fix: re-key hallways/tunnels on wing-name migration so the graph follows renamed wings

[ggettert]

Summary

Fixes #1938. migrate_wing_names re-keyed drawer/closet/topics_by_wing metadata but never touched the associative graph, so hallways.json and tunnels.json kept pointing at the pre-normalization wing.

What changed

• hallways.rekey_hallway_wings(rename_map) — rewrites hallway wing + regenerates ids; merges records that collapse onto the same (wing, entity-pair).
• palace_graph.rekey_tunnel_wings(rename_map) — rewrites tunnel endpoint wings + regenerates canonical ids; drops tunnels whose endpoints coincide after the rename; merges id collisions. Covers entity and topic tunnels (shared store).
• migrate_wing_names calls both after the metadata re-key, using the same rename map, and reports counts.
• Extracted _hallway_label so freshly-computed and re-keyed/merged records format identically (no behavior change).

Why re-key, not recompute

A wing-name migration is a pure rename — drawer membership and entity co-occurrence are unchanged. Re-keying is sufficient and preserves the #1578 dynamics fields (strength/stability/last_activated/access_count); a recompute from drawers would reset them. Collision merges keep the stronger/more-recent dynamics.

Tests

• Unit: rename regenerates ids + preserves dynamics; collision merges counts/rooms/dynamics; untouched wings/tunnels left alone; self-referential tunnels dropped.
• Integration: seed _alpha+beta drawers, build hallways/tunnels, run migration → graph moves to alpha; two-wings-into-one merges to a single hallway with summed count.

Known limitations

• The re-key runs after the drawer/closet/topics_by_wing re-key without a shared transaction across the four stores. If the graph step fails mid-way, the drawer re-key is intentionally not rolled back (a successful data migration shouldn't be undone because a derived index failed); the failure is surfaced as a warning and the operation is re-runnable — a subsequent migrate re-keys any orphaned graph records. Full cross-store atomicity is out of scope for this fix.

Fixes #1938.

[ggettert]

Rebased this onto current develop after v3.6.0 and updated the graph re-key path to pass the explicit MempalaceConfig(palace_path=...) through to hallway/tunnel side-files. Local focused tests and full GitHub CI are green.