A shard killed by SIGNAL reports as a bare 'test run failed' — 11,757/40,750 not run on yoga-build3, nextest printed the number, nothing reads it; must report Unknown{Signal} with ran/not-run/total #3625

A shard killed by signal reports as a bare test run failed — indistinguishable to every consumer from a test failure. nextest prints the number of tests it did not run and nothing reads it.

Measured on #3621's workspace-test-shard (1/3), attempt 2 of run 35530342053, job 106147852695, runner yoga-build3, started 20:55:09Z. The log, in order:

573:  Cancelling due to signal: 4 tests still running
574:  TRY 1 TERM [   0.007s] aprender-serve gpu::tests::tests_06::test_forward_gpu_with_cache_out_of_bounds
576:    (test aborted with signal 15: SIGTERM)
578:  TRY 1 TERM [  12.227s] … test_super_long_context_memory_bound
585:  TRY 1 TERM [  81.902s] aprender-core::spec_checklist_q_qwen_coder q1_qwen25_coder_imports
593:  TRY 1 TERM [  13.500s] … test_mega_long_context_memory_bound
601:     Summary [ 242.884s] 28993/40750 tests run: 28993 passed, 217 skipped
602:  warning: 11757/40750 tests were not run due to signal

Line 573 precedes every TERM. Something on the host sent SIGTERM to the nextest process itself ~4 minutes in — no runner-shutdown message, no step timeout, job conclusion failure not cancelled. The four TERMed tests are nextest killing its own children as part of cancelling; one of them (test_forward_gpu_with_cache_out_of_bounds) had run for 0.007 s. Nothing had failed.

The step then exits 100 with error: test run failed, and the shard is red.

What this is not

It is not fail-fast. nextest's fail-fast wording is Cancelling due to test failure; that string is absent. nextest cancels on SIGTERM regardless of [profile.ci] fail-fast. Setting fail-fast = false (#3587's remaining half, a separate row) would not have run one of the 11,757. An earlier reading on this board attributed them to fail-fast; the log's ordering refutes it.

It is not a test defect. It is the runner-lottery class (#3607): yoga-build3 is a laptop, two of the four in-flight tests were *_memory_bound, and the host killed the process — the systemd-oomd-on-build-pressure shape. Land the same shard on intel and it runs.

What it is

#3587's doctrine — skipped is Unknown, not Fail — applied to the receipt. A shard that was killed has a verdict of Unknown{Signal} over 11,757 tests and a verdict of Pass over 28,993. Today it reports one bit: red. A real test failure on intel and a SIGTERM on a starving laptop produce the same red, and every consumer — the fan-in workspace-test, gate, the merge queue, the person reading the rollup — treats them identically.

The number is already in the log. 11757/40750 tests were not run due to signal is nextest's own line. The reducer discards it.

done_when

1. The shard's receipt carries ran / not-run / total and the reason for not-run (signal, timeout, test-failure), read from nextest's own summary lines rather than re-derived.
2. A shard cancelled by signal reports Unknown{Signal}, exit 2 — not Fail, not Pass. The fan-in workspace-test propagates Unknown as Unknown.
3. Runner name in the receipt. A signal-kill without the runner is a red nobody can attribute; with it, yoga-build3 becomes a named line on infra#813 rather than a mystery.
4. Falsifier, both arms: send SIGTERM to a running shard in a fixture and require Unknown{Signal} naming the count; a genuine test failure in the same fixture must still report Fail. Both, or the verdict is not discriminating.
5. The 11,757-class number appears in the fan-in's output, so the next person sees "11,757 unmeasured" instead of "shard 1/3 red".

Relationship to the other rows

• nextest fail-fast discards verdicts: one failure hid 15 unmeasured commands and the shard read as '1 failure', not '1 failure + 15 Unknown' #3587 / PMAT-3587: fail-fast discards verdicts — the default flips, and skipped becomes a number #3616: the explicit-commands runner now defaults to --no-fail-fast with ran/skipped/total. This row is the same shape for the nextest shards.
• [profile.ci] fail-fast = false (separate, being built): prevents a test failure from cancelling the rest. Its number is "7 rounds for 4 tests", not this one.
• #3540's red set varies across heads that differ by an 11-line quorum JSON: determinism (X64) and a PATH-resolution guard-cargo step — a runner lottery, not a flake #3607: the runner side. This is the receipt side.

Credit: the ordering reading and the correction of the attribution are the train/fleet lane's.

Refs #3587, #3616, #3607, #3621, infra#813.