results.tsv.
+#
+# Usage: experiments/run_f1d_w3_ab.sh
+set -u
+cd "$(dirname "$0")/.."
+
+WORKERS=4
+OUT=scratch/f1d_w3_ab; mkdir -p "$OUT"
+TSV=scratch/f1d_w3_ab_results.tsv
+[ -f "$TSV" ] || printf 'programme\tseed\truin\tweight\tbudget\tfails\tfitness\telapsed_s\n' > "$TSV"
+
+run() {  # programme seed budget
+  local prog="$1" seed="$2" budget="$3"
+  local dom="$OUT/${prog}_rr1w3_s${seed}.dom"
+  local log="$OUT/${prog}_rr1w3_s${seed}.log"
+  echo ">>> $prog seed=$seed ruin_recreate=1 weight=3.0 budget=$budget"
+  local t0; t0=$(date +%s)
+  homemaker-evolve "examples/$prog/init.dom" \
+    --budget "$budget" --workers "$WORKERS" --seed "$seed" \
+    --ruin-recreate --output "$dom" > "$log" 2>&1
+  local t1; t1=$(date +%s)
+  local fitness fails
+  fitness=$(sed -n 's/^best *: \([0-9.e+-]*\) .*/\1/p' "$log")
+  fails=$(sed -n 's/^best *: [0-9.e+-]* (\([0-9]*\) fails).*/\1/p' "$log")
+  ( cd "examples/$prog" && homemaker-fitness "$(realpath "../../$dom")" > /dev/null 2>&1 )
+  printf '%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\n' \
+    "$prog" "$seed" "1" "3.0" "$budget" "${fails:-ERR}" "${fitness:-ERR}" "$((t1-t0))" >> "$TSV"
+  echo "    -> ${fails:-ERR} fails, fitness=${fitness:-ERR}, $((t1-t0))s"
+}
+
+for seed in 1 2 3; do run harbor-house "$seed" 2500; done
+for seed in 1 2 3 4 5; do run programme-house "$seed" 3000; done
+
+echo "=== f1d weight=3.0 follow-up complete ==="
+column -t -s $'\t' "$TSV"
--- a/src/homemaker_layout/driver.py
+++ b/src/homemaker_layout/driver.py
@ -103,7 +103,8 @@ def _reqs_for(programme_dir: str) -> dict:
 # (homemaker-py-qjg, DESIGN.md §22) found no total-fail benefit and MORE
 # trajectory-divergence-induced new not-connected fails than at the
 # uniform default weight -- reverted, left at implicit uniform weight.
-_MUTATION_WEIGHTS = {"level_add": 0.2, "level_delete": 0.2, "place_missing": 2.0}
+_MUTATION_WEIGHTS = {"level_add": 0.2, "level_delete": 0.2, "place_missing": 2.0,
+                     "ruin_recreate": 3.0}


 def _worker_init() -> None:
@ -241,6 +242,7 @@ def search(
    enable_reassociate: bool = False,
    enable_shape_repair: bool = False,
    enable_bridge_circulation: bool = False,
+    enable_ruin_recreate: bool = False,
    feasibility_filter: bool = False,
    feasibility_max_shape_fails: int | None = None,
    circ_divisor: int = 3,
@ -323,6 +325,16 @@ def search(
    because it needs no ``fitness.Fitness`` instance — only the tree's own
    adjacency graph — so it is otherwise unconditionally live once landed in
    ``operators.MUTATIONS``.
+
+    ``enable_ruin_recreate`` (homemaker-py-f1d, EXPERIMENTAL, default off) un-
+    mutes ``operators.mutate_ruin_recreate``: a large-neighbourhood-search move
+    that un-divides one wing of a storey and rebuilds it with the same
+    adjacency-aware constructor the seeders use (``operators.
+    _assign_adjacency_aware``, seeded from the surviving circulation bordering
+    the wing), instead of relying only on the small local mutation operators to
+    discover an improving rearrangement. Gated like ``reassociate`` (zero
+    mutation weight unless enabled) — it needs only ``reqs``, no
+    ``fitness.Fitness`` instance.
    """
    from .oracle import DEFAULT_URB_ROOT

@ -337,6 +349,8 @@ def search(
        mutation_weights["reassociate"] = 0.0
    if not enable_bridge_circulation:
        mutation_weights["bridge_circulation"] = 0.0
+    if not enable_ruin_recreate:
+        mutation_weights["ruin_recreate"] = 0.0
    # homemaker-py-161: shape_rotate/deslim are gated by operators.mutate itself
    # (fit_ops go to zero probability when fit=None) — only build the Fitness
    # instance, and thus only let them fire, when explicitly enabled.
@ -936,6 +950,7 @@ def search_staged(
    enable_reassociate: bool = False,
    enable_shape_repair: bool = False,
    enable_bridge_circulation: bool = False,
+    enable_ruin_recreate: bool = False,
    feasibility_filter: bool = False,
    feasibility_max_shape_fails: int | None = None,
    circ_divisor: int = 3,
@ -993,6 +1008,7 @@ def search_staged(
                      enable_reassociate=enable_reassociate,
                      enable_shape_repair=enable_shape_repair,
                      enable_bridge_circulation=enable_bridge_circulation,
+                      enable_ruin_recreate=enable_ruin_recreate,
                      feasibility_filter=feasibility_filter,
                      feasibility_max_shape_fails=feasibility_max_shape_fails,
                      circ_divisor=circ_divisor,
@ -1031,6 +1047,7 @@ def search_staged(
            enable_reassociate=enable_reassociate,
            enable_shape_repair=enable_shape_repair,
            enable_bridge_circulation=enable_bridge_circulation,
+            enable_ruin_recreate=enable_ruin_recreate,
            feasibility_filter=feasibility_filter,
            feasibility_max_shape_fails=feasibility_max_shape_fails,
            circ_divisor=circ_divisor,
@ -1078,6 +1095,7 @@ def search_staged(
        enable_reassociate=enable_reassociate,
        enable_shape_repair=enable_shape_repair,
        enable_bridge_circulation=enable_bridge_circulation,
+        enable_ruin_recreate=enable_ruin_recreate,
        feasibility_filter=feasibility_filter,
        feasibility_max_shape_fails=feasibility_max_shape_fails,
        circ_divisor=circ_divisor,
--- a/src/homemaker_layout/evolve.py
+++ b/src/homemaker_layout/evolve.py
@ -114,6 +114,16 @@ def _parse_args(argv=None) -> argparse.Namespace:
                        "to circulation, directly clearing a 'level N not "
                        "connected' fail instead of relying on the qi6 graded "
                        "comparator key (measured negative, §18) (default: off)")
+    p.add_argument("--ruin-recreate", dest="ruin_recreate",
+                   action=argparse.BooleanOptionalAction,
+                   default=_env_bool("HOMEMAKER_RUIN_RECREATE", False),
+                   help="homemaker-py-f1d: large-neighbourhood-search repair "
+                        "mutation that un-divides one wing of a storey and "
+                        "rebuilds it with the adjacency-aware constructor "
+                        "(seeded from the surviving circulation bordering the "
+                        "wing), applying the one construction technique with a "
+                        "track record repeatedly during search instead of only "
+                        "at seeding (default: off)")
    p.add_argument("--collapse-insearch", dest="collapse_insearch",
                   action=argparse.BooleanOptionalAction,
                   default=_env_bool("HOMEMAKER_COLLAPSE_INSEARCH", True),
@ -192,6 +202,7 @@ def main(argv=None) -> int:
    print(f"superpose    : {args.superpose}", file=sys.stderr)
    print(f"conn grade   : {args.conn_grade}", file=sys.stderr)
    print(f"bridge circulation : {args.bridge_circulation}", file=sys.stderr)
+    print(f"ruin recreate      : {args.ruin_recreate}", 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)

@ -243,6 +254,7 @@ def main(argv=None) -> int:
            superpose=args.superpose,
            conn_grade=args.conn_grade,
            enable_bridge_circulation=args.bridge_circulation,
+            enable_ruin_recreate=args.ruin_recreate,
            collapse_insearch=args.collapse_insearch,
            log=lambda m: print(m, file=sys.stderr, flush=True),
        )
--- a/src/homemaker_layout/operators.py
+++ b/src/homemaker_layout/operators.py
@ -849,7 +849,8 @@ def _assign_adjacency_aware(lvl: dom.Node, room_codes: list[str], reqs,
                            rng: np.random.Generator, door_width: float = 1.2,
                            fixed_circ: "list[dom.Node] | None" = None,
                            interior_outside: bool = False,
-                            n_outside: int = 1) -> None:
+                            n_outside: int = 1,
+                            scope: "set[dom.Node] | None" = None) -> None:
    """Assign leaf types so rooms cluster around a connected circulation spine.

    s44 (DESIGN.md §11.2 follow-up): random type assignment leaves rooms stranded
@ -870,13 +871,23 @@ def _assign_adjacency_aware(lvl: dom.Node, room_codes: list[str], reqs,
    ``lvl`` already has the right number of leaves grown; their types are
    (re)written in place. Stochastic where it is free (room order, tie-breaks) so
    a bootstrap batch stays diverse.
+
+    ``scope`` (homemaker-py-f1d): restrict retyping to this subset of ``lvl``'s
+    leaves — used by the ruin-and-recreate LNS move to rebuild one wing of an
+    already-typed storey in place. ``fixed_circ`` may then name leaves OUTSIDE
+    ``scope`` (the surviving circulation bordering the wing) purely as
+    dominating-set seeds; they anchor the spine but are never retyped, and the
+    dominating-set growth and room/outside placement only ever touch ``scope``.
+    ``None`` (default) reproduces the unrestricted whole-``lvl`` behaviour
+    exactly — every existing caller is unaffected.
    """
    from . import geometry

    reqs = reqs or {}
    leaves = lvl.leaves()
-    n = len(leaves)
    idx = {leaf: i for i, leaf in enumerate(leaves)}
+    assignable = scope if scope is not None else set(leaves)
+    n = len(assignable)
    R = len(room_codes)
    n_circ = max(1, n - (R + max(1, n_outside)))  # leftover after rooms + outside
    seeds = [c for c in (fixed_circ or []) if c in idx]
@ -893,16 +904,19 @@ def _assign_adjacency_aware(lvl: dom.Node, room_codes: list[str], reqs,

    # Greedy connected dominating set of size n_circ: seed from the fixed core (or
    # the most central leaf), then repeatedly add the frontier leaf that newly
-    # dominates the most leaves (keeping the set connected).
-    circ = set(seeds) if seeds else {max(leaves, key=lambda L: (deg.get(L, 0), -idx[L]))}
+    # dominates the most leaves (keeping the set connected). Growth is confined to
+    # ``assignable`` so a scoped call never annexes a leaf outside the wing.
+    circ = (set(seeds) if seeds
+            else {max(assignable, key=lambda L: (deg.get(L, 0), -idx[L]))})
    dominated = set().union(*( _nbrs(s) | {s} for s in circ))
    while len(circ) < n_circ:
-        frontier = (set().union(*(_nbrs(s) for s in circ)) - circ) if circ else set()
+        frontier = ((set().union(*(_nbrs(s) for s in circ)) - circ) & assignable
+                    if circ else set())
        if frontier:
            pick = max(frontier, key=lambda L: (len(_nbrs(L) - dominated),
                                                deg.get(L, 0), -idx[L]))
        else:  # disconnected remainder — seed a new component by degree
-            rest = [L for L in leaves if L not in circ]
+            rest = [L for L in assignable if L not in circ]
            if not rest:
                break
            pick = max(rest, key=lambda L: (deg.get(L, 0), -idx[L]))
@ -910,9 +924,10 @@ def _assign_adjacency_aware(lvl: dom.Node, room_codes: list[str], reqs,
        dominated |= _nbrs(pick) | {pick}

    for s in circ:
+        if s in assignable:  # never retype a fixed_circ seed outside scope
            s.type = "C"

-    noncirc = [L for L in leaves if L not in circ]
+    noncirc = [L for L in assignable if L not in circ]
    if interior_outside:
        # ld2 (§13.6): seed ``O`` as INTERIOR light wells instead of one
        # peripheral leaf. A landlocked room (no plot facade, no uncovered-O
@ -1196,6 +1211,78 @@ def lift_base_to_storeys(base_root: dom.Node, upper_buckets: list[dict[str, int]
    return _finalise(child)


+def mutate_ruin_recreate(root: dom.Node, rng: np.random.Generator,
+                         types: list[str], reqs=None) -> tuple[dom.Node, str]:
+    """LNS ruin-and-recreate: rebuild one wing of a storey with the constructor.
+
+    homemaker-py-f1d (DESIGN.md's experiment log): every "search machinery"
+    change tried so far (niching+restarts, graded objective, Wong-Liu
+    reassociation, granularity, island model, grain annealing, circulation-
+    repair ops) has come back null-to-negative, while construction/seeding
+    quality (adjacency-aware seeding, proportion-aware seeding) is the only
+    lever that has ever moved the fail count. ``_assign_adjacency_aware``
+    currently only runs once, at seeding. This move reuses it repeatedly
+    during search: pick a divided, live-cut subtree ("wing") of one storey
+    holding a genuine partial neighbourhood of that storey's leaves (at least
+    2, at most half), un-divide it back to a single leaf, then regrow and
+    retype it with the same adjacency-aware constructor the seeders use —
+    seeded (``fixed_circ``) from whichever already-typed circulation leaves
+    border the wing, exactly the mechanism ``lift_base_to_storeys`` uses to
+    grow an upper storey off an inherited core (ld5, §11.7), so the rebuilt
+    interior spine reconnects to the surviving one instead of growing a
+    disconnected island.
+
+    The wing's programme room-code budget (the multiset of required-space
+    types already inside it) is preserved exactly; only its internal
+    circulation/outside counts and split are rebuilt, at the same
+    circ_divisor=3/outside_divisor=3 ratio the constructive seeders default
+    to (not threaded from the run config — an experimental repair op, like
+    ``bridge_circulation``, kept parameter-light).
+    """
+    if not reqs:
+        return _finalise(copy.deepcopy(root)), "ruin_recreate noop"
+    from . import geometry
+
+    child = copy.deepcopy(root)
+    _finalise(child)
+    lvls = dom.levels(child)
+    totals = {li: len(lvl.leaves()) for li, lvl in enumerate(lvls)}
+    cands = [(li, n) for li, n in _owned_branches(child)
+             if totals[li] >= 4 and 2 <= len(n.leaves()) <= max(2, totals[li] // 2)]
+    if not cands:
+        return _finalise(child), "ruin_recreate noop"
+    li, wing = _pick(rng, cands)
+    lvl = lvls[li]
+
+    G = geometry.leaf_graph(lvl)
+    wing_leaves = set(wing.leaves())
+    border_circ = sorted(
+        {nb for lf in wing_leaves for nb in G.neighbors(lf)
+         if nb not in wing_leaves and nb.type and nb.type[0].lower() == "c"},
+        key=lambda n: n.id or "")
+
+    rooms = [lf.type for lf in wing.leaves() if lf.type in reqs]
+    n_circ_total = max(1, -(-len(rooms) // 3))  # circ_divisor=3
+    n_o = max(1, round(len(rooms) / 3))  # outside_divisor=3
+    n_new = len(rooms) + n_o + max(0, n_circ_total - len(border_circ))
+
+    wing.left = wing.right = None
+    wing.division = None
+    wing.type = None
+    _grow_leaves(wing, max(1, n_new), rng, balance=True)
+    dom._link(child)
+
+    _assign_adjacency_aware(
+        lvl, rooms, reqs, rng, fixed_circ=border_circ or None,
+        interior_outside=True, n_outside=n_o, scope=set(wing.leaves()))
+    dom._link(child)
+    _size_divisions_from_targets(wing, reqs)
+
+    return _finalise(child), (
+        f"ruin_recreate {li}/{wing.id or 'root'} "
+        f"({len(rooms)} rooms, {len(border_circ)} anchors)")
+
+
 def mutate_reassociate(root: dom.Node, rng: np.random.Generator,
                       types: list[str]) -> tuple[dom.Node, str]:
    """Wong-Liu M3 associativity move: ``(a|b)|c <-> a|(b|c)`` on parallel cuts.
@ -1418,6 +1505,7 @@ MUTATIONS = {
    "level_delete": mutate_level_delete,
    "shape_rotate": mutate_shape_rotate,
    "deslim": mutate_deslim,
+    "ruin_recreate": mutate_ruin_recreate,
 }


@ -1435,7 +1523,7 @@ def mutate(root: dom.Node, rng: np.random.Generator, types: list[str],
    names = sorted(MUTATIONS)
    p = np.array([(weights or {}).get(n, 1.0) for n in names], dtype=float)
    # these operators need programme reqs; disable them when not available
-    reqs_ops = ("level_fix", "level_compound_fix", "place_missing")
+    reqs_ops = ("level_fix", "level_compound_fix", "place_missing", "ruin_recreate")
    # also takes reqs (to avoid displacing a required room) but works without
    # it — never zero-weighted, unlike reqs_ops above
    reqs_optional_ops = ("bridge_circulation",)