Yard v0.5.0

Added

• Rule auto-learn + yard harness review (harness H4 completion). The
  learning loop already auto-recorded worker-proposed skills (S3); now a
  run's harness_suggestions of kind "rule" are auto-recorded too, as
  .agents/rules/learned-<slug>.md — an always-apply constraint H1 inlines
  into every packet (the worker proposes, Yard's deterministic core writes; no
  clobber; gated by auto_rule, default on). Because a rule is always-on it
  has no per-task attribution to score, so learned rules are kept until removed
  (git-reversible) rather than auto-pruned like skills. New yard harness review shows the learned rules and the learned skills with their eval
  scores in one place. (Deterministic-observation candidate mining — failure
  themes into candidates — remains the open part of H4.)

• Workspace hooks (harness H3) — deterministic guards that bind every
  worker. Executables in .agents/hooks/pre-run.d/* run before a worker
  spawns; a non-zero exit blocks the run (the task fails with the hook's
  reason, so the drain stops on it — fix the cause and re-run). Executables in
  post-run.d/* run during evaluation; a non-zero exit is a fatal check
  that blocks Done, folded into the evaluation. Each hook runs in the repo
  root with YARD_TASK_ID / YARD_RUN_DIR / YARD_WORKER, a 30s timeout
  (longer is killed and fails), and stdout/stderr captured to
  <run_dir>/hooks/<phase>/. Only executable files run, in sorted filename
  order. Unlike a single CLI's hooks, these bind Codex, Claude Code, and any
  generic-adapter worker alike. Yard ships no enabled hooks — yard init
  lays down empty pre-run.d/post-run.d and a documented README.md. Off
  with hooks: false in yard.yaml.

• Explicit skill authoring: yard skill research / create / apply (S2/S3).
  On-demand skills without hand-writing a SKILL.md. yard skill research "<topic>" runs a researcher-role worker that drafts a candidate skill to a
  run dir and installs nothing; yard skill apply <run-id> installs that
  draft; yard skill create <name> [--from "<topic>"] authors and installs in
  one step. The run is queue-isolated — like the planner it spawns one
  worker, but derives no intent/queue, so authoring a skill never disturbs the
  live intent (the gap that deferred this). The worker proposes the content;
  Yard (the deterministic core) is the sole writer. Authored skills are tagged
  source: created (not learned), so they are user-chosen and never
  auto-pruned — they persist like a library equip until unequip.

Fixed

• Recover a task wrongly stuck Failed by a dead orchestrator. If Yard
  exited after a worker finished but before the result was evaluated, the
  task could end up Failed even though its run produced a clean done
  result — and neither restart-recovery nor yard recover could salvage it
  (recovery only looked at Running tasks), forcing a wasteful full re-run.
  Recovery now also re-evaluates such a task's stranded result, detected by an
  unfinalized orphan run (worker.pid still on disk — a finalized run
  removes it — with the process gone). It routes through the evaluator, so a
  genuinely-bad result stays failed; only real, completed work is reclaimed.
  Surfaced by dogfooding Deadline12, where a completed map task sat Failed.