Loki Mode

Fix: memory panel is correct and self-consistent on JSON-backed projects

Builds on the fix contributed in PR #178 by @iizotov (memory summary reporting 0
on a JSON-backed project) and closes the two gaps its review surfaced, plus the
same class of bug in the sibling endpoints.

• /api/memory/summary no longer reports 0 on a JSON-backed project. The
  SQLite fast path returned all-zero stats whenever the module imports even with
  no .db file, short-circuiting the JSON fallback. It now only trusts SQLite
  when it actually reports data, else falls through to the JSON counts (the core
  PR #178 fix). Episodes and skills are counted by recursing the dated
  subdirectories (episodic/YYYY-MM-DD/*.json, skills/*/*.json) instead of a
  flat glob that found nothing.

• The summary count and the drill-in lists can no longer disagree. PR #178
  fixed only /api/memory/summary; /api/memory/episodes, /api/memory/skills,
  the single-item lookups, and the JSON stats fallback still used a flat glob, so
  the panel showed a count of N but the list returned 0. All of them now go
  through two single-source-of-truth helpers, _memory_episode_files() and
  _memory_skill_files(), so the count and the list are always computed the same
  way. .lock files are ignored everywhere.

• latestDate is the newest episode by recorded timestamp, not by filename.
  The intra-day filename tiebreak is a hash, not a time, so a filename sort could
  report the wrong episode (observed hours off). The summary and the episodes
  list now select the newest by parsed timestamp, and an honest null/empty is
  preferred over a confidently-wrong value when no timestamp is readable.

• Verified against a real JSON-backed project (dated episodic subdirs, no .db):
  the summary count equals the episodes-list length and the reported latestDate
  matches the true maximum timestamp. Regression coverage:
  tests/dashboard/test_memory_json_parity.py (summary/episodes parity,
  summary/skills parity, timestamp-not-filename latestDate, .lock exclusion).