← board

Track T's own pushes destroy Track T's own breadth coverage

The finding

make_preempted() aborts idle work when any commit since the tested sha needs_test(), and:

NOTEST_PREFIXES = ("devdocs/", "docs/")

So tools/** is testable. That is correct and should not change — a testmgr.py change can change what every job reports, and a watcher that ignored its own tooling would be publishing verdicts about code it did not test.

The consequence is structural: T is the only lane whose ordinary work systematically destroys T's own coverage. Every other lane pushes compiler/** or lib/** and preempts a run that was testing a previous sha; T pushes the tool doing the testing, and the discard is total.

Measured 2026-08-19, self-inflicted and confirmed: pushing 8ec77190c (a twatch ticket-text fix) preempted the first full-on-HEAD run in 5h13m at ~207 of 2765 jobs. Abort count 52 → 53. The run had been reachable at all only because pin verify had retired minutes earlier — see bug-t-the-push-rate-starves-breadth-coverage-entirely.

Why a habit is not the fix

The obvious answer is "check .testmgr/watch.json for phase: testing, fast: false before pushing, and batch tooling commits". That is what T is doing now, and it is exactly the shape this repo keeps rejecting: a property that holds only because an agent remembered is a habit, not a property. It also fails open — a T agent that does not know the rule, or a cron-driven face 2, pushes straight through it.

Shape (sketch, not prescribed)

Ordered cheapest first; 1 and 2 are worth having regardless of the others.

  1. Refuse to be silent about it. tools/sync.sh (or a pre-push check) warns when an idle-phase full tier is in flight, naming the sha and the job count at risk. Cheap, no behaviour change, converts an invisible cost into a visible one.
  2. Make the cost recoverable rather than avoided — this is bug-t-the-push-rate-starves-breadth-coverage-entirely shape 2 (resumable backfill). A resumable run turns a T push from "discard 207 jobs" into "pause 207 jobs", which dissolves this ticket rather than managing it. If shape 2 lands, close this as fixed-by.
  3. Distinguish tool changes that can move a verdict from ones that cannot. A change to ticket-text generation (file_cascade_ticket, write_report_md prose) cannot change any job's result; a change to tier composition or timeouts obviously can. Tempting, and probably wrong to attempt: the classification is a judgment call the watcher would get silently wrong in the direction of under-testing, which is the one direction that matters. Recorded so the next person does not re-derive it as new.

Gate

Whatever lands: a devtest driving make_preempted over a scratch bare repo with a tools/-only commit, asserting the chosen behaviour explicitly — so the rule is a guard rather than a note.

Filed by Track T (plexus-T) 2026-08-19, after doing exactly this to itself and predicting it in the ticket before confirming it.

Update 2026-08-19 (plexus-T): option 2's premise was false, and is now repaired

This ticket says "if shape 2 lands, close this as fixed-by." Shape 2 had landed — e2449adc5, before this ticket was even filed — and it did nothing. Over its entire life the watcher saved 9 partials worth 1520 jobs and carried zero. There was one partial slot, and resume_arg() deleted a partial belonging to other work instead of declining it, so the very push this ticket is about destroyed the partial as well as the run.

So the fix-by condition was never met, and reading "shape 2 shipped" as "option 2 is available" would have closed this ticket on a mechanism that did not exist — the same substitution of code landed for effect measured the breadth ticket made. Root cause and fix: bug-t-a-saved-partial-is-evicted-by-the-next-run-of-different-work.

What is true now: partials live in a keyed store and provably survive an interleaved run of different work (devtest). What is not yet true: that a resumed slice has actually carried anything on the live watcher. That is one number — carried_runs leaving zero — and it is this ticket's fix-by condition as much as the other's. Until then option 1 (warn on a T push while a full tier is in flight) is still worth its low cost, because it is the only one of the three that does not depend on a measurement nobody has taken yet.

RESOLUTION 2026-08-26 — dissolved, but NOT by the mechanism this ticket

nominated

The ticket says: "If shape 2 lands, close this as fixed-by." Shape 2 landed on 2026-08-19 (e2449adc5). Closing on that would have been wrong, and the numbers say why: over the mechanism's whole life, 73 of 22,280 saved job-results were ever reused — 0.33% (saved_partials 83, carried_runs 3, superseded 70).

Not because resume is broken. superseded: 70 is the explanation and it is correct: a partial is keyed on (sha, tier), and on abort the watcher re-targets to the new HEAD — so the partial it just saved belongs to a sha nobody will ask about again. Resumability can only pay where the same (sha, tier) is retried, and a push-driven ladder almost never retries one. That is a structural ceiling, not a defect. It is why "a T push turns discard 207 jobs into pause 207 jobs" did not happen: the next run was not for that sha.

What actually dissolved it

A commitment point (572524c7c, 2026-08-25): an idle-ladder breadth run that is full_commit_secs (default 60s) in stops accepting preemption, and a reserved breadth run (7457f6aee) commits at zero seconds, because a reservation exists precisely for the case where pushes never stop.

So the premise — "any T tooling push aborts an in-flight idle-phase full tier and discards 100% of it" — is now false except inside a 60-second window at the very start of an idle-ladder run. The motivating incident, a push landing at ~207 of 2765 jobs, is far outside it and would survive today.

Outcome, measured across 3,828 run records: breadth full-to-full gap median 1.3h over the last 24h (worst 3.1h), against the 5h13m drought that opened the sibling ticket. See bug-t-the-push-rate-starves-breadth-coverage-entirely for the dated before/after.

Where resume DOES pay, so it is not written off

The one phase that retries a single sha: pin verify. It is deliberately left fully preemptive — the commitment point is breadth's alone — and it relies on resume instead, which works for it exactly because the key is stable. The live log: kept 432 decided job(s) from the aborted full run, and resume: partial accepted — 56 job(s) already decided against this exact binary. Resume was built for the wrong ticket and turns out to be load-bearing for a different one.

Shape 1 was not implemented, deliberately

"tools/sync.sh warns when an idle-phase full tier is in flight." With the commitment point in place, the cost it would make visible is a 60-second window on one phase — and a warning that fires on a cost that is usually zero trains people to ignore warnings. If breadth staleness regresses, this is the cheap thing to add back.

And the ticket's own instinct was right

a property that holds only because an agent remembered is a habit, not a property.

That is exactly why it should not have closed on shape 2: the property it wanted — a T push does not cost breadth — was never delivered by resume, and only looked delivered because carried_runs had left zero. A close condition can be about the wrong subject just as easily as a test can.

Log