feat(meta-orchestrator): tri-layer classifier L2 jaccard REUSE + L3 done.md convention

[artyhoo]

Summary

Stage 2-extend of meta-orch-no-arg-overview umbrella: tri-layer umbrella-completion classifier extends L1 branch-matcher (#272) with Layer C2 jaccard REUSE + Layer C3 done.md convention. Closes the §1.5c empirical 13% branch-match coverage gap surfaced in PR #273 §4 verdict.

• L2 (C2): REUSE jaccard from dup-detect.sh:62 via sub-shell call — adds ~4% additional coverage per §4 estimate
• L3 (C3): NEW committed done.md per umbrella — ADAPT Cline Memory Bank pattern (SSOT docs(research): cross-worktree gitignored coord-doc sync — verdict ADAPT #77, ~85% match on storage sub-pattern)
• CLAUDE.md «Umbrella closure convention» section codifies done.md schema + write-trigger
• Paired-negative test (6 cases) under packages/core/hooks/done-md-completion-filter.test.ts
• Naming: existing "L1 extension" synthetic-candidate comments renamed to "synthetic-candidate extension" to prevent reader confusion with new completion-detection layers C1/C2/C3

Backed by PR #273 §4 binding verdict (multi-signal classifier confirmed; §0.5 falsifier P2 closed for multi-signal interpretation). dup-detect.sh NOT modified (Stage-4-owned per §6 Option B — Stage 2 calls it via sub-shell only).

§1.7 Forward-check applied

• build-first-reuse-default.md §3 — C2 jaccard is REUSE (zero new algorithm code), C3 done.md is ADAPT Cline SSOT docs(research): cross-worktree gitignored coord-doc sync — verdict ADAPT #77. file:line: .claude/skills/meta-orchestrator/helpers/dup-detect.sh:62 (jaccard source reused) + docs/meta-factory/prior-art-evaluations.md:145 (SSOT docs(research): cross-worktree gitignored coord-doc sync — verdict ADAPT #77 row).
• phase-research-coverage.md §1.12 — every verdict backed by research(meta-orch-no-arg-overview): Stage 0-remainder R-phase — full §1 checklist + Stage 3 UX design #273 §4 empirical 13% finding. file:line: docs/meta-factory/research-patches/2026-05-29-meta-orch-no-arg-overview-s0-remainder.md:84 («Match? ~15%. §1.5c empirical = 13% branch-matchable. NOT a problem-class fit as sole signal»).
• no-paid-llm-in-ci.md §1 — all mechanism is deterministic bash (file-presence, sub-shell call, grep) + vitest. file:line: .claude/skills/meta-orchestrator/helpers/priority-score.sh:1 (bash shebang; no claude -p in changed code).
• doc-authority-hierarchy.md §2 — CLAUDE.md edit adds section «Umbrella closure convention», preserves existing Authoritative-for header. file:line: CLAUDE.md:1 (top-level header untouched) + CLAUDE.md:102 (new section start).

§1.7 Backward-check applied

• priority-score.sh C1 branch-matcher from feat(meta-orchestrator): branch-matcher для umbrella completion #272 preserved verbatim as completion-detection Layer C1; C2 + C3 are ADDITIVE filters layered after. file:line: .claude/skills/meta-orchestrator/helpers/priority-score.sh:158 (C1 block start: «completion Layer C1 — branch-prefix match»).
• dup-detect.sh NOT modified — called via sub-shell (approach a) from priority-score.sh:109; dup-detect.sh callers unaffected (verified: grep -rn 'dup-detect.sh' .claude/skills/ shows SKILL.md:173 unchanged). file:line: .claude/skills/meta-orchestrator/helpers/dup-detect.sh:62 (jaccard source, untouched).
• CLAUDE.md «Umbrella closure convention» is a NEW section — no existing convention superseded. file:line: CLAUDE.md:102 (new section; existing PR-strategy / build-vs-reuse / Artifact Ownership sections verbatim preserved at lines 87/18/68).
• Existing tests dup-detect-empty-arg.test.ts (feat(meta-orchestrator): Stage 4 P4 fix — bash-quoted helper invocations + dup-detect empty-arg #266) + priority-score-branch-matcher.test.ts (feat(meta-orchestrator): branch-matcher для umbrella completion #272) untouched — new test is separate file. file:line: packages/core/hooks/done-md-completion-filter.test.ts:1 (new file; does not modify existing tests).
• Cline SSOT docs(research): cross-worktree gitignored coord-doc sync — verdict ADAPT #77 T16 problem-class check: upstream class = committed-markdown for AI session memory across turns (Memory Bank progress.md/activeContext.md); ours = umbrella closure marker (single done.md per umbrella, machine-readable schema, written at last-stage merge). Match: ~85% on storage pattern (committed markdown as load-bearing state); diverges on update trigger (Cline = on-demand AI-signalled, ours = explicit at-merge convention). ADAPT verdict preserved from SSOT row docs(research): cross-worktree gitignored coord-doc sync — verdict ADAPT #77. file:line: docs/meta-factory/prior-art-evaluations.md:145.
• done.md NOT written for this umbrella — Stage 3 still pending; recursive T15 validation deferred until full umbrella closure per kickoff §2 stage table.
• "L1 extension" in synthetic-candidate section renamed to "synthetic-candidate extension" at priority-score.sh:35,207 to prevent naming collision with completion-detection layers. Narrow scope comment-rename; no logic changed.

Test plan

• npm run typecheck green (all workspaces)
• npm test — 1 pre-existing failure on staging (hooks/pre-push.test.ts — unrelated to this PR; verified pre-existing via git stash before/after comparison); all other tests pass including new 6-case test
• New paired-negative test passes — 6 cases (C2±, C3±, combined, all-fail)
• L1 paired-negative test (feat(meta-orchestrator): branch-matcher для umbrella completion #272) still passes — preserved verbatim (5/5)
• dup-detect.sh tests unchanged + passing (14/14: dup-detect.test.ts + dup-detect-empty-arg.test.ts)
• All 21 principle tests pass
• CLAUDE.md edit landed; new section follows doc-authority-hierarchy header format
• (deferred) done.md NOT written for this umbrella — Stage 3 still pending; recursive T15 validation at full umbrella closure