Proposal (rationale)

link-integrity strips #anchor from a link target before resolving it and does not strip a trailing :line. So a markdown link to ../../web/src/app/features/forge/forge-store.ts:222 is reported broken while forge-store.ts sits exactly there.

That form is a deliberate code-reference convention, and it is the one CLAUDE.md mandates in this repo: “Reference code as file_path:line_number”. A gate that refuses the convention its own repo requires is not enforcing anything; it is generating work.

Measured 2026-08-16 on rift-forge (never installed, 3,623 commits), by rungs doctor --explain:

WI-038 set the threshold that this breaches — “a mis-framed rate above roughly one in five means the pass under-reports further before it ships” — and shipped anyway, because its triage was run against the three repos that happen not to use the convention, with a script that made the same assumption as the engine and therefore could not detect the class. That is corrected in WI-038’s Review; this item fixes the cause.

It affects rungs check in every consumer repo, not only --explain. It has been latent since WI-008 made link checking per-link — before that, any file containing a {{token}} had every link exempted, which masked it.

Decision

accepted — 2026-08-16. Opened from WI-038’s post-merge correction, as a child of WI-037. The epic’s requirement that no child add a new module or pattern holds: this changes one resolution rule in an existing engine.

Plan

Requirements

• A link whose target is <path>:<line> or <path>:<line>:<col> resolves against <path>, and is not reported when <path> exists.
• A link whose target is broken with or without the suffix is still reported. Stripping must narrow what is reported, never widen it.
• A path that genuinely contains a colon is not broken by the change.
• The self-test fixtures state both directions, as gates-self-tests-both-directions requires.
• The measured false-positive rate is re-derived across all four source repos after the fix, with a triage script that does not share the engine’s assumption.

Impacts

• src/engines.ts — link-integrity resolution only. gates-paths-exist runs the same scan (F-007), so both gates change together.
• modules/gates/gates/structural.toml — a self-test fixture for the new direction.
• Every consumer repo’s rungs check reports fewer link findings. That is the point, and it is a behaviour change worth stating in the item rather than discovering in a release.
• Risk: stripping too eagerly hides a real broken link whose name ends in digits after a colon. Bounded by only stripping when the stripped path exists.

Approach

Strip and re-test, never strip and assume. Resolve the target as written first; if that fails, strip a trailing :\d+(:\d+)? and resolve again; report only if both fail. A link is reported broken only when no reading of it resolves, which keeps the change monotone — it can only remove findings, never add them.

Rejected: treating :line as an anchor and stripping it unconditionally alongside #. It reads tidier and it is wrong — it would silence a genuinely missing foo.ts:12 by resolving a foo.ts that also does not exist, and worse, would make the two cases indistinguishable in the output.

Acceptance criteria / tests

1. A link to an existing file.ts:387 is not reported; a link to a missing file.ts:387 still is. Both as unit tests.
2. Self-test fixtures in structural.toml state both directions for the new rule.
3. The four source repos are re-triaged with a script that resolves :line independently of the engine; the false-positive rate is recorded per repo.
4. rungs check on this repo passes with no change in finding count (this repo’s links do not use the form outside code spans, so a change here would mean the fix is too broad).
5. npm test passes.

Out of scope

• Collapsing gates-links-resolve and gates-paths-exist. Still F-007; this item changes what both resolve, not that there are two.
• #L387 GitHub-style anchors. Already handled by the existing # strip. Nothing deferred.
• The 2,057 genuinely broken links in rift-forge. They belong to that repo, not to rungs.

Execution

Branch feature/WI-042-link-line-references, cut from main at d77c6bc.

One change of substance: src/engines.ts gains resolvesHere, which resolves a link target as written and only then retries without a trailing :line / :line:col. linkIntegrity calls it in place of its inline existsSync. Two self-test fixtures added to structural.toml, one unit test added, and linkIntegrity is now exported so the test can reach it without a registry.

No deviation from the plan. The strip-and-retest shape held; nothing needed widening.

Review

Verified 2026-08-16 on feature/WI-042-link-line-references.

1 · Existing file.ts:387 not reported; missing one still reported. Unit test link-integrity resolves a :line code reference but still reports a missing target covers four targets in one fixture — an existing file with :222, the same with :222:14, a missing file with :222, and a missing file bare. Exactly the two missing ones are reported. Met.

2 · Self-test fixtures state both directions. Added to structural.toml; gates-self-tests-both-directions passes. Worth naming what that gate does and does not do: it confirms both fixtures exist, and F-006 records that no fixture is ever executed. So criterion 2 is met as written, and the unit test in criterion 1 is what actually runs the assertion. Met — with the standing caveat that the self-tests remain declarations.

3 · Four source repos re-triaged, independently of the engine. The triage script resolves :line itself rather than reusing the engine’s assumption, which is the flaw that let this ship:

1,794 findings removed, all from rift-forge, none from anywhere else — which is the monotone property the approach was chosen for. Met.

Four of the 2,057 survivors were then checked by hand against the filesystem, and all four are genuinely broken — policy-boundary.md, ../backlog/FINDINGS.md and ../specs/data-sources.md are absent, and one is a literal unfilled ](url) placeholder in docs/research/lol-mechanics/README.md. The fix removed the false class without blunting the gate.

4 · rungs check unchanged on this repo. 20 pass · 0 fail · 0 unimplemented · 0 error, with the same finding count as before the change. This repo writes code references inside code spans, which WI-008 already excluded, so an unchanged count here is the evidence that the strip is not too broad. Met.

5 · npm test. 11 pass, 0 fail — up from 10. Met.

What this leaves

The 2,057 remaining rift-forge findings are that repo’s, not rungs’. F-007 (two gate ids, one scan) and F-006 (self-tests declared, never executed) are both untouched and both still open — F-006 in particular is why criterion 2 is weaker than it looks.