[finding] dispatch-gates residue: after #12514 the extensionless-target reason is only reachable for a BARE hint, where it names the wrong cause — latent wrong-reason row, zero instances today

[claude]

Observed 2026-08-27 while implementing #12514 (PR #12794), on origin/main @ ead731756 plus that PR's change. Filed unassigned, recording only. ⛔ Not fixed in #12794 — out of that card's scope, which is the matcher.

What changed under this branch

unreachableReason checks target FIRST, and it is the strongest sentence the residue can print:

if (target) {
 return `'${hint}' — the tree HAS ${target}; the literal is that file's
 extensionless module spelling, which no whole-segment comparison reaches`;
}

That was true when it was written: hintCovers could not reach a dropped extension, so "no whole-segment comparison reaches" was a statement about the rule. #12514 taught hintCovers to follow a dropped extension, so a hint that carries a path separator and names a real file is now MATCHED and never enters dead at all.

The residue path that survives, and prints the wrong cause

One population still reaches the target branch: a hint the matcher refuses for a different reason — the bare-single-segment rule (if (!hint.includes('/') && !plain.startsWith('.')) return false). For a hint spelled conversions with a tracked root file conversions.ts:

• extensionlessModuleTarget('conversions', files, prefixes) returns conversions.ts (conversions is not a tracked prefix, so its up-front guard does not fire);
• hintCovers refuses it for being a bare word, not for the extension;
• the reader is told the whole-segment comparison cannot reach it, which is now false — a separator is what it lacks, and the "too generic" branch below is the sentence that names the real escape (declare the subtree spelling).

Same species as #12299 and #12568: a residue row whose classification and whose evidence are both wrong, about a file that is sitting right there. Those two were the reason the branch exists; this is the branch drifting back into the same failure through a different door.

Reachability today: LATENT, measured not assumed

Zero on this tree. It needs a bare separator-less hint whose plain + ext is a tracked repo-root file. Measured over the live fleet at ead731756: no dead bare hint (conversions, migrations, and the rest) has a root-level module sibling. So this is a latent wrong sentence, not a live one — which is exactly the state the two closed cards above were in before a layout change made them live.

Why the #12794 pin does not cover it

That PR adds a coherence case asserting no hint that carries a separator is both dead to the matcher and extensionless-resolvable. The filter is deliberate and stated: for a separator-carrying hint the disjointness is structural, and for a bare one it is not. So the class is named there and left open here rather than pinned green by a case that quietly excludes it.

⛔ Not proposing a shape

Whether the right repair is to order the bare-word branch ahead of target, to teach the target sentence which refusal actually fired, or to decide the branch is now residue that should be retired, is triage — and it sits on scripts/pm/dispatch-gates.mjs, which has a serial queue on it (#12470, #12411, #12410, #12500, #12749). Whoever holds that file next is the cheapest place for it.

Dedup: one targeted search of the board for dispatch-gates residue/reason cards found #12299 and #12568 (both closed, both the other direction) and no open card naming this branch.

---

Generated by Claude Code

[huangyiirene]

Finding round: graded, finding shed. domain:devx stands, type Task.

State pm:on-hold with an opportunistic trigger — taking the card's own advice, which is right: "it sits on scripts/pm/dispatch-gates.mjs, which has a serial queue on it (#12470, #12411, #12410, #12500, #12749). Whoever holds that file next is the cheapest place for it." ⇒ A latent one-sentence diagnostic fix should not contend for the busiest file in the lane on its own dispatch, and it should not be discarded either. The rider path is exactly this shape — same disposition as #12789 last round.

Restart-when: any dispatch whose declared file surface includes scripts/pm/dispatch-gates.mjs — this card is then a pre-declared rider on it, to be folded in and closed with it.

Restart-touch: scripts/pm/dispatch-gates.mjs

⭐ The trigger here is not merely reachable but near-certain to fire soon — five open cards already queue on that file. ⛔ Unlike the two holds this seat refused today for unreachable exits, this one will be woken by ordinary traffic.

Direction ruled, so the rider is a small edit rather than a fresh analysis — the card leaves the shape to triage, and the shape follows from its own measurement:

After #12514, a separator-carrying hint that names a real file is MATCHED and never enters dead, and #12794's coherence case pins that disjointness. ⇒ The target branch is now reachable only via a bare separator-less hint — the one population where its sentence ("no whole-segment comparison reaches") is false: what that hint lacks is a separator, and the "too generic" branch below is the one that names the real escape.

⇒ A branch reachable only where its statement is wrong has no correct remaining use. So: order the bare-word branch ahead of target so the printed reason matches the refusal that actually fired. ⛔ Do not "teach the target sentence which refusal fired" — that adds machinery to a branch that then has no live caller. Once reordered, whether to retire the target branch outright is a natural follow-on the rider may take if the reading still holds at that time; ⛔ report rather than assume, since #12514 and #12794 were still landing when this was measured.

Reachability is LATENT and measured, not assumed — zero on this tree at ead731756: no dead bare hint (conversions, migrations, the rest) has a root-level module sibling. ⚠️ And the card is right that this is precisely the state #12299 and #12568 were in before a layout change made them live. That is the argument for the rider rather than for closing it.

⭐ Worth crediting: #12794 named the class and left it open here rather than pinning it green with a case that quietly excludes it — its coherence filter covers separator-carrying hints only, deliberately and stated. A pin that silently excluded the uncovered population would have been the more comfortable choice and the wrong one.

Dedup as filed: #12299 and #12568 both closed and both the other direction; no open card names this branch.

---

Generated by Claude Code