bug(T): the full-suite hook keys on the TIER NAME, so it refuses the repro line every auto-filed regression ticket prints

.claude/hooks/no-full-suite.sh:118-120 matches testmgr\.py.*--tier[[:space:]]+[a-z]+ and exempts only --tier quick. It has no aperture for --job, which narrows a tier to a single case — so tools/testmgr.py --tier native --job 'test-core#src:...', a ~0.5s run of exactly one test, is refused with the text "every testmgr tier except quick is Track T's sweep — native, slow and opt cost the same ten minutes as full and limited." That sentence is true of the tier and false of the command.

Why it matters more than a one-off annoyance

Every auto-filed regression ticket prints this exact command as its ## Repro. Measured 2026-09-03 by folder (not by a glob across all of them):

backlog/ is where twatch files regressions, so the population is not "some tickets" — it is the entire auto-filed regression population, 100%. The watcher writes a repro the repo then declines to run.

The shape, which is the reason to fix it rather than route around it

The hook is a speed guardrail: its own header says so, and its refusal text prints the hatch. It is aimed at ten-minute sweeps. But it decides by reading the tier NAME, and a name is not a duration — --job changes the cost by three orders of magnitude without changing the word the hook matches on. This is the handbook's "the name is not the thing": an instrument that does not error, answers confidently, and is correct about the tier while being asked about the command.

It is not dangerous, because the hatch works and Track T's own commit trail shows agents lifting it correctly. What it costs is that the first thing an agent does with a fresh regression ticket is get refused, and the refusal argues at length that the command is expensive when it is not — so the agent must disbelieve a specific, confident message to proceed. A guardrail that is wrong in its reasoning teaches agents to discount guardrails that are right.

Suggested fix

Exempt a command that carries --job (and, likely, --only/--test if those exist) the same way --tier quick is exempted, at no-full-suite.sh:118-120. Keep the tier refusal otherwise. Add a positive control: a --tier native WITHOUT --job must still be refused, asserted — an exemption widened by hand is exactly the kind that stops being able to fail.

Provenance — the instrument caught itself

Filed by the coordinator after handing frankh-15 a verification command it then could not run. The command was written from the ticket's own ## Repro line and never executed by its author; frankh-15 ran it under PXX_ALLOW_FULL_SUITE=1, noted why in the commit (4a0de8788), and reported the refusal back. Two things worth keeping from that: a repro line copied out of a ticket is not a command anyone has run, and quick could not have substituted here — it does not cover test-core's native rows, so it would have returned GREEN on a job that was RED.

Related, all open, none covering this: decide-t-the-full-suite-hook-refuses-prose-about-the-suite, bug-a-no-full-suite-hook-refuses-make-n-and-misses-half-the-long-tiers, bug-t-no-full-suite-refuses-prose-in-a-non-git-compound-command. Four tickets against one 200-line hook is the smell the handbook names — count how many mechanisms serve one concept. Whoever takes this should read all four before editing, and consider whether the answer is a fifth patch or one rewrite that decides on the ESTIMATED COST of the command rather than on the words in it.

Log

• 2026-09-03 — resolved, commit wildcard-free --job aperture, plus the hook's first test wired into gate.sh quick.