fix: rebuild_profiles hermeticity + bridge_finder ValueError on real session data (#174)

[cdeust]

Closes #174

The defect, two layers

1. Real bug (production data path). scanner.discover_all_memories() returns a list[dict] of memory records, but bridge_finder.find_bridges fed that list straight into dict.update. A non-empty list of 9-key records raises:

ValueError: dictionary update sequence element #0 has length 9; 2 is required

Empty homes skip the if memories: branch, so CI (empty $HOME) stayed green while every developer machine with real ~/.claude memory files failed. This is a type-contract mismatch reached by legitimate production data — classified as a real bug, not flakiness.

2. Test isolation (hidden shared state). Three test_rebuild_profiles cases called handler({"force": True}) with no mocking, reading the developer's live ~/.claude session history. Results depended on whose machine ran them.

Minimized failing input

A single scanner-shaped record (the 9 keys file, path, project, name, description, type, body, modifiedAt, createdAt) wrapped in a list:

find_bridges(profiles, None, [{"file": "note.md", "path": "...", "project": "proj-a",
    "name": "note", "description": "", "type": "note", "body": "",
    "modifiedAt": "...", "createdAt": "..."}])   # -> ValueError pre-fix

Root-cause classification

bridge_finder should handle that shape — it is exactly what the production producer (discover_all_memories) emits. Fixed there, not by rejecting the data.

What changed

• core/bridge_finder.py: new _index_memories() normalizes the scanner list into an id-keyed map (keyed by path/file/name); the brain_index dict shape is passed through unchanged. Signature widened to dict | list | None.
• core/profile_assembler.py: annotations corrected to the honest dict | list[dict] | None through the call chain.
• tests_py/core/test_bridge_finder.py: TestFindBridgesScannerListShape pins the minimized 9-key record and the list+brain_index merge.
• tests_py/handlers/test_rebuild_profiles.py: hermetic_claude_home fixture redirects every ~/.claude path constant at a synthetic tmp_path seeded with real sessions + memory files; the three previously-live-home tests now drive the full pipeline against fixtures and consume the list-of-records shape that triggered the ValueError.

Gates (on this machine, live ~/.claude present)

• ruff check + ruff format --check: pass.
• Full pytest: 5304 passed, 9 skipped. The 3 previously-failing test_rebuild_profiles cases are green. Two unrelated pre-existing failures (test_cold_start real-DB SessionStart hook; test_I2_canonical_writer order-dependent, passes in isolation) reproduce on clean main without this change and are out of scope for test: test_rebuild_profiles depends on live ~/.claude data — 3 failures reproduce on main #174.

🤖 Generated with Claude Code