What is actually broken

check is a reporter wearing a gate's exit code, and the two populations it reports are not comparable:

class	what it means	acting on it
DUP-SLUG	one ticket file exists in two folders	corruption — two copies, one board row, `resolve` ambiguous
STALE-PARK / -HELD	a park's prose cites a now-closed ticket	read it; the slug matched, not the question
NEAR-DUP	two slugs are similar	usually fine
DANGLING-LINK	a wiki-link resolves to no ticket	five possible repairs, deleting rarely right
PROSE-EDGE-NOT-IN-FRONTMATTER	prose states a block the frontmatter lacks	fix on promotion, not before
DEAD-COMMIT	a citation names a sha absent from origin	historical, 350 of them

Every row but the first needs a human to read something before acting. The first is a fact that is always wrong and always mechanically fixable.

Measured, so the ratio is not a guess

$ tools/progress.sh check ; echo rc=$?
rc=1
   9 STALE-PARK-HELD    3 STALE-PARK    7 NEAR-DUP
   2 PROSE-EDGE-NOT-IN-FRONTMATTER    2 DANGLING-LINK    1 DEAD-COMMIT
   0 DUP-SLUG
$ grep -rn "progress.sh check" tools/gate.sh tools/sync.sh .claude/hooks/*.sh
(nothing)

The fix

Split severity, then wire only the hard half:

hard (DUP-SLUG, plus anything later found to be always-wrong-and-mechanical) → distinct exit code, or check --errors-only
advisory (everything above) → unchanged output, exit 0 in that mode

Then gate.sh quick can run the hard half for the cost of a directory walk, or Track T's watcher can run it on its ticket paths. Do not wire the whole check — 24 advisory findings would redden every gate in the fleet on the first run, and "a gate that cries wolf gets ignored, and then it is not a gate".

Positive control — BOTH directions, and the second is the one that matters

Copy a ticket into a second folder in a scratch tree; assert the hard mode exits nonzero and names it. (A control that only tests this direction passes on a mode that returns nonzero for everything.)
On a tree carrying only advisory findings — i.e. this repo today — assert the hard mode exits ZERO. This is the direction the whole ticket exists for, and a hard mode that still fails here is unwireable and has changed nothing.

What this is not

Not a request for new checks; check already finds the right things. Not a claim any current finding is wrong. The defect is that one channel carrying two severities collapses to the stricter one, so the output is unusable by a machine and therefore is read by nobody — the classic shape where a guard exists, works, and guards nothing.