perf: make repository monitoring non-interfering

[fieldsofland]

Summary

Stackmap can now monitor large repositories without competing with normal Git commands or consuming a CPU core while idle. Passive inventory work cannot acquire the Git index lock, repository notifications collapse into one refresh after activity settles, and recoverable exits leave no monitoring workers or child processes behind.

Design

• Run passive Git commands with GIT_OPTIONAL_LOCKS=0, while preserving normal locking and a separate 30-second deadline for explicit checkout and deletion mutations.
• Filter .git notifications to state that affects the branch map, ignore transient locks and object-store churn, and debounce relevant bursts for 200 ms.
• Use five-minute reconciliation only as a dropped-event safety net instead of repeatedly scanning every 30 seconds.
• Supervise subprocess groups and join watcher, refresh, and upstream workers during normal quit, Ctrl-C, SIGHUP, SIGTERM, and recoverable error exits.
• Give timed-out commands a graceful TERM window before forced termination.

Measurements

FactMachine monorepo	Before	After
Settled CPU per runaway instance	88–98%	0.0%
Passive monitoring can create index.lock	Yes	No
Process remains after q	Observed	None
Fallback reconciliation	30 seconds	5 minutes

Checkout and guarded deletion retain their existing mutation semantics and integration coverage.

Verification

• cargo fmt --check
• cargo clippy --offline --all-targets --all-features -- -D warnings
• cargo test --offline --all-targets --all-features (176 tests)
• Offline release build
• Live startup, idle CPU, index-lock, and shutdown smoke test in the FactMachine monorepo

---