docs(skill): repair topology (doors/windows) from the access graph

[ChanMeng666]

Summary

Documents a verified agent-layer procedure for making every room reachable and every bedroom lit by adding doors/windows — the design choice the core deliberately won't make (a door's position is a choice among valid options; per ADR 0005 that belongs above compile()). No core code change.

arch repair fixes furniture; this closes the remaining gap — the topology faults (W_ROOM_UNREACHABLE, W_ROOM_DISCONNECTED, W_NO_ENTRANCE, W_BATH_VIA_BEDROOM, W_BEDROOM_NO_WINDOW) — by teaching the agent to drive off the facts describe --json already exposes.

The procedure (in SKILL.md, pointer in arch spec)

1. Read describe: access.rooms[].reachable, room bbox/uses/adjacent, building extent = min/max of room boxes (a room edge on it = exterior wall).
2. Connect each unreachable room — priority: exterior entrance into a cut-off living/kitchen > door on a shared wall to a reachable non-bedroom > never route a bath only through a bedroom.
3. Exact on-centerline at arithmetic for exterior doors, shared-wall doors, and bedroom windows.
4. Re-repair (furniture clears new swings/landings) → validate --strict until ok: true.
5. Addendum: an existing off-wall door/window (W_DOOR_OFF_WALL etc.) is a separate mis-coordinate — snap onto a wall or delete.

Verified end-to-end (two different topologies)

• Example B (bath+living cut off; bedroom held the only entrance) → exterior entrance into the living/kitchen + bedroom window → repair → validate --strict ok: true, fully clean.
• Example C (bath reachable; living+bed cut off) → exterior entrance + bed↔living door + bedroom window → all rooms reachable, 2 entrances (residual W_DOOR_OFF_WALL is a pre-existing generator mis-coordinate — hence the addendum).

This completes the loop: geometry-blind AI plan → arch repair (furniture) + access-graph door/window procedure → a sound, fully-validated drawing.

Verification: 466 tests still pass; no generated-file drift.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[github-actions]

⏱️ Benchmark (median ms, informational — not a gate)

Plan	Stage	Current	Baseline	Δ
BALANCED	compile	10.91	12.92	-16%
BALANCED	parse	0.09	0.08	+10%
BALANCED	resolve	0.00	0.00	—
BALANCED	render	8.11	12.83	-37%
ROOM_HEAVY	compile	2.58	4.14	-38%
ROOM_HEAVY	parse	0.10	0.09	+7%
ROOM_HEAVY	resolve	0.00	0.00	—
ROOM_HEAVY	render	2.33	4.05	-42%
OPENING_HEAVY	compile	16.83	26.11	-36%
OPENING_HEAVY	parse	0.08	0.07	+10%
OPENING_HEAVY	resolve	0.00	0.00	—
OPENING_HEAVY	render	16.72	23.21	-28%

No stage regressed past the warning threshold.