← board

Track A from the job NAME test-emit-obj, not from its source. This job names a MECHANISM rather than a subject — the source it was fed (test/c_obj_data_dup_a.c) is what the mechanism was run ON, not what is being tested, so a lane guessed from it would be wrong by construction. The ranker reads frontmatter, so this line decides who works it; re-lane it if this job has changed what it covers.

origin/master has advanced 5 commit(s) since this sha. Re-verify at current HEAD before acting — the callback is tagged to the sha that was tested, which may no longer be the state of the tree.

regression: test-emit-obj#src:test/c_obj_data_dup_a.c at 1b59d1bbc364 in step 1/52, ./compiler/pascal26 --emit-obj test/c_obj_data_dup_a.c /tmp/cods_dup_a.o (auto-filed by twatch)

Repro

tools/testmgr.py --tier full --job 'test-emit-obj#src:test/c_obj_data_dup_a.c' at 1b59d1bbc364eed83812303246711467fd2dd7c5

Range

The named sha 1b59d1bbc364 CANNOT be the cause — it touches no buildable file (docs / tickets / tstate only). It is the sha that was TESTED, i.e. the upper bound of an untested range; the cause is somewhere below it.

bad 1b59d1bbc364, last good 1e5371209512, 2 commit(s) in range — the watcher narrows this by idle bisect; check tstate/TSTATE.md for the current range.

Log tail

pascal26: error: compiled successfully but wrote no output file: /tmp/testmgr-scratch-4008039/cods_dup_a.o
(tail)
pascal26: error: compiled successfully but wrote no output file: /tmp/testmgr-scratch-4008039/cods_dup_a.o
  the code was generated; the artefact is not on disk or is empty.
  usual cause: a missing or unwritable directory in that path.

Stub ticket: signal only. Track T agent (face 2) enriches or a dev track takes it from the repro line.

Log

RESURRECTED BY ACCIDENT, AND MERGED BACK — read this before the triage below

This ticket was closed by the seven watcher at 54964cd97 (04:52:10Z) as a clean rename backlog => done. Four minutes later dae71ffb2 added the triage below to devdocs/progress/backlog/… — a path that no longer existed. That tree predated the rename, git re-created the file at the old path, and the rebase merged silently because the two commits touched different paths. The result was one slug open and closed at once, indistinguishable from a live ticket in every instrument except a whole-tree duplicate census, and inflating exactly the open-bug count the owner wants at zero. Found by frankuser's census at 2a534ba32: 512 open slugs, 3770 terminal, one intersection — this one. Stub deleted, content merged here. See tools/twatch_duplicate_slug_devtest.py for the guard that now catches it.

AND THE CLOSE CORROBORATES THE TRIAGE RATHER THAN REPLACING IT

"closed: job green again" and "closed: proven non-causal" are different claims and only the first was on the record. Both are true here and the second is the one that generalises.

The job passed at d0e253df7e7a with no source change in the range — ten commits, zero touching compiler/, lib/ or tools/. A job that recovers on its own with nothing built differently is precisely what a non-causal range predicts, so the green is evidence FOR the finding below, not a reason to stop reading it. Had it stayed red, the range would still have been unsound; that it went green tells you which of the two remaining explanations — harness-side, or nondeterministic — is likelier.

Triage: this is not a source regression, and the range is not evidence

Added by frank-coordinator 2026-09-08, read-only measurement. The RED stands — only the attribution is impossible.

Every commit in the range is irrelevant to this job. Over 1e5371209512..1b59d1bbc364:

The header already says the named sha touches no buildable file. The range-wide statement is the one that ends the search: a reader told the accused is innocent still assumes one of the others is guilty.

Two of the ten commits are the coordinator's own prose-docs pushes. A seat that could not have caused a compiler defect was inside its blame population, and an idle bisect over this range would have converged on a pxx.skip edit and named it for an --emit-obj failure, with the confidence of a real bisect.

Disk: MEASURED and EXCLUDED, not merely unconsidered

Recorded because an untested hypothesis and an excluded one look identical to the next reader, and this one will be reached for again — the compiler's own message ("usual cause: a missing or unwritable directory") points straight at it, and ENOSPC presents identically to a write that silently produced nothing.

frankuser, read-only on seven at 04:50Z, thirteen minutes after the RED:

df -i /tmp -> 349,311 of 1,048,576 inodes used (34%)
df -h /tmp ->        2.3G of 47G used          (5%)

Neither resource was near a ceiling. The climb is monotonic between reaper firings and systemd-tmpfiles-clean.timer last fired 17:48Z the previous day, next at 17:48Z — so nothing freed inodes between 04:37Z and the reading, and whatever the count was at the failure it was at most 349k. ENOSPC is out in both dimensions, and the reaper did not race the run.

This exclusion has a date on it and does not generalise. Inode use was 834 after the previous cleanup and 349,311 a day later — roughly 350k/day against a 1,048,576 ceiling, so the volume reaches it in about two days and the outage recurs on its own schedule. /tmp/testmgr-* is the dominant producer at 192,118 inodes across 42 directories (~9,070 per tier run), not tstate-at.* at ~678 per directory — which the earlier outage census named because it ranked by standing total rather than by inodes per run. Appended to the prio-80 producer ticket at b617c3acb. If ENOSPC is out today it will not be out on Thursday, and a range guard that excluded the disk once must not be read as having excluded it always.

What is left

No source change, no disk, no reaper. That leaves harness-side or nondeterministic. Do not bisect this range; look at the box, the harness, and whether the job reproduces at HEAD.

Guarded in the tool as of this commit: range_non_causal() in tools/twatch.py classifies exactly this shape, the stub banner says it, and bisect_step declines. Positive control in tools/twatch_range_causality_devtest.py — a range containing a compiler/ commit must still bisect.