Wire the self-test runner in, and fix the four defects that blocked it

Proposal (rationale)

Closes F-018, and with it F-006’s substance: gate self-tests were declared and never executed, so every fixture in the repo was documentation shaped like a test.

Three previous rounds (WI-045, WI-056, WI-057) built the runner and took the mismatches 17 → 7 → 3, each time finding real defects but never reaching zero, so the runner stayed unwired and no fixture ran on a normal rungs check.

Decision

accepted — 2026-08-16, directed by the user.

Plan

Requirements

Zero mismatches, then wire the runner into gates-self-tests-both-directions.
A fixture that disagrees with its engine fails the gate.
Fixtures that cannot be reproduced are reported, never counted as passes, and do not fail the gate — an unbuildable fixture is not a defect in the gates.

Approach

F-018 said the direct call was the oracle and the wiring had to be made to agree with it. Follow that, and check the oracle itself before trusting it.

Acceptance criteria / tests

Zero mismatches across all executable fixtures.
The wired gate fails on a fixture that disagrees with its engine.
Unrun fixtures are reported and do not fail the gate.
rungs check and npm test pass.

Out of scope

Making the remaining 45 fixtures runnable. Their shapes need context the format does not carry. Recorded, not attempted.

Execution

Branch feature/WI-059-selftest-wiring, cut from main at 0425173.

My own diagnosis was wrong, and checking it was the first step

F-018 recorded that “the runner is sound; the wiring is not”, on the evidence that session-sections-present returned ok, ok called directly and mismatch through gateMeta.

The oracle was wrong. I had called the runner with engine = 'sections' — inferred from the gate’s name. The registry says frontmatter-schema. So the two paths were not running the same thing, and the wiring had never been at fault.

Instrumenting gateMeta step by step is what showed it, in one line of output: engine: "frontmatter-schema".

Four defects, all of the same family


`session-sections-present` declared an engine whose table it does not have. `session.toml` holds `[sections]` and `[filename_schema]`; the gate asked for `frontmatter_schema`, got `undefined`, scanned nothing and passed by examining nothing on every run since it shipped. The handoff’s seven required sections had never been checked	fixed: `engine = "sections"`
`[register_schema.open]` was read by nothing. `registerSchema` only ever used the top-level spec, so the Open table’s `non_empty = ["Sev", "Pri", "What", "Evidence"]` and its Sev/Pri enums had never been enforced	fixed: sub-specs are iterated
The table matcher was a substring test. Once the Open spec ran, every Closed row failed it — because `## Closed — 2026-08-16 by [WI-044](archive/WI-044-resolve-open-findings.md)` contains “open”, inside a filename	fixed: the heading must start with the table name
The runner did not bridge `opted_in` → `extensions_opted_in`, so the fixture for an opted-in extension asserted the opposite of what it said	fixed

The first is the one worth pausing on. pass session-sections-present 0ms — no examined count — had been printed on every run of this repo for as long as the gate existed. The tell was in the output the whole time.

Review

Verified 2026-08-16.

1 · Zero mismatches. Across the registry: ok 17 · mismatch 0 · unrun 45. Met.

2 · The wired gate fails on a disagreeing fixture. Flipped one session-sections-present fixture from fail to pass and ran rungs check:

FAIL gates-self-tests-both-directions   23 examined
   gate 'session-sections-present' has no self-test expecting 'fail'
   self-test for 'session-sections-present' expected pass, fired: missing section 'Up next'

Both halves — the declaration check and the execution check — caught it. Restored, passes. Met.

3 · Unrun fixtures reported, not failing. 45 are named in a line on every run and the gate passes. Making them a finding was tried first and reverted: it would leave rungs check permanently red for fixtures whose shapes the format cannot express, which is how a gate gets disabled. Met.

4 · Gates and tests. rungs check 24 pass · 0 fail · 0 unimplemented · 0 error; npm test 24 pass. session-sections-present now reports 1 examined, and firing it was verified by renaming a heading in .ai/session.md: missing section 'Up next'. findings-disposition-has-reason went from 0 examined to 19. Met.

What this leaves

45 fixtures still have no builder — worktrees, workflows+similarity, values, changed, and a dozen other shapes that describe git state or multi-file scenarios. They are counted and named on every run rather than quietly skipped, which is the honest position: 17 of 62 executable fixtures now assert something, and the other 45 are visibly not asserting anything.

That is the number to improve next, and it is a format question rather than a harness one — but this time with a working runner to measure against, rather than a guess about what the format needs.