feat(workshop): port spec→plan→compile pattern (4th instance) (v5.59.0)

[edwinhu]

Summary

Ports the spec → plan → compiled-runner pattern to the workshop workflow — the 4th instance after ds/dev (exit-code) and writing (judgment). Workshop is writing's twin: flat-parallel section fan-out, compile = DATA work-list, no per-project run.js, no DAG. Distinct contribution: a rich-mechanical floor (typst compile + check-all.py real exit codes) + a semantic ceiling — a new gate-type data point for the pass #9 run-core extraction.

What changed

• ONE shared parser scripts/workshop/workshop_slide_table.py (canonical 7-col TABLE + legacy 4-field PROSE; tolerant-at-parser, canonical-at-emitter) replaces the doubled LLM Discover and emits a DATA work-list.
• workshop-generate.js: discFromIndex full-replaces the LLM Discover when slideIndex is passed; grep-grounded inventory fidelity; notes-compile gate (notes were previously ungated — could ship malformed).
• workshop-verify.js: keeps the slides.typ enumeration + an unbiased free-read semantic join (a candidate menu biased it → appendix over-match, caught by n=3 variance); the parser contributes a JS inventory whitelist applied outside the agent; scope:{checked,notChecked} disclosure (D1 / doctrine feat: improve skill scores (+46% avg across 5 skills) #3).
• Guard reconciled to validate = build_index().violations (S6, −89 lines). Prose now passes → the shipped opv deck is no longer rejected (parity regression fixed by construction); stale-approval = allow+WARN.
• Skills wired to run the parser + pass args.slideIndex (back-compat fallback). Born-canonical table emitter reconciled; structure-reorder PAUSE fixture.

Parity validation (real ~/projects/opv deck, via the opv-parity partner)

• Track A — compile-fail short-circuit byte-identical (zero wasted agents).
• Track B — n=3 clean after fixing an appendix over-match regression: COVERED [21,21,21], over-match [0,0,0].
• Parser A/B: generate 21/21 exact + verify 21/21 via semantic join; grep-fidelity independently confirmed file-backed; generate end-to-end PASS.
• 8 real issues surfaced + folded (parity-regression, cardinality, semantic-join, appendix-over-match, paperPath ~, Track-A count nondeterminism, permanently-red floor, notes gate-gap).

Doctrine deepenings (recorded for pass #9)

• A join turns semantic exactly when the work-list enumerates from >1 source; the parser owns enumeration, never a drifting-identifier join.
• Don't feed the deterministic artifact into the semantic step — even as a "helpful menu" it biases it; a JS post-filter > a prompt instruction.
• Single-sample parity is insufficient for judgment seams — n≥3 variance required.
• Gate only what you compile (notes are a first-class deliverable).

Tests

parser 37/37 · guard 6/6 · engines 23/23.

Out of scope (separate fast-follow — D-w-8)

Pre-existing mechanical-floor bugs deliberately preserved for parity, fixed separately: check-all.py rides cross-domain phantom constraints (→ permanently-red gate on every deck), divider-naive overflow, summary.critical≠findings.length.

DESIGN: docs/DESIGN-workshop-spec-plan-compile.md

🤖 Generated with Claude Code