epic-to-spec-terminology

Summary

spc-8 delivers itd-43's implementable remainder: the GL002 forbidden-synonym lint gate, built detector-first against the glossary, and the live-prose sweep it drove — 19 flagged uses of epic as the concept-noun cleared to spec across four record files, with the detector (not human judgment) deciding what counted. The criterion that held the spec open, a reviews-subsystem spec-review token, is dropped from itd-43 as moot (product thinker, 2026-09-23): itd-28 was re-scoped to pin plus staleness on 2026-09-21, and nothing classifies reviews by kind. The spec closes on the four criteria that remain.

Approach

Fix-the-detector, literally: the rule was armed before any prose changed, its 19 live-corpus findings were captured verbatim as the watched-fail evidence, the sweep then drained exactly what it flagged, and a regression test pins the live corpus at zero GL002 findings from here on.

• The rule (internal/core/lint): checkForbiddenSynonyms reads the glossary term files (.abcd/development/brief/glossary/) as the source of truth, flags an enforced synonym used as a standalone word in live prose, and errors if an enforced word is not actually declared in some term's forbidden_synonyms (the gate cannot drift from the glossary). Matching is case-insensitive with explicit Unicode word boundaries (utf8.DecodeRune + a word-rune class), deliberately not Go's ASCII-only regexp \b; a test proves epics, epicenter, and unicode-adjacent text do not match while bare and parenthetical uses do.
• Enforcement is a deliberate subset — epic only, for now. Most of the glossary's forbidden synonyms (sprint, milestone, project, feature) are common English words whose false-positive rate would sink the gate; each is opt-in via the rule config as the corpus is readied for it.
• Exemption scope (config in .abcd/record-lint.json, not silent skips): YAML frontmatter, code spans and fenced blocks, allow_context phrases (epic-to-spec, epic->spec, epic→spec — the rename's own name), and historical paths (research/, decisions/, dated plans/, intents/shipped/ + superseded/, the glossary term files themselves). Live surfaces — brief body, roadmap, drafts/planned intents, open specs, docs, commands — are in scope.
• Wiring: registered in .abcd/record-lint.json, so it runs under cmd/record-lint inside make preflight and CI — the gate is a gate, not a claim.

Milestones as delivered

1. GL002 rule + seven hermetic unit tests, watched-fail first (23c0e86, hardened by a1e1467).
2. The armed-detector sweep: 19 hits → 0 across itd-34-three-intent-kinds.md (11 lines), itd-6-rp-mcp-only-integration.md (6), itd-48-intent-fidelity-reviewer-roles-2-3.md (1 line, 2 tokens), roadmap/phases/phase-0-substrate.md (1) (2a6e9cd).
3. CHANGELOG entry for the new gate (6c1a88a).

Acceptance-criteria satisfaction

AC as ordered in itd-43 when this spec was minted → status and evidence:

1. No live epic-as-noun reference remains — gap-filled, detector-driven: GL002 armed pre-sweep flagged 19 lines (captured verbatim); post-sweep the corpus is at 0, pinned by TestForbiddenSynonymsRealGlossary (loads the real config + glossary over the real corpus). Historical records exempt per the AC's own carve-out.
2. The glossary resolves to spec with epic forbidden — met already: brief/glossary/core/spec.md carries term: spec and epic in forbidden_synonyms; no stale epic.md term file exists anywhere (verified by filesystem sweep). The intent's open question (rename vs stub) was answered historically — the rename happened with no stub.
3. Reviews subsystem classifies against spec-review — dropped from itd-43 as moot (product thinker, 2026-09-23). The Go tree has no reviews subsystem that classifies reviews by kind, and itd-28 was re-scoped on 2026-09-21 to pin plus staleness, so no epic-review token exists to rename and no spec-review emitter exists to build. The intent's criteria renumber: its fourth and fifth are this list's 4 and 5.
4. issue.schema.json uses related_specs — satisfied-by-adjudication: no issue.schema.json exists in the Go tree (old-system reference); the native validator (internal/core/capture/validate.go) and capture engine already use related_specs exclusively, with no related_epics anywhere. Nothing dead was built.
5. internal/core/lint raises no GL002 violation for epic — gap-filled: the rule now exists (it did not before), is wired into the preflight/CI record-lint, and the live corpus passes at zero (same regression pin as AC1).

Judgment call on record

Two of the 19 hits sat inside internal self-quotes (a working-log line and spc-12's overview, quoted in itd-48). Both are abcd's own records, not external citations, so they were swept for the single-word consistency this intent exists to enforce — recorded here rather than silently.

Out-of-scope confirmations: the spc- ID prefix is untouched (the intent renames the concept word, not the ID scheme); epic_id→spec_id was the prior atomic change; no schema files were altered.