feat: composer workflow restructure and factory hardening

[FrkAk]

Summary

Task Reference: [MYMR-237]

Three stacked rounds of composer work, making it a PR-terminal software factory: a Mymir task goes research → plan → implement → CI gate → review → bounded fix loop → opened PR, with crash-safe recovery and a GitHub-feedback rework round-trip. HOTL owns merging; composer terminates at the opened PR.

Round 1 — workflow restructure (63b7e91..4e3c51d):

• skills/composer/SKILL.md rewritten as a lean workflow: shared STATUS vocabulary (DONE / DONE_WITH_CONCERNS / NEEDS_DECISION / BLOCKED), todo-anchored loop with digraph, continuous execution, structural stop conditions, red-flags table, CSO-fixed description. /goal harness removed entirely.
• Bounded review→fix loop (2 rotations, then HOTL escalation); implementer fix mode.
• Slim per-phase reference extracts under skills/composer/references/ replace full mymir-spec force-loads in the four phase agents (researcher spec context ~4,950 → ~2,650 words; reviewer ~6,500 → ~900); extract headings carry canonical source prefixes so citations resolve.

Round 2 — factory hardening (9b5e173..97c9c15):

• 13 reliability rules from an adversarial audit: plannable-pick exit + dependency pre-flight, Mymir-transport stop condition, BLOCKED-because-terminal exception, claimed-task entry, PR-link recovery with [taskRef] bracket verification, environmental gh failure handling, backlog transient retry + stranded-task report, stale-claim sweep, headless gate skip, provisional propagation on escalated verdicts, branch-collision check, foreign-commit re-evaluation, terminal-write status re-read.
• Implementer runs worktree-isolated (isolation: worktree, live-verified incl. signed-commit push from inside a subagent worktree); default-branch derivation replaces hardcoded main; merge-forward policy before PR and each fix rotation; claim-ownership semantics with branch-evidence fallback.
• CI gate: orchestrator waits bounded (timeout 600 gh pr checks --watch, exit-code mapped) before dispatching review; pending/unresolved CI caps the verdict at request-changes.
• Crash-safe append-only run log at .mymir/composer-<project>.md (event vocabulary, increment-before-dispatch rotation counter, grep-derived counters, recovery algorithm, archive rotation).
• Rework mode: /mymir:composer rework <taskRef|pr-url> — reviewer-led intake fetches unresolved GitHub review threads (GraphQL, outdated-anchor re-location), re-verifies against HEAD, feeds the existing fix loop with a fresh 2-rotation budget; implementer may post one summary comment, never resolves threads.
• Context/cost: researcher single depth='agent' fetch; estimate-based model selection with 7 opus-forcing guardrails (reviewer never downgraded); app fix: formatCriteria now renders acceptance-criterion ids (lib/context/format.ts), closing a bug where the researcher's documented by-id AC rewrite appended duplicates — TDD'd in tests/context/format.test.ts, 5 golden snapshots regenerated.
• Flag-gated research-ahead pipelining (--pipelined): conservative variant, lookahead 1, 7-row brief-invalidation table, kill switch after two consecutive invalidations.
• scripts/check-plugins.ts gains two CI gates: @-include target resolution across every plugin (a dangling include silently strips an agent's rules at runtime) and canonical hash pins for the composer extracts (references/sources.json) — any edit to a pinned mymir reference fails CI until the extracts are reviewed and the pin refreshed via bun run sync:plugins.

Round 3 — fresh-eyes review fixes (74eb69b, cf3fcb9): 10 confirmed defects, each verified against server code or reproduced shell behavior before fixing:

• Implementer branch setup ran git checkout "$DEFAULT_BRANCH", which git refuses (exit 128) inside the worktree isolation its own frontmatter mandates; now fetches and branches from origin/$DEFAULT_BRANCH without checking out.
• Rework fix dispatches never said "rework", but the implementer's fix-mode contract requires the word to accept an in_progress entry (HOTL flip); dispatches now carry a Rework. prefix.
• Completion Protocol AC payload example was schema-invalid ({id, checked}; the MCP schema requires text); now {id, text, checked}.
• Orchestrator's reviewer dispatch instructed an up-front depth='review' fetch that defeats the reviewer's own two-phase isolation; dropped to the contract's 3-line dispatch shape.
• review.md pre-flight conditioned on files, which depth='working' mechanically excludes; reworded to the missing-PR-handle check.
• --pipelined parsed as single-task under the invocation rule; the prefetch brief had no run-log event and risked breaking the last-PICK recovery invariant (new BRIEF event, prefetch logs no PICK).
• Stale researcher/planner frontmatter descriptions (researcher claimed it "does not write to Mymir"); canonical artifacts.md §2 tag-vocab guidance corrected overview → meta with extract + mirror sync and pin refresh.

Committed regression suite (skills/composer/tests/scenarios.md, now 20 scenarios with RED→GREEN provenance) — all pass against this branch. Platform mirrors (codex / cursor / antigravity) synced for the review skill, reviewer-rules extract, and mymir references.

Type of change

• New feature
• Bug fix
• Refactor / cleanup

Testing

• Tested locally with bun test tests/context tests/api/task-context.test.ts (36 pass)
• Linting passes (bun run lint)
• Typecheck passes (bun run typecheck)
• bun run check:plugins (mirror sync, include resolution, extract pins) and bun run format:check pass

Plugin markdown verified by the committed 20-scenario pressure suite dispatched as fresh subagents (loopholes found during runs were countered and re-tested). Live worktree-isolation smoke test confirmed branch/commit/push from inside a plugin-agent worktree.

Notes for reviewer

• Live end-to-end acceptance (a real /mymir:composer <task> run + rework round-trip) requires this branch installed as the plugin source and a fresh session; checklist in the repo-external plan doc.
• decompose-* agents still force-load full mymir specs; converting them to slim extracts is deliberate future work. The rework GraphQL intake reads the first 100 review threads (no hasNextPage pagination yet) — documented future work.
• No caller-id MCP surface exists today, so task claims use branch-evidence ownership (assigneeIds wiring documented as the upgrade path).