Loki Mode

Reliability + data-durability + repo hygiene wave (12 fixes)

This release is a quality/reliability pass: dashboard concurrency hardening,
memory data-integrity fixes, durable SWE-bench instance persistence, doctor UX,
healing-hook safety, and test-suite hygiene (no more leaked branches/temp dirs).
No user-facing behavior changes; everything is additive or a bug fix.

Memory data integrity

• Consolidation merge no longer drops entries (memory/consolidation.py):
  when two newly-generated patterns merged into the same existing pattern in one
  consolidation run, the second merge read a stale in-memory copy and the full
  overwrite silently dropped the first merge's conditions, source episodes, and
  confidence bump. The in-memory list is now refreshed after each merge, so no
  entry is lost. Regression-tested (proven non-vacuous: the test fails against
  the old code with the exact data-loss symptom).
• Memory store hardened against malformed input (memory/storage.py,
  memory/retrieval.py): non-dict entries in a pattern list are now skipped
  instead of raising AttributeError; a store missing the patterns key
  degrades to empty instead of KeyError; re-saving a pattern stays idempotent
  (no index double-count). Each failure mode has a regression fixture.

SWE-bench instance durability

• Completed SWE-bench instances now persist durably and resume
  (benchmarks/swebench-pro-pilot/run_batch.py): the pilot harness wrote
  per-instance results only under /tmp and rewrote a single monolithic ledger
  every iteration, so a reboot or a crash mid-write could lose every completed
  instance. Results now write incrementally, one atomic file per instance
  (temp + fsync + rename), to a durable repo-relative ledger under
  benchmarks/results/swebench-pro-pilot/ (git-ignored data, tracked
  .gitignore, so it survives reboots but stays internal). Re-runs skip
  already-completed instances; --force/--rerun re-runs from scratch. Infra
  failures and budget placeholders are not frozen as "done", so they stay
  re-runnable. Cleanup globs (/tmp, test-branch pruning) structurally cannot
  reach the durable results.

Dashboard concurrency hardening + heredoc CI guard (reliability)

• Dashboard no longer stalls under concurrent requests (dashboard/server.py):
  several FastAPI endpoints did blocking filesystem work (event-log reads,
  learning-signal scans, memory blobs, cost/timeline aggregation, council
  transcripts, checkpoint listing, log tails) directly on the asyncio event
  loop. A single slow read could freeze every other in-flight request, so the
  status poller and live panels could hang while one endpoint did heavy I/O.
  Those blocking call sites now run via asyncio.to_thread, keeping the event
  loop responsive. Verified: /api/status stayed sub-millisecond while a
  deliberately slow endpoint blocked for seconds; all endpoints still return
  200. The shared compose-discovery TTL cache reachable from the new concurrent
  path is threading.Lock-guarded (read-modify-write bracketed); the wrapped
  helpers are pure reads, and no shared-state writer was newly threaded.
• CI guard against the v7.41 heredoc $<digit> footgun
  (tests/check-heredoc-dollar-digit.sh, wired into scripts/local-ci.sh):
  scans autonomy/loki and autonomy/run.sh for unescaped $1..$9 inside
  python3 -c "..." bodies, which set -u bash silently expands as positional
  params (the class of bug that once crashed loki plan --json for every user).
  Non-vacuity proven (the check fires on a planted violation). Runs on every
  local-ci pre-push.

Doctor UX

• loki doctor now prints the exact install command for a missing provider
  (autonomy/loki): a missing optional provider CLI (claude/codex/cline/aider)
  previously printed only "not found" with no remedy; it now prints the precise
  install command under the warning. Bounded network probes (ChromaDB, MiroFish)
  gained --connect-timeout/--max-time so a doctor run can no longer hang on a
  dead endpoint, with transient stderr-only progress feedback (stdout stays
  byte-identical for bash/Bun parity).

Healing-hook safety + test hygiene

• Healing hooks regression test for live-tree safety
  (tests/test-healing-hooks-safety.sh): locks in that a healing revert
  preserves unrelated uncommitted edits (snapshot-scoped restore, no blanket
  git checkout) and that friction path matching is exact (no substring false
  hits). The source already had these guards; this test prevents regression and
  proves non-vacuity against the old destructive behavior.
• Test suite no longer leaks git branches or temp dirs
  (tests/test-e2e-features.sh, tests/test-stop-process-group.sh): added
  unconditional trap ... EXIT INT TERM cleanup so an interrupted test can no
  longer leave a loki-test-wt-* worktree/branch in the real repo or a
  loki-pgtest-* temp dir behind. New scripts/clean-test-branches.sh safely
  prunes only test-artifact branches from a sandbox repo (refuses the real repo,
  protected branches, or while a live run is active).