skip is published as pass, so green does not mean "ran"

• Type: bug (Track T — tools/twatch.py, tstate schema)
• Opened: 2026-08-03 by claude@xeon, split out of [[bug-t-corpus-regex-invents-phantom-tree]] item 3. That ticket fixed the two self-contained halves (the phantom corpus name, and one absent corpus taking an unrelated regression test down with it) and deliberately left this one, because it changes the tstate schema and wants a migration rather than a drive-by.

The defect

tools/twatch.py:587:

now = {job_key(j): ("pass" if j["status"] == "skip" else j["status"]) ...}

A job skipped because its corpus tree is absent on this box is published as pass. The run-time warning is loud — !! CORPUS MISSING — 33 job(s) will SKIP — but the published state is silently green, and every consumer of tstate/*.json (cross-host comparison, --status, the dashboard, the cutover decision) reads it as covered.

The comment's reasoning is sound as far as it goes: mapping skip to pass closes an open regression when a box legitimately cannot run a job. The defect is that it does so by destroying the distinction, rather than by treating a third state as non-gating.

Why it matters

Green must mean "ran and passed". Concretely:

• On xeon, 33 of the full tier's jobs skip when the corpus trees are unfetched — including all 24 c-testsuite conformance jobs. The verdict is still GREEN.
• A per-host skip count is the only thing that makes host-to-host coverage differences visible at cutover time, which is precisely when the fleet decides to trust one box's green over another's.
• It was also how the phantom-corpus bug stayed invisible for so long: the affected job read pass on every host while having never executed anywhere.

Shape of the fix

1. Publish skip as its own status in tstate/<host>.json.
2. Keep it non-gating: new_red / still_red must not fire on it, and a job that goes red → skip still closes the open regression (today's behaviour).
3. Migration: existing borg.json / xeon.json carry pass for jobs that were skips. Readers must tolerate a missing/unknown status, so the safe order is (a) teach every reader the third state, (b) start writing it.
4. Report the per-host skip count in the tstate summary and the dashboard, so the coverage gap is visible without diffing two json files.

Gate

tools/testmgr.py --tier quick green, plus a devtest over diff_jobs covering: skip is not new-red; red → skip closes the regression; skip → red opens one.