v2.6.0

Added

• mq-hal actions shows a local, append-only ledger of what HAL ran,
  cancelled or refused. The router, execute steps and confirmed
  open/fix each write one mq_hal.action.v1 entry to
  $MQ_HAL_STATE_DIR/actions.jsonl with intent, repo, routing, command,
  confirmation, outcome, exit code and duration (newest 1000 kept). Prompts
  and arguments are secret-redacted before write. A failed ledger write only
  warns; it never blocks the command.

• audit --verification and stack-status --verification present repo-signal's
  existing changed-file evidence as PASS, FAIL, STALE or NOT_VERIFIED, separately
  from structural/runtime status. Shared bounded read-only collection supports
  canonical and legacy stack views; missing or malformed evidence never passes.
  Default and sample paths do not start verification providers. No tests or
  receipt writes are performed by HAL.

• mq-hal compatibility shows the stack compatibility report produced by
  mq-agent stack compatibility --json (mq.stack-compatibility.v1):
  repositories, producer/consumer relationships, findings with every
  implicated repo, and mq-agent's next action. --json prints the producer's
  bytes. Transport differs from provenance in one way: mq-agent exits 2 for
  FAIL and 3 for UNAVAILABLE while still printing a report, so a non-zero exit
  with a valid report is a report, and the view passes mq-agent's exit code
  through. HAL derives no severity, blocking rule or remedy of its own.

• mq-hal route readiness and mq-hal route divergence expose mq-agent's
  evidence-gated learned-routing state as read-only operator views. HAL validates
  the producer schema IDs, preserves mq-agent's recommendation/status, current
  analysis era, route counts and safety flags, and never recomputes eligibility,
  ranks a route or grants promotion. Divergence defaults to mq-agent's current
  era and docs-review, with optional --era / --task-class filters.

• Consumer contracts for mq-agent stack cockpit, release-check and
  provenance --json: the schemas (and runtime_identity, which provenance
  references) are vendored in schemas/vendor/, and
  tests/stack-contract-smoke.sh checks them against mq-agent (CI checks out
  the four files), validates sanitized real-shaped fixtures, and feeds them
  through hal.stack, hal.release and hal.provenance. CI installs
  jsonschema for it.

• Consumer contract for mqlaunch doctor --json (mq.doctor-status.v1,
  owned by macos-scripts, vendored in schemas/vendor/).
  tests/doctor-status-contract-smoke.sh checks the copy against macos-scripts
  (CI checks it out), validates a fixture and doctor-summary --sample
  against the schema, and feeds the real shape through brief and
  doctor_summary.

• mq-hal changes compares the current check with the previous local
  snapshot and lists new, resolved, persisting, and unverified problems by
  stable check ID with source and timestamp. A check that is missing or
  SKIPPED after a previous problem is reported as unverified, never as
  resolved. Works without Ollama, supports --json, and reuses the dashboard
  collectors instead of adding new probes.

• mq-hal feedback is a read-only operator view over mq-agent Feedback
  Engine status, reports, comparisons and candidates. Machine JSON is passed
  through unchanged; human output renders producer verdicts and future unknown
  states neutrally without recomputing them. Mutating feedback actions remain
  outside HAL.

Changed

• Default local Ollama model is qwen3.5:9b for every profile (router,
  planner, critic, code), the planner, routing, doctor and model checks,
  replacing qwen3:4b-instruct, qwen3:8b and qwen2.5-coder:7b. Install
  hints in README and guides pull qwen3.5:9b.

Fixed

• Router grep_repo passes . to rg, so it no longer waits on stdin when
  stdin is not a terminal (pipes, scripts, CI).

• mq-hal open --confirm with a missing editor reports exit 127 instead of
  raising FileNotFoundError, and the attempt is recorded in the action ledger.

• Router output on a pipe: HAL's own line now comes before the command's
  output, and mq-hal "..." | head no longer ends with BrokenPipeError.

• release-check.sh no longer prints "all smoke tests passed" when a smoke
  test fails after an earlier gate has already failed.

• Stack verification resolves the cockpit label mqlaunch to the configured
  macos-scripts repo, preserving an explicit mqlaunch override. Register
  mqobsidian in the default repository map so both components are checked
  instead of reported as unconfigured.

• hal.py, doctor_summary.py, fix_planner.py and model_test.py send
  think: false to Ollama, as planner.py already did. qwen3.5 thinks by
  default: model_test.py took 25 s against its 30 s timeout, now 0.8 s, and
  an intent request took 46-80 s instead of 1.1 s.

• hal/operator.py is now hal/operator_actions.py. Run as a script it put
  hal/ first on sys.path and shadowed the stdlib operator module, so on
  any Python that does not import collections at startup (a python.org
  build, or python3 -S) mq-hal crashed on import with a circular-import
  error. tests/stdlib-shadow-smoke.sh forbids stdlib names in hal/ and
  scripts/ and runs every hal/ module without site.

• mq-hal release showed readiness unknown for every repo on real
  mq-agent stack release-check output: the parser looked for ready /
  status on each repo, and real items carry blockers and warnings
  lists instead. A repo with blockers is now no, one without is yes
  (warnings do not block).

• doctor-summary miscounted real doctor output: walking every string read
  the top-level status as a second warning and a check's detail: missing
  as a serious issue — 1 serious and 2 warnings for a run with 0 failures and
  1 warning. A mq.doctor-status.v1 document is now read from its own summary
  and checks; other doctors keep the heuristic. The --sample document used a
  message key real output never had and is now shaped like v1.

• mq-hal changes --no-save no longer reports "baseline recorded" when there
  is no previous snapshot; it says the baseline was not saved.

• mq-hal stack reads the cockpit contract mq-agent actually emits. The view
  still expected components[] / name / status / overall, while
  mq-agent stack cockpit --json emits repos[] / repo / gate /
  overall_gate under mq_stack_cockpit.v1. Cockpit placeholder dashes are
  now read as absent rather than as operator statuses, and the shared feedback
  normalizer knows the cockpit vocabulary — GO, NO-GO, REVIEW, DRIFT,
  BLOCKED, fresh, aging, stale. The smoke case is copied from the real
  cockpit-v1 shape, so the test no longer agrees with an imaginary producer.
  No routing authority moves into mq-hal; this stays a presentation fix.

• mq-hal brain and mq-hal runtime find truth exports where mq-agent writes
  them. mq-agent writes stack truth under memory/stack-truth/ while mq-hal
  probed only a top-level truth/, so brain reported zero truth exports and
  runtime warned missing: truth with the producer output sitting there.
  Accepted locations are centralized in BRAIN_FOLDER_SOURCES and both commands
  read the same mapping. Freshness and aging remain a separate concern; no
  scheduler was added.

• Two module invariants that the rebuilt ports had reintroduced: hal/stack.py
  regained a shebang despite having no __main__ and never being executed
  directly, and the script-mode fallback imports in hal/brain.py and
  hal/runtime.py lost their explicit type: ignore[no-redef, import-not-found] annotations. Behaviour is unchanged.