Loki Mode

Memory subsystem hardening (wave-6 bug-hunt)

A dedicated hardening pass on the memory subsystem (storage.py, engine.py,
retrieval.py, consolidation.py, token_economics.py). The headline is a real
data-integrity fix; the rest is defensive robustness against corrupt or
hand-edited records. Reviewed by a 3-reviewer council (2 Opus + 1 Sonnet) to
unanimous approve. Full Python test suite: 1248 passed, 11 skipped.

Data-integrity (the real HIGH):

• Lost-update under concurrent writes: storage._decay_semantic /
  _decay_episodic / _decay_skills and engine._update_index_with_episode /
  _update_index_with_pattern previously read a JSON state file under one lock
  scope, mutated in memory, then wrote under a separate lock scope. A concurrent
  save_pattern / store_episode landing between the read and the write was
  silently clobbered (dropped patterns, under-counted index totals). All sites
  now hold ONE exclusive file lock across the full read-modify-write, mirroring
  the already-correct save_pattern template. This completes the v7.58.0
  lost-update fix, which had been applied unevenly. Cross-process safe (fcntl
  flock on the sidecar lock file).

Defensive robustness (corrupt / hand-edited records; no live writer emits these
values, schema validation enforces ranges):

• Null-field guards across every score/keyword path that does arithmetic or
  .lower() on a raw record field: boost_on_retrieval, apply_decay,
  calculate_importance (importance, phase, category), _score_result,
  _keyword_search_semantic / _anti_patterns / _skills / _episodic,
  optimize_context (confidence, usage_count), and the engine search /
  update-pattern confidence + usage_count paths. An explicit null field no
  longer crashes a query with TypeError / AttributeError. The guard preserves a
  legitimate stored 0.0 (is-None check, not truthiness) where 0.0 is
  meaningful (importance, confidence).
• Path-traversal hardening: engine.store_skill and storage.save_episode
  now sanitize the filename / validate the date component before building the
  on-disk path, matching save_skill. A poisoned skill name or episode
  timestamp (../../../tmp/evil) can no longer escape the memory root.
• _cleanup_stale_locks now probes a stale lock with a non-blocking flock
  before unlinking, so a long-running (>5min) live writer's lock is not removed
  out from under it.

Correctness:

• consolidation: cluster-merge dedup now appends the saved pattern to the
  in-memory set so two similar clusters in one run dedup correctly (no more
  near-duplicate patterns); zettelkasten link-count is incremented only for
  links that actually persisted.
• retrieval Layer-2 progressive budget now trims topic summaries to the
  remaining budget (mirroring Layer 3) instead of all-or-nothing admission.
• retrieval topic-relevance ranking now reads the keys the writer actually
  emits (id / summary), so the previously-inert layer-1 ranking works.
• token_economics optimize_context is now strict-by-default (never exceeds
  the stated budget); a previously-hardcoded 10% overage is now an explicit,
  documented, opt-in slack_ratio parameter (default 0.0).

Test infrastructure:

• pytest.ini now collects hyphen-named Python test files
  (python_files = test_*.py test-*.py). The repo had long carried hyphen-named
  tests (e.g. test-memory-retrieval-bugs.py, test-memory-consolidation-merge.py)
  that the default test_*.py-only pattern silently skipped, so their
  regression guards never ran under pytest tests/. They now do.