fix(hooks): guard fcntl import on Windows — _shared + session_start

[Huntehhh]

Summary

truememory/ingest/hooks/_shared.py and truememory/ingest/hooks/session_start.py
both called bare import fcntl at module top with no try / except ImportError
guard. fcntl is POSIX-only — on Windows every import raised
ModuleNotFoundError before any function ran. That cascaded to every Claude
Code hook subprocess that imports from _shared (SessionStart,
UserPromptSubmit, Stop), silently breaking the hook-based memory-extraction
pipeline on Windows.

Fix mirrors the established _HAS_FCNTL flag pattern already used by
ingest/pipeline.py, hooks/core.py, and ingest/hooks/user_prompt_submit.py.
Both call sites (_shared.check_extraction_budget and
session_start._scan_stale_sessions) now gate their fcntl.flock calls on
_HAS_FCNTL. POSIX behavior is unchanged; Windows falls through to the
non-locking path with documented worst-case race semantics (1-2 extra
extractions per hour for the budget check, deduplicated overlap for the stale
scan).

This is a sibling fix to the WNOHANG guard in the cold-start resilience PR
(if that one is open) — same POSIX-only-API class of bug, same hasattr/
try-except-guard pattern.

Changes

File	Change
truememory/ingest/hooks/_shared.py	Bare import fcntl → try / except ImportError + _HAS_FCNTL flag; check_extraction_budget guards fcntl.flock(LOCK_EX) on the flag; documented fallback semantics
truememory/ingest/hooks/session_start.py	Bare import fcntl → same _HAS_FCNTL guard; _scan_stale_sessions guards fcntl.flock(LOCK_EX | LOCK_NB) on the flag with inline comment explaining the dedup path
tests/ingest/test_hooks_windows_portability.py	New file (8 tests): _HAS_FCNTL flag contract on both modules, check_extraction_budget behaves correctly without fcntl (allows, enforces cap, resets hourly), POSIX-only regression lock that fcntl.flock(LOCK_EX) is still acquired on platforms where it's available
CHANGELOG.md	Unreleased entry documenting the bug, fix, and test coverage

Test Plan

• python -m pytest tests/ingest/test_hooks_windows_portability.py -v → 7 passed + 1 skipped on Windows (the POSIX regression lock); 8 passed on POSIX
• On Windows: python -c "from truememory.ingest.hooks import _shared, session_start, user_prompt_submit; print('OK')" → exits 0
• Trigger a Claude Code SessionStart hook on Windows and confirm no ModuleNotFoundError: No module named 'fcntl' in stderr — the hook should run to completion and inject memory context
• On POSIX, spawn two concurrent python -c "from truememory.ingest.hooks._shared import check_extraction_budget; print(check_extraction_budget())" and confirm the budget counter increments correctly (i.e. the lock is still doing its job)

Design Notes

• Why fall back to non-atomic read-modify-write on Windows instead of a
  named mutex? The Windows-specific Mutex API (win32event.CreateMutex /
  pywin32) would add a non-stdlib dependency for a code path whose
  worst-case bug (1-2 extra extractions per hour, never a corruption) is
  far below the threshold that justifies a new dependency.
• Why skip the non-blocking scan lock on Windows instead of polling a
  filesystem-marker file? The orphaned-session scan only runs every 15
  minutes per session-start; even if two scans run concurrently, the
  backlog drainer's atomic .json → .processing rename deduplicates any
  overlapping work. Adding a marker-file lock would be more code with
  similar race characteristics.
• Why not delete the entire scan on Windows? Hooks should have the
  same observable behavior across platforms wherever the trade-off is
  acceptable. Disabling functionality on Windows would create a
  silent-feature-gap that's harder to debug than a documented race window.

Co-Authored-By: claude-opus-4-7 wontreply@getfucked.ai

Merge ordering

Status: MERGEABLE (clean against origin/main).

Blocks: #351 (Windows subprocess portability — session_start.py:~226 Popen edit needs this PR's _HAS_FCNTL guard at file top to be in place first).

Depends on: none. Disjoint from #344 / #346 / #347 / #348 / #349 / #350 / #352 / #353.

Recommended sequence position: between #344 and #346.

Full sequence (10 PRs from a 3-agent coordination):

#353 (CI runner) → #344 → #345 (this) → #346 → #351 → #348 → #352 → #347 → #349 → #350