[Bug]: ephemeral lock holder sidecars race directory walks and match *.json globs, causing Windows-only test flakes #6128

Preflight

• I searched existing issues and discussions for this behavior.
• This report is not an unpatched security vulnerability.
• The report and attachments contain no credentials or private/internal material.

Issue origin

Observed or reproduced in a real environment

LoopX version or commit

Observed on trees near 641855d9c and c2b17b3e (October main); the sidecar mechanism itself is older (loopx/file_lock.py, LOCK_HOLDER_SCHEMA_VERSION = "file_lock_holder_v0").

Host or runtime surface

Not host-specific (reproduced on native Windows 10, zh-CN, cp936)

LoopX area

Build, packaging, installer, or CI

Problem

exclusive_cross_runtime_file_lock writes an ephemeral holder sidecar next to the locked file (lock_holder_path: <locked>.lock.holder.json). Two properties of that sidecar combine into a class of Windows-only test flakes in suites that walk runtime state directories:

1. The name ends in .json, so glob("*.json") and directory walks that collect JSON state files match it even though it is not state.
2. The lifecycle is ephemeral and racy against walkers: it is created around the locked operation and removed on release. On Windows the create/delete window is wide enough that a walker can list the sidecar and then fail read_bytes() with FileNotFoundError, or assert against a directory snapshot that still contains it.

Two concrete reproductions from reviewing current PRs (both native Windows, cp936):

• tests/test_loopx_turn_executor.py::test_reserved_managed_start_recovers_after_death_before_the_first_journal_write - the "no journal written" assertion at the death point uses glob("*.json") and matches the persisted ...turns/<hash>.json.lock.holder.json; after a (simulated) crash the holder legitimately survives, so the assertion trips even though no Turn journal was written.
• tests/test_peer_collaboration.py::test_canonical_stopped_goal_rejects_peer_request_before_any_write - a helper that walks .local/manager-context/entries/** collecting {relative_path: read_bytes()} listed a ....lock.lock.holder.json and raised FileNotFoundError when the sidecar was deleted before the read. Flaky across runs (passes on rerun).

Both are test-side failures, but the same walk pattern exists in product-adjacent code paths that enumerate runtime directories, so the class is broader than these two rows.

Minimal reproduction

1. On Windows, run a suite that walks a runtime directory containing locked files (e.g. python -m pytest -q tests/test_peer_collaboration.py::test_canonical_stopped_goal_rejects_peer_request_before_any_write) - repeat until the walker lists a .lock.holder.json that disappears before its read.
2. Or directly: hold exclusive_cross_runtime_file_lock over a file in a loop and, from another thread, walk the directory - observe transient FileNotFoundError on ....lock.holder.json and *.json glob matches.

Expected behavior

Directory walks over runtime state should be deterministic: either the holder sidecar is not named like a state file, or walkers (and the tests that assert on walks) skip/ignore it consistently on every platform.

Actual behavior

On Windows the sidecar is caught mid-lifecycle: glob assertions see persisted holders after simulated crashes, and walkers raise FileNotFoundError on holders deleted between listing and reading.

Sanitized diagnostics

FAILED ... test_canonical_stopped_goal_rejects_peer_request_before_any_write
E  FileNotFoundError: [Errno 2] No such file or directory:
   '...\.local\manager-context\entries\<hash>\<hash>.lock.lock.holder.json'

FAILED ... test_reserved_managed_start_recovers_after_death_before_the_first_journal_write
E  assert not [...\turns\<hash>.json.lock.holder.json]   # glob("*.json")

Additional context

Possible directions, maintainer preference welcome:

1. Rename the sidecar so it does not end in .json (it is lock metadata, not state), or move holder files into a dedicated locks/ subdirectory;
2. Ship a small shared walk helper (or fixture) that skips .lock.holder.json and tolerates per-entry FileNotFoundError, and use it in the two affected suites;
3. Minimally, patch the two test sites (glob filter + per-entry tolerance) so Windows runs are deterministic.

Happy to implement whichever direction is preferred.