v0.3.0 — the world-class revamp

The world-class revamp: corrected economics, context discipline, a deep surface, the
learning moat wired end-to-end, a falsifiable benchmark, and engine-owned worktrees.

Added

• Worktree Phase 1, engine-owned (pqa/worktrees.py): one isolated git worktree
  per branch on ephemeral pqa/<run>-bN branches, with a write-ahead registry in
  .pqa/state.json so strays survive even a mid-run SIGKILL; rollback on partial
  spawn; reconcile() merges --no-ff, aborts on conflict preserving the survivor
  branch, and always prunes. Branch.workdir + run(workdirs=...) thread isolation
  through the engine; spawn_branches.sh / reconcile.sh became thin engine callers;
  the orchestrator and reconciler honor branches_mode = "worktree" with a stray
  sweep at init. Zero-orphan recovery is a tested invariant.
• Locked eval benchmark: 8 tasks under evals/tasks/ (each task.toml + LOCKED
  verify.py + reference.py must-pass + sabotage.py must-fail) and
  scripts/eval_harness.py (deterministic score/report/smoke, zero model
  calls); /eval and pqa-eval-runner wired to it; nightly eval-smoke workflow
  re-proves verifier integrity. The README documents the methodology; live numbers
  land only from a live run — losses included.
• The learning moat, wired: conviction signals get their outcomes back-filled
  post-collapse; pqa/instincts.py synthesizes instincts from precipitates+failures
  (overlap clustering; confidence from support and contradictions); prior-art injects
  top instincts; RunReport carries instincts_injected and per-instinct agreement;
  the dashboard gains calibration + instincts sections; the self-reflector reads the
  engine's calibration().
• Run resume: crash-resumable run journal (pqa/state.py → .pqa/state.json,
  atomic tmp+rename) — /pqa --resume re-enters at the first incomplete stage.
  Journal writes preserve foreign top-level keys (the file is shared with the
  worktree registry).
• Generated configuration reference: docs/configuration.md rendered from
  pqa/config.py by scripts/generate_config_doc.py, drift-pinned by tests.

Changed

• Economics corrected; tokens primary: cost-model defects fixed, model aliases
  (fable/opus/sonnet/haiku) wired to real pricing/dispatch via
  pqa.cost.resolve_model, budgets token-primary with USD secondary, and a
  pre-flight would_abort gate before every dispatch (not just after the spend).
• Model routing per role: Fable 5 where output quality is decided (generators,
  unknown-scout, adversary, collapse-judge, baseline control); sonnet/haiku for
  mechanical and bookkeeping tiers. "Every agent on Opus" is gone from docs and
  dispatch.
• Context discipline in the orchestrator: branch payloads live on disk and are
  read only by the subagent that needs them; the orchestrator holds ≤200 tokens of
  state per branch (digests only) and reports per-stage context telemetry.
• Surface: depth over breadth — 34 agents · 59 skills · 27 commands trimmed to
  14 agents · 12 commands · 12 deep skills (each skill a protocol + worked
  example + anti-patterns playbook); validate_components.py gained census, depth,
  and description-budget gates and drift-gates docs/catalog.json.
• Memory retrieves by relevance under a hard token budget (not recency), and
  every injected memory id is reported per run.
• Docs truth pass: README (counts, hook claims, stage wording, workflow count,
  status), architecture.md (rewritten to the shipped reality), CONTRIBUTING and the
  plugin manifests; hooks language unified with SECURITY.md (two blocking hooks,
  the rest fail open; the binding guarantee lives in CI).

Fixed

• Hook hardening: per-hook kill-switches (PQA_DISABLED_HOOKS), once-per-session
  research gate, fail-closed fixes on the security/secrets gates.