docs(refactor): §S74 — room_walker.js / compile_rooms.py refactor survey (spec only)

[red1oon]

Spec-only survey — no product code touched

Appends §S74 to prompts/SCRIPT_LENGTH_REFACTOR_SEAMS.md (base = fable/meshdb-livewire, where the canonical file lives; §S67 was already taken by 4D_GANTT_TM_REFACTOR.md, so the next free number in the shared sequence is §S74).

What was measured (all runs logged, Log Mandate)

• witness_room_walker_parity.js as-is: vacuous — SKIP ×6, pass=0 fail=0, exit 0 (dead /tmp/wt-fable-livewire path, the §S61.3 class-B bug). Same dead path in witness_room_fill.js / witness_room_wellformed.js.
• Scratch copy repointed at ~/bim-ootb/modeller: pass=6 fail=0, byte-identical py↔js on all six buildings — the port is still verbatim; only the lock's front door is dead.
• scripts/witness_geomap_tier3.py: RED today — stale exact-lock (baseline recall claimed 1/21, measures 3/21; the flood baseline improved and the lock was never re-recorded).
• Zero slice-based consumers of either file (no §S61.3-A rot possible); the API export map is the contract.
• Copy drift measured: 3 algorithm generations across 4 JS copies (+127 lines in deploy/dev/lib, +131 in bim-ootb viewer/lib — §CONTAINMENT-ALIAS join never ported to the py "ground truth" despite its own LOCKSTEP comment).

Verdict

• Top candidate = priority 1, NOT an extraction: revive the dead-path witness trio + re-lock the geomap baseline at the measured number (§S63 fix class).
• Copy-drift consolidation (room_walker_ext.js) is priority 2, blocked on candidate 1 per the file's governing rule (low witness coverage disqualifies).
• No line-count split of either file recommended — both are cohesive and already the model output-state (§S59 schedule_author.js / §S60 room_graph.js precedent).

No auto-merge — left for human review.

🤖 Generated with Claude Code