Proposal (rationale)

gates-self-tests-both-directions executes a fixture only when its engine is on the runner’s context-free list and its shape has a builder. Recomputed 2026-09-06 with a scratch script over all fifteen module tables (.scratch/fixture-inventory.mjs, run from this checkout): 147 fixtures, 71 execute and agree, 69 are unrun, 7 disagree with their engine. This repository’s meta-gate reports 45 unrun and zero mismatches because it only examines the 26 gates registered here; the 7 mismatches live in modules this repo does not install, so no gate anywhere has ever run them.

Reading the seven, none is a fixture typo. Four are rules a shipped table configures and no engine reads — the F-007 shape, again: sections ignores required_subsections (specs) and the per-tier when condition (workflows, which therefore flags every tier-1 plan for tier-2 sections in three profiles); register-schema ignores requires_note (specs); file-population ignores exempt_marker (audit) and the structural similarity_* keys (ci, which therefore fires on any repo with eight workflows however different they are). Two are a builder writing a register row under no heading when the spec names one (design-sync). One is the builder joining paths with the platform separator, so on Windows the audit population never matches its own glob.

Among the unrun, the same reading finds findings-ids unable to detect a duplicate finding id (the engine extracts one id per file, and a register is one file), adr-index-current checking that a block exists and never that it is current (derives = "file-index" is unread), instructions-render-current passing a repo that has rule sources and has never rendered anything, and skills-description-routes reading a field_shape block nothing implements — so that gate has examined every consumer’s skills and checked nothing about their descriptions. These are consumer promises in shipped profiles, and the fixtures that would have caught them were documentation.

Decision

accepted — 2026-09-06 under WI-085. Give every fixture an explicit scenario — context declared in the fixture, builders keyed on fixture fields, adapters for engines that need a Git repository or a module root — and implement the table rules the fixtures assert where they are bounded. A rule that stays unimplemented is named as such in the runner’s output and in a producer-side test, never relabelled obsolete.

Plan

Requirements

• Every [[self_test]] in every bundled module resolves to one of: ok, mismatch (a gate failure), error (the engine threw — a gate failure, no longer converted into unrun), or unrun with a reason naming the unimplemented rule. no builder is not an acceptable reason at the end of this item.
• Fixture construction never branches on expect. Context a fixture needs — files that exist, an item’s status, a Git branch state, a registry row — is declared in the fixture and built the same way for both directions.
• The rules fixtures assert are implemented in their engines: when and required_subsections (sections), requires_note (register-schema), exempt_marker and structural similarity (file-population), field_shape (frontmatter-schema), row ids in a register (id-integrity), derives = "file-index" (render-freshness), no rendering anywhere (render-freshness), and the hook engine from WI-086.
• A producer-side test runs the inventory over all fifteen modules and asserts zero mismatch, zero error, and an unrun set equal to an explicit allowlist with reasons, so the seven invisible mismatches cannot recur invisibly.
• Existing verdicts on this repository and on the four source repos’ --explain output are preserved except where an implemented rule now legitimately fires; every such change is listed.

Impacts

• src/selftest.ts (builders, adapters, outcomes), src/engines*.ts (the rules above), src/engine-table.ts if a mapping moves, module gate tables (fixture context, never expected verdicts), test/core.test.js.
• Risk: implementing similarity or row-level ids may fire on real repositories. Each new rule is run against this repo and the four source repos before it ships; a rule above WI-053’s one-in-five false-positive threshold on any single repo is narrowed or left unrun with its reason.

Approach

Fix the seven mismatches first, because each is a shipped defect. Then builders by fixture shape, smallest first: files with N lines, populations, filenames, registers, then the Git-backed shapes (merged status, worktrees, merge drivers) reusing the repository builder change-requires-file already has. The meta-gate’s own fixtures need a module root the fixture controls; the ejected runner’s root override from WI-077 provides it and is reset in finally. design-mirror-not-edited depends on a rungs design pull command that does not exist and is recorded as unsupported with that reason, not built here.

Acceptance criteria / tests

1. The inventory script reports mismatch: 0, error: 0, and lists every remaining unrun with a reason that names an unimplemented rule; the count is recorded with the command and date.
2. rungs check here reports the unrun count that the allowlist predicts for the gates registered here, and gates-self-tests-both-directions still passes.
3. Each implemented rule fails on a seeded violation and passes when corrected, through the production runGates path, in test/core.test.js.
4. doctor --explain on hexguard, hexguard-templates, rift-forge and axiom-mesh changes only where a listed rule now fires; the diff is recorded per repo with its commit.
5. Full npm test, all registered gates and rungs modules pass.

Out of scope

• Building rungs design pull or any command a fixture presupposes.
• warn_at thresholds: the runner has no warning state, and inventing one is an ADR-0005 question, not a fixture repair. Recorded as a finding if still unimplemented at close.
• Rewriting fixtures to match what an engine happens to do today.

Execution

Executed 2026-09-06 on feature/WI-087-executable-self-test-coverage, cut from main 1670d1f (WI-086 landed). The scratch inventory (.scratch/fixture-inventory.mjs, gitignored) was the oracle throughout: 147 fixtures · 71 ok · 69 unrun · 7 mismatch at the start.

Runner. src/selftest.ts was rewritten around one rule: a builder is keyed on what the fixture declares and never on the verdict it expects. Every shipped shape has one — files of N lines, managed blocks, rule sources and renderings, derived index blocks, item files and the next-id marker, register rows, story files, workflow files sharing a stated fraction of structure, pointer documents, open findings with detail, skills naming or not naming peers, one skill’s description, a topic restated in a document, worktrees holding branches, merge drivers declared and installed, a merged or unmerged item branch, a board row and its item, the meta-gate examining a module directory the fixture builds (through a module-root override in src/ejected.ts, reset in finally), and content fixtures with tier, subsections and a spec-derived table heading. Context a fixture needs is now declared in it: link inputs name the files that exist, stale-blocker inputs name the finished item, rule-propagation inputs name the registry row, board inputs name the item’s status. Paths are written and returned with forward slashes — the audit population had never matched its own glob on Windows. An engine that throws is error, a gate failure; the two fixtures no engine can serve are unrun with the unimplemented rule named.

Rules the fixtures asserted and no engine read, now implemented: when and required_subsections (sections), field_shape (frontmatter-schema), requires_note (register-schema, note column configurable), exempt_marker and similarity_on/similarity_min (file-population, Jaccard over job ids, uses: and run: tokens), register-row ids (id-integrity), derives = "file-index" currency and never-rendered detection (render-freshness). render now fills a file-index block from its sources (renderDerivedBlocks), because a gate that compares rows to records would otherwise turn every consumer’s ADR index red with no command able to fix it.

Shipped defects the executing fixtures exposed, fixed in the same change:

1. Every exemption-marker check (engines.ts, engines2.ts, engines3.ts, the closure gate) matched marker\s*\S, and --> is a non-space character — so <!-- owner-ok: --> exempted. The reason must now be on the marker’s line and must not be the comment closer.
2. findings-self-declared-closure’s row pattern (engine default and the findings table) required a ] after the id; real rows are | F-054 | …, so the gate had examined nothing on any consumer.
3. Two tables shared one spec entry between two gates: instructions-render-current also judged the repo-map block (entry rules, renamed render), and skills-spec-pure would have judged descriptions once field_shape existed (entry skills, split into spec-pure and description-routes).
4. skills-description-routes, once real, failed two shipped skills: harden-rule and design-align named no trigger phrase. Both descriptions now say “Use when …”; this repo’s identical copy of harden-rule is updated with the module’s.
5. docauth-working-rules’ fourth fixture said “We used to …”, which the negation window rightly permits regardless of the marker, so it could never have failed for its stated reason; it now states a present-tense violation with a bare marker. specs-status-evidence’s rows lacked the Story column the register requires, so they fired for the wrong reason; rows now carry it, and a pass fixture with a note joins the fail fixture without one.

This repository. adr-index-current fired the moment currency was checked: the adr-index block in docs/decisions/README.md had been empty beside ten records since it was written, with a hand-kept table doing the index’s job below it. rungs render now fills the block; the hand-kept table is gone. rungs check reports 30 pass and, for the first time, no unrun fixtures.

Deviations. One the plan did not foresee. Filling the ADR index block at install made the packed journey call docs/decisions/README.md diverged: the record hashes the emitted bytes, and a filled block is not the emitted bytes. Ownership now hashes a module-owned file with every managed block’s body stripped (ownershipHash beside contentHash; used by the install record, upgrade planning and application, detection and the eject launcher check), with a recorded raw hash from an older record still accepted through either reading. The five owned files carrying blocks — the ADR and specs indexes, the ci trigger block, the session archive index, the shared registry — all hold generated bodies, so the rule is consistent; a deleted block still reads as an edit because the markers stay in the hash. warn_at stays unimplemented per Out of scope.

Found on the way. F-062: design-sync ships a gate that cannot function (local-modification detection and rungs design pull, neither of which exists).

Review

Against each acceptance criterion, 2026-09-06, Windows 11, Node v22.22.3, npm 10.9.8:

1. Inventory. node .scratch/fixture-inventory.mjs (scratch, gitignored): 148 fixtures · 146 ok · 2 unrun · 0 mismatch · 0 error. The two unrun are both design-mirror-not-edited, each with the detail “rule detect = "local-modification" is unimplemented, and the rungs design pull it presupposes does not exist”. Started the day at 147 · 71 · 69 · 7 (one fixture was added: the requires_note pass case).
2. This repository. node src/cli.ts check: 30 pass, and for the first time no “fixture(s) have no builder” line — every fixture of the 26 registered gates executes. The count the allowlist predicts here is zero, because design-sync is not installed.
3. Production path. test/core.test.js “the table rules the fixtures exposed…” seeds a violation and its correction through runGates for when (tier-1 plan passes, tier-2 plan fails), required_subsections, field_shape (and proves skills-spec-pure ignores wording), exempt_marker (bare vs reasoned), register-row ids, the closure row pattern (real row, bare marker, reasoned marker), file-index (0 rows fail; renderDerivedBlocks writes the two rows and the gate passes; a second render changes nothing) and never-rendered detection. “every shipped self-test fixture executes…” runs the inventory inside the suite and holds unrun to the explicit two-entry allowlist. “ownership ignores a filled managed block…” covers the ownership deviation.
4. --explain on the source repos, captured before any engine change and after all of them, ANSI stripped, diffed: hexguard-templates d24cf0aa, rift-forge 846cfa06 and axiom-mesh 3e1508a8 byte-identical. hexguard 51b25dac differs by two lines: the “29 self-test fixture(s) have no builder” line is gone because they run, and ci-workflow-proliferation reports 97 matching files instead of 99 — the similarity rule now excludes the two workflows whose structure differs from the rest, which is the count the gate’s message always promised.
5. Suite and gates. Focused: the eight ownership-related tests and the packed journey, 8 pass (1 host skip). Full serial node --test --test-concurrency=1 test/*.test.js on the final tree: 149 tests, 146 pass, 0 fail, 3 platform skips, 244 s. node src/cli.ts check: 30 pass, 0 fail. node src/cli.ts modules: audit clean.

Pending. The exact-SHA OS/Node matrix has not run: the branch is not pushed.