Release v0.44.0

Added

• Strict spec contract review (#1650): on spec/* branches the local AI
  reviewer runs a second, non-blocking pass against the spec checklist and
  records STRICT_SPEC_* evidence without changing the ordinary verdict.
• Strict implementation-plan review (#1655): a second strict checklist
  beside the spec pass, with plan documents read at the reviewed head and
  STRICT_PLAN_* evidence in loop history.
• Missed-finding telemetry (#1651): when an external reviewer reports
  blocking findings after a local-clean pass, the loop history records the
  miss without changing the review outcome.
• Review doctrine for the local AI reviewer (#1654): five generalized
  finding patterns, a doctrine linter, and REVIEW_DOCTRINE_* supply-state
  reporting in the local reviewer bundle.
• Reviewer effectiveness report (#1657): read-only
  reviewer-effectiveness-report.sh summarizes loop history per pull request
  and across a recent window.

Changed

• Current-head reviewer evidence (#1648): summaries record which commit
  each reviewer actually reviewed; a stale local-clean result no longer
  satisfies ready-for-human-review.
• Expensive reviewers after local clean (#1649): codex-github waits for
  current-head local-clean evidence, prior-reviewer evidence, resolved
  threads, and green non-reviewer checks, and fails closed when that
  evidence is missing or stale.
• Small-finding terminal policy (#1652): blocking findings on spec, plan,
  protocol, or the review contract are never "small"; contract-surface
  findings on other docs escalate; counted rounds must be on the current
  head.
• Stage-specific local reviewer prompts (#1653): the local reviewer
  resolves workflow stage from the head branch, can add the Workflow Policy
  checklist, and emits REVIEW_STAGE* evidence.
• Second local pass before ready-phase (#1656): the loop may re-run the
  local reviewer when its latest verdict is not clean on the current head,
  and re-reads the live PR head on every proceed path (fail closed on
  mismatch).

Fixed

• Workflow test harnesses in consumer repositories (#1631): suites now
  read the live .ai-dev-workflow.yaml or skip, so downstream replacements
  of placeholder workflows keep the harness check green; the harness also
  installs PyYAML before YAML-parsing suites.
• Protocol 91 pgrep wait-loop guidance (#1659): document the bracket
  trick so a wait loop cannot match its own command line, and capture
  pgrep's exit code in pgrep_rc instead of reading $? after the loop.