Proposal (rationale)

Seven findings were open, four of them opened during WI-037’s run and three inherited from 2026-08-15. A register that only grows stops being a register: FINDINGS.md states that it “counts what was recorded, not what was noticed”, and the same logic applies one level up — a row nobody ever closes is a row people stop writing.

This item takes each to a disposition. Not each to a fix: fixed, promoted and dismissed are all resolutions, and forcing everything into fixed is how a batch like this quietly becomes six half-done features.

Decision

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

Plan

Requirements

• Every row leaves the Open table with a written reason, per FINDINGS.md’s three dispositions.
• A finding whose fix is a design change is promoted, not attempted here.
• No fix ships without evidence it works and evidence it does not over-fire — the register’s own standard, and the one WI-042 was opened for missing.
• rungs check and npm test pass.

Impacts

• src/engines.ts, src/engines2.ts, src/cli.ts, modules/gates/gates/structural.toml.
• site/astro.config.mjs and a new site/scripts/check-vendored.mjs.
• One new repo-owned gate in .ai/gates.toml. Gate count 20 → 21.
• Two new items, WI-045 and WI-046.

Approach

Work each finding in ascending cost, and let measurement decide the ones the finding itself left open. F-007 offered two options and named the cheaper; I implemented the more expensive one, measured it at ten findings, all ten false, and narrowed it until it was clean rather than falling back — the fallback was still available the whole time and the measurement is what justified not taking it.

Acceptance criteria / tests

1. The Open table is empty; every row appears in Closed with a disposition and a reason.
2. Each fixed finding has a test or a self-test asserting both directions.
3. Each promoted finding names its item and what scope that item takes.
4. rungs check and npm test pass; the site builds with 0 broken links.

Out of scope

• Archiving the 39 done items. rungs backlog archive, which backlog/README.md §8 instructs contributors to use, does not exist — rungs backlog is an unknown command. Recorded as F-015 rather than worked around by hand: moving 39 files and recomputing every link repo-wide is exactly the mechanical repo-wide edit that goes wrong silently.
• Fixing what the promoted items cover. WI-045 and WI-046 have their own plans.

Execution

Branch chore/WI-044-resolve-open-findings, cut from main at 7e5f3c6.

F-001 — merged-status fires on a branch that landed nothing · fixed

The finding proposed “is an ancestor and has at least one commit of its own”. Implemented as base..branch > 0, it was wrong, and the probe caught it: after any merge a branch is zero commits ahead, so that version never fires again — it deletes the gate while looking like a fix. Had the test only covered the false-positive direction it would have shipped.

What distinguishes them is the merge commit. This repo merges --no-ff, so a branch that landed work is some merge commit’s second parent; one that landed nothing never is. Verified three ways in a throwaway repo — branch cut with no commits (silent), same with main advancing underneath (silent), branch with a commit then merged (fires). Known gap, stated in the code: a fast-forward merge that keeps the branch is read as having landed nothing. That is a workflow this repo does not use, and it is the direction to be wrong in.

F-003 — nothing verifies the vendored design-system shas · fixed

New site/scripts/check-vendored.mjs re-hashes all 33 files listed in VENDORED.md; registered as site-vendored-unedited, a repo-owned command gate placed outside every managed block so upgrade cannot rewrite it. The hashing is duplicated from the generator rather than imported, deliberately: a checker sharing the producer’s hash cannot catch the producer hashing wrongly, which is the mistake WI-042 was opened for.

F-007 — backticked_paths configured and unimplemented · fixed

The finding offered two options and called the collapse “smaller and honest”. I implemented the check instead, because gates-paths-exist’s provenance is a distinct measured incident — hexguard’s instruction files naming paths across 105 packages — that the markdown-link scan does not cover, and collapsing would have deleted it.

First run: ten findings on this repo, all ten wrong — slash commands (/work-item), branch placeholders (feature/WI-###-slug), illustrative directories (.cursor/rules/), bare filenames whose anchor is unknowable. Narrowed to the incident’s actual shape — contains /, has a file extension, resolved against both the repo root and the citing file’s directory — which takes it to zero, and it still catches a genuinely stale path.

The table is now one entry per gate id, so the two gates no longer run the identical scan and report everything twice. It fired immediately on real content: the why text of the gate added for F-003, in this same branch, cited src/design-system/VENDORED.md when the file is at site/src/design-system/VENDORED.md.

F-012 — ledger questions printed from check · fixed

Moved to doctor. The ADR does not merely name the command, it gives the reason — ADR-0005: “They must be pull (doctor), never push; no output during normal runs.” check is the normal run, so tier B was pushing, inside the feature that forbade pushing. The README’s example also named check-findings-register, which is not a gate id here; now audit-output-is-rows, which is.

F-013 — astro dev serves no stylesheet on Windows · fixed

fileURLToPath(new URL("../", import.meta.url)) instead of .pathname. Verified in the browser: 3 stylesheets load, h1 computes to Barlow 46px on the themed background, 0 console errors — against 49 consecutive 403s before.

F-006, F-011 — promoted

Both need a design change rather than a fix, and both are sized wrong for a register row:

• F-006 → WI-045. 27 fixtures are runnable text; the rest are ~8 bespoke structured shapes ({workflows, similarity}, {worktrees}, {values}, {matching_files}), each needing its own synthesizer. A runner covering only the easy half would reproduce the finding’s own complaint one level down.
• F-011 → WI-046. The Console component renders the literal label real output · <command> and is vendored — the directory is generated, now sha-gated by F-003’s own fix, and the export it comes from is gitignored and not in this checkout. So the cheap interim the finding suggested (rename the attribute) is not available from here.

Deviations from the plan

None of substance. Two fixes were harder than the finding predicted (F-001’s discriminator, F-007’s narrowing) and both are recorded above with what the measurement changed.

Review

Verified 2026-08-16 on chore/WI-044-resolve-open-findings.

1 · Open table empty, every row closed with a reason. Seven rows moved: five fixed, two promoted. Checked by reading FINDINGS.md after the change; findings-disposition-has-reason passes over all of them, which is the mechanical half.

2 · Both directions asserted per fix.

3 · Promoted findings name their item and scope. F-006 → WI-045, F-011 → WI-046, both with the reason the fix is a design change stated in the row itself. Met.

4 · Gates, tests, site. rungs check 21 pass · 0 fail · 0 unimplemented · 0 error (20 + the new vendored gate). npm test 14 pass, 0 fail, up from 12. Site builds; link check clean.

What this did not close

F-015, opened here rather than resolved: rungs backlog archive is documented in backlog/README.md §8 as the way to archive finished items, and the command does not exist. 39 done items sit in items/ and archive/ holds only its README. That is the same class as F-012 — a surface promising behaviour the code does not have — and it is why this item did not “mark items closed” by hand.