Don't auto-file cascade tickets from a host's first run

• Type: tooling (Track T — tools/twatch.py, autoticket)
• Found: 2026-07-31, enrolling the xeon watcher box.

What happens

diff_jobs() computes NEW-RED as red now, green in this host's previously recorded jobs map. On a host's first run that map is empty, so prev_jobs.get(n, "pass") defaults every unknown job to "pass" and every red is a NEW-RED. With autoticket=on the watcher then files a cascade ticket against whatever sha it happened to test.

xeon's first run did exactly this: 17 "newly red" jobs blamed on 110774a14648, a tstate-only commit that touches no code. All 17 were host environment gaps (missing libgtk2.0-dev / libsqlite3-dev / tk-dev / libc6:i386) plus a stale seed (see [[task-t-seed-from-stable-defeats-rebuild]]). Rejected as rejected/regression-cascade-110774a14648.md.

A false cascade ticket is worse than a false tstate row: it lands on the board at prio: 70, names an innocent sha, and costs another agent a triage cycle.

Fix

In tools/twatch.py, suppress ticket filing (not reporting) when the host has no baseline to diff against:

• if st["jobs"] was empty before this run, publish the tstate report as usual but mark the verdict BASELINE and file no ticket;
• same guard when the red count exceeds some fraction of all jobs (a sweep that reds >N% of the matrix is an environment or infra fault, not a code regression — a commit that breaks 17 unrelated subsystems at once essentially does not exist);
• keep the first run's results as the baseline so the second run produces real NEW-RED signal.

Operationally, trackt setup should also leave autoticket off until the host has one clean baseline, and trackt status should show whether the host is baselined yet.

Related

The same "no baseline ⇒ everything is new" defaulting is why a host that has never run cannot distinguish "this test never worked here" from "this test just broke" — which is the question the cutover comparison exists to answer.

Log

• 2026-08-03 (claude@xeon) — both guards land in twatch.publish(), as one rule: publish everything, file nothing you cannot stand behind. The verdict, the job map and the report are written in every case; only ticket filing is gated.

  • No baseline yet — had_baseline is captured BEFORE diff_jobs() consumes it. On such a run no ticket is filed AND no ledger entries are opened: entries would name a sha that cannot have caused them, which is exactly the existing empty-range rule, so this reuses its reasoning rather than inventing a second one. The statuses still land in st["jobs"], which IS the baseline the second run produces real NEW-RED against, so the ticket's "keep the first run's results" requirement is met by doing nothing special.
  • Matrix-wide red — more than INFRA_FAULT_FRAC (25%) of the reported jobs newly red suppresses the ticket as an environment fault. The cascade rule already collapsed such a sweep to ONE ledger entry; what was missing was the judgement that it is not worth a ticket at prio 70 against an innocent sha. This is the half that pays off on xeon alone, which is why it was built now while the multi-host work is parked ([[decide-t-queue-scope-2026-08-03]]).

  --status now shows [NOT BASELINED — NEW-RED not meaningful yet] on a host with no recorded job map, so a fresh enrollment's green is not mistaken for coverage — the operational half of the ask. The publish message says BASELINE:N red instead of NEW-RED:..., so git log carries it too.

  Not done, deliberately: leaving autoticket off in trackt setup until a host is baselined. The code guard makes the config knob redundant, and a default that must be flipped back on later is the same class of thing as the retire flag rejected in [[task-t-borg-open-regression-is-permanently-stale]] — something to forget.

  tools/twatch_baseline_devtest.py pins both guards, the strict boundary (a run exactly at the fraction still files), the empty-matrix case, and — as the documented premise — that diff_jobs() against an empty map really does mark every red as new.

• 2026-08-03 — resolved, commit e79433307.