Loki Mode

Memory file-lock concurrency fix + CI test deflake

Hotfix for two CI failures surfaced by the v7.66.0 test matrix (the shipped
v7.66.0 runtime is unaffected; both issues are in the memory locking primitive
and a test harness, not in v7.66.0's completion-council changes).

• memory/storage.py _file_lock: removed the os.remove(lock_path) on lock
  release. Unlinking the lock file on every release is a flock+unlink
  inode-replacement race: with 3+ contenders, a holder can unlink the inode a
  waiter just acquired, after which a third contender opens the path, finds it
  gone, creates a NEW inode, and flocks that -- entering the critical section
  while the waiter is still inside. This dropped index.json topics under
  concurrent store_pattern / store_episode. Reproduced deterministically on
  Linux CPython 3.13 (16 threads: 3/25 runs dropped a topic pre-fix; 25/25 pass
  after). Persistent lock files are the standard flock pattern; stale ones are
  GC'd by _cleanup_stale_locks, which is itself flock-safe (probe-before-unlink,
  v7.65.0). This completes the v7.65.0 lost-update hardening (the lock body was
  correct; the release path was not).
• loki-ts/tests/gate-failures-cap-parity.test.ts: removed the execFileSync("head")
  subprocess spawns. They were non-hermetic (returned 0 bytes on some CI runners,
  e.g. macos-latest + bun 1.3.13) and unnecessary -- the head-vs-tail semantics
  the test pins are fully captured by comparing Bun's cap against the
  deterministic expected slice. No loss of coverage.