perf: reuse logging locks and cache source paths

[codeforester]

Persistent logging now reuses its private sidecar descriptor and caches human-log source paths per invocation. Repeated source paths avoid filesystem resolution, forked children reopen their lock descriptor, and sidecar failures use logging's error handler so commands can complete.

Fixes #381.

Branch maintenance

Refs #426. Targets the branch for #415. Retarget and refresh after that parent is squash-merged; preserve the ordered stack.

The branch was refreshed without rewriting history to include main at a576cc279739eae5e4cfc33ffab2a7fb56de24de.

Current-head validation

At 66ce2689a0b5859ad25927aaa9274c7a4ff5ab19: uv lock freshness and baseline, runtime, strict typing, style, and contracts passed locally with all declared extras. Runtime result: 616 passed, 1 warning, 262 subtests passed in 8.57s.

Hosted checks: 7/7 required checks passed; 0 checks pending; 0 unsuccessful checks at 2026-10-04T14:19:40.760978+00:00. See the PR Checks tab and #426 for subsequent results.

[codeforester]

Additional finding (unchanged line outside this diff's hunks, so not postable as an inline comment):

lib/python/base_cli/logging.py:353 — format() now calls self._source_path(record) (the new method), but the old module-level _source_path() plus its helpers _active_project_root()/_active_application_home() are left in place, byte-for-byte unchanged, and are no longer called from anywhere in this file or elsewhere in the repo. Two functions named _source_path now coexist implementing diverging root-resolution logic — confusing for a future maintainer, and a likely source of a bug if someone edits the wrong copy believing it's live. Worth deleting the dead module-level trio in a follow-up.

[codeforester]

Addressed the blocking sidecar-loss regression in the pushed head 8a2ef793.

The handler now stats the sidecar before permission repair, treats a deleted/replaced sidecar as stale, reopens it, and preserves later records. The dead module-level _source_path/active-context helpers were also removed; the cached formatter method is the sole implementation. Added a regression test for deleting the sidecar during a live handler.

Validation: the supported runtime gate passes (626 tests, 262 subtests), and the style gate passes.

[codeforester]

Follow-up: the first hosted Windows matrix exposed that the new deletion regression test itself cannot unlink an open sidecar on Windows. The test is now explicitly POSIX-only in c4e4b9b9; the production recovery code remains unchanged and cross-platform. The focused logging tests and Ruff pass locally, and the hosted matrix has been restarted on the corrected head.

[codeforester]

Branch maintenance complete.

• Synced with current main (including security: centralize secret-name redaction heuristics #400) in merge commit 250bbe40; Git reported no textual conflicts.
• Targeted logging/performance tests and the full test suite pass locally.
• Ruff, formatting, and strict mypy pass.
• Fresh hosted checks are queued; no failures reported yet.

The PR remains open and unmerged.