Filed by the os-dev agent working #15602, from a measurement taken while landing that card. Filed unassigned, finding, for triage to price.
The gap
scripts/pm/bare-root-worklist.mjs reads a gate's own declared hint array back out of its source in exactly one place — the declaration pin — and that pin is scoped to one verdict:
for (const [key, v] of TRIAGE) {
if (v.verdict !== 'DECLARED-NARROWER' || !v.spelling) continue;
contradictedRows is the only other mechanism that compares a recorded refusal against what the gate declares, and it fires on r.covered — a row the sweep finds REACHABLE, i.e. one whose declaration reaches an arbitrary file at the top of the root.
So a row that records a REFUSAL while its gate declares a hint at that root that is genuinely NARROWER than the bare word falls between the two: the declaration pin skips it by verdict, contradicted skips it by reachability, and every remaining pin asks the TREE, which a declaration does not move. The record can assert "the population is a filter the idiom cannot spell" while the gate's own source spells one, indefinitely, with --self-test green.
That is not hypothetical. It is what #15602 was: check-declaration-mirrors declared both sides of its mirror when PR #15601 landed, the row went on recording REFUSE-UNSPELLABLE, and nothing anywhere reddened. That card fixes the ROW. This one is about the missing mechanism, which the same docblock already names in the general case: "only re-measuring catches this class".
Measured, at commit e581457b
Asking the question the missing pin would ask — for every REFUSE-* row, does its own gate declare any hint at that row's root, read through auditSource — gives exactly two rows today:
DECLARES REFUSE-UNSPELLABLE check-declaration-mirrors.mjs SCRIPTS_DIR scripts
DECLARES REFUSE-UNSPELLABLE check-driver-conformance.mjs CASE_SETS_DIR packages
The first is #15602 and is being fixed there. The second is a FALSE POSITIVE, and it is the whole design problem this card has to solve.
Why the obvious pin is wrong
Hints are declared per FILE, one array beside the constants. check-driver-conformance.mjs carries TWO rows at the packages root:
DRIVERS_DIR packages — DECLARED-NARROWER, and the declaration is ITS declaration;CASE_SETS_DIR packages — REFUSE-UNSPELLABLE, whose recorded reason already says so in as many words: "Its sibling constant took the escape; this one has nothing honest to declare".
A file-level read cannot tell those two apart, so a naive "a REFUSE row whose gate declares at that root is contradicted" reds a row that is correct, and correct for a reason already recorded. Attributing a declared hint to one CONSTANT rather than to the file is information the ROOT_DIR_WATCH_HINTS idiom does not carry today.
⇒ So this is a real edit with a decision in it, not a pin to bolt on. Three routes a fixer should price rather than inherit:
- Attribute by coverage, not by file — ask whether any declared hint COVERS the population that row's constant names. Needs the row's population, which the map has only as prose plus a number.
- Extend
omits downward — let a REFUSE row name the sibling declaration it is deliberately not claiming, the way a DECLARED-NARROWER row names a hint it passes over. Cheapest, and it puts the existing recorded sentence under a pin that re-checks it every run.
- Attribute in the idiom — carry the constant beside the hint at the declaration site. Widest blast radius: it touches every gate that declares, and
check:watch-hint-literal with them.
⛔ Not started here: #15602's scope was the row, and adding a pin that forces a re-decision on rows nobody measured is wider than the authorisation sentence that card ran under.
Why the priority is low, stated rather than assumed
Nothing reds today and nothing mis-runs: this map feeds no dispatch prompt, and the CI cost of the class is zero. What it costs is the map's trustworthiness — a recorded refusal that is false at its own gate's source, which is the defect class the file exists to make visible, arriving one level up inside the instrument.
Generated by Claude Code
Filed by the
os-devagent working #15602, from a measurement taken while landing that card. Filed unassigned,finding, for triage to price.The gap
scripts/pm/bare-root-worklist.mjsreads a gate's own declared hint array back out of its source in exactly one place — the declaration pin — and that pin is scoped to one verdict:contradictedRowsis the only other mechanism that compares a recorded refusal against what the gate declares, and it fires onr.covered— a row the sweep finds REACHABLE, i.e. one whose declaration reaches an arbitrary file at the top of the root.So a row that records a REFUSAL while its gate declares a hint at that root that is genuinely NARROWER than the bare word falls between the two: the declaration pin skips it by verdict,
contradictedskips it by reachability, and every remaining pin asks the TREE, which a declaration does not move. The record can assert "the population is a filter the idiom cannot spell" while the gate's own source spells one, indefinitely, with--self-testgreen.That is not hypothetical. It is what #15602 was:
check-declaration-mirrorsdeclared both sides of its mirror when PR #15601 landed, the row went on recording REFUSE-UNSPELLABLE, and nothing anywhere reddened. That card fixes the ROW. This one is about the missing mechanism, which the same docblock already names in the general case: "only re-measuring catches this class".Measured, at commit
e581457bAsking the question the missing pin would ask — for every REFUSE-* row, does its own gate declare any hint at that row's root, read through
auditSource— gives exactly two rows today:The first is #15602 and is being fixed there. The second is a FALSE POSITIVE, and it is the whole design problem this card has to solve.
Why the obvious pin is wrong
Hints are declared per FILE, one array beside the constants.
check-driver-conformance.mjscarries TWO rows at thepackagesroot:DRIVERS_DIR packages— DECLARED-NARROWER, and the declaration is ITS declaration;CASE_SETS_DIR packages— REFUSE-UNSPELLABLE, whose recorded reason already says so in as many words: "Its sibling constant took the escape; this one has nothing honest to declare".A file-level read cannot tell those two apart, so a naive "a REFUSE row whose gate declares at that root is contradicted" reds a row that is correct, and correct for a reason already recorded. Attributing a declared hint to one CONSTANT rather than to the file is information the
ROOT_DIR_WATCH_HINTSidiom does not carry today.⇒ So this is a real edit with a decision in it, not a pin to bolt on. Three routes a fixer should price rather than inherit:
omitsdownward — let a REFUSE row name the sibling declaration it is deliberately not claiming, the way a DECLARED-NARROWER row names a hint it passes over. Cheapest, and it puts the existing recorded sentence under a pin that re-checks it every run.check:watch-hint-literalwith them.⛔ Not started here: #15602's scope was the row, and adding a pin that forces a re-decision on rows nobody measured is wider than the authorisation sentence that card ran under.
Why the priority is low, stated rather than assumed
Nothing reds today and nothing mis-runs: this map feeds no dispatch prompt, and the CI cost of the class is zero. What it costs is the map's trustworthiness — a recorded refusal that is false at its own gate's source, which is the defect class the file exists to make visible, arriving one level up inside the instrument.
Generated by Claude Code