fix(witness): §S74 candidate 1 — revive room-walker witness trio + re-lock geomap baseline

[red1oon]

Executes prompts/SCRIPT_LENGTH_REFACTOR_SEAMS.md §S74's priority-1 verdict: revive the dead witness locks before any structural refactor. Witness-side only — zero product lines changed. Full evidence trail appended as §S74's 2026-08-23 addendum in the same file.

Fix 1 — the dead LIVEWIRE front door

witness_room_walker_parity.js / witness_room_fill.js / witness_room_wellformed.js had front-doored on the pruned /tmp/wt-fable-livewire/modeller since that worktree was removed — every run printed all-SKIP pass=0 fail=0, exit 0 (green-but-vacuous, §S61.3 class B).

• LIVEWIRE now resolves: ARC_DB_DIR env override → ~/bim-ootb/modeller → the old /tmp path (first existing dir wins); per-building SKIP kept for genuinely absent DBs.
• Hardcoded absolute sql.js requires → the loadSqlJs() candidate loop already used by witness_disc_walk_shim.js (§S74 landmine 6).
• Parity's hardcoded /home/red1/bim-compiler/scripts/compile_rooms.py exec → ROOT-relative (it was silently testing the PRIMARY checkout's py from any worktree).
• KNOWN_TYPES += SUSPECT_ELONGATED/SUSPECT_LARGE (specced latent-misfire fix).

Before/after: all three all-SKIP vacuous → parity pass=6 fail=0 (6/6 buildings byte-identical on spatial_structure + rel_contained_in_space; totals 51/33/207/6/201/45), fill pass=18 fail=0, wellformed pass=19 fail=0, all zero-SKIP.

Revival exposed two witness assumptions gone stale during the dead-gate window (product moved 2026-07-12/13 while the witnesses stayed frozen at 2026-07-11; py+js remained byte-identical throughout, so these are test-side, not product defects):

• fill F2 "SUSPECT rooms have exactly ONE row": 7 multi-rect suspects fleet-wide, each explained by R-MERGE rect-concat (room_walker.js:990–998, merged_from 2–3) or R-REJECT's post-decompose SUSPECT_OPEN flag (:1063, enclosure 0.391 < 0.50). Re-locked to exempt exactly those two; unexplained multi-row suspects still FAIL.
• wellformed W3 "zero wall cells inside any rect": red on ALL SIX buildings because §WALL-SNAP (room_walker.js:432–467, 2026-07-13) puts every rect edge ON its wall's continuous near face. Depth-classified fleet-wide: 16,391 hits at depth ≤1 (the snap boundary band), ZERO at depth ≥2. Re-locked: band contact reported (snapBandHits=), never failed; depth ≥2 — a wall genuinely inside a room — still FAILs.

Fix 2 — stale geomap exact-lock

scripts/witness_geomap_tier3.py was RED on baseline_iou_recall == 1 while the flood-fill baseline measures 3/21 over 3 candidates (confirmed fresh before editing). Re-locked 1→3 with a dated comment; assertion shape untouched (exact-lock + BEATS strict inequality both kept). Now RESULT: GREEN, exit 0. The artifact's db_md5 line is re-recorded to the fixture DB's actual hash (DB rebuilt 2026-07-10 after the artifact was committed; all scored numbers re-verified identical).

Perturbation proof (revived gates can still go red)

• mergeRooms group order flipped to Object.keys() in a scratch copy → parity pass=4 fail=2 exit 1 (Clinic + Hospital same-count/wrong-room-per-guid — the :974–985 historical red re-demonstrated).
• One KNOWN_TYPES entry dropped → wellformed pass=14 fail=5 (badType 9/23/1/8/13).

No auto-merge — left for review per dispatch instructions.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GmNVMG9ZGcKygz4zXuCe6m