Loki Mode

Parallel-worktree cleanup batch (9 isolated fixes, file-disjoint)

Nine fixes developed in parallel isolated git worktrees, each with a non-vacuous
regression test, merged conflict-free (zero file overlap). Full pytest gate: 1176
passed; all 6 standalone tests pass.

CI hardening:

• Release gate now runs bun typecheck + test (release.yml): publish-npm and
  publish-docker ship the loki-ts Bun dist, but the gate was python-only. A broken
  TS build or failing bun test can no longer reach npm/Docker.
• bun-parity stops blanking doctor pass/fail/warning counts (bun-parity.yml):
  the blocking gate normalized away the exact count-shift class that broke v7.58.0
  (doctor crash off-TTY). It now keeps the counts, matching the nightly parity-drift.

Reliability:

• App Runner non-compose watchdog runs a health check (app-runner.sh): a
  hung-but-alive HTTP app (passes kill -0, no longer serving) is now detected and
  restarted; the HTTP-fail path writes ok:false instead of a misleading ok:true.
• Benchmark harness: atomic results write + resume + dead-flag removal
  (run-benchmarks.sh): intermediate results now write via temp+os.replace (a crash
  mid-write no longer corrupts the file), a new --resume skips already-solved
  instances (no redundant paid calls), and the misleading unused --parallel flag is
  removed.
• Checkpoint prune sorts by basename (run.sh): the index rebuild sorted full
  metadata paths with -t- -k3, so a hyphen in the cwd path (the repo is literally
  loki-mode) shifted the epoch field and corrupted rollback ordering. Now sorts on
  the checkpoint dir basename.

Memory:

• Valid ISO timestamps + Layer-2 budget gating (cross_project.py,
  knowledge_graph.py, layers/loader.py): dropped the double-Z suffix that produced
  invalid +00:00Z timestamps (broke fromisoformat), and gated Layer-2 timeline
  loading on the remaining token budget so an oversized timeline no longer overspends
  max_tokens.
• Index/timeline layer dead cache removed + non-mutating relevance boost
  (layers/index_layer.py, timeline_layer.py): removed a _cache field that was written
  but never read (3 disk reads per query; a real cache would serve stale data across
  the separate writer/reader processes), and the keyword match boost now lives on a
  transient match_score so a topic's returned relevance_score reflects its stored
  value.

Audit + docs:

• verifyManifestLink detects trailing-truncation (src/audit/crosslink.js): a
  truncated audit chain (dropping the manifest-link anchor) re-linked cleanly and read
  as valid; it now cross-checks the witness file's agentEntries high-water mark and
  flags truncation.
• Template complexity buckets + stack row corrected (templates/README.md): the
  simple-todo-app stack was wrong (listed React/Express/SQLite; it is vanilla
  JS/localStorage), and the Simple/Standard/Complex gallery now matches how
  detect_complexity actually classifies each template by section count.