Status of the evidence, separated

Measured:

• RED on borg in every native-tier report from cc03b4a (2026-09-17T16:06) through e5614b0 (2026-09-18T08:39). 19 native reports in the window.
• GREEN by hand on plexus at HEAD: prints nine lines ending OK, rc=0.
• The assertion is tools/expect_same.sh test_object_value_type26 "$(... | tail -1)" "OK" — Makefile:20130. The compiler's ok: line is not an input to it.
• 578907347 is not an ancestor of ad85bf019 (merge-base --is-ancestor), and the ad85bf019 report already lists the row red quoting the pre-change spelling. The ok:-line change did not cause this.

Inferred, and by someone who could not check it: that the program produced no stdout on borg. The report's failure detail contains only the compile line, which is consistent with empty program output — but the extractor quotes whatever line it found, so this is a reading of an absence.

NOT evidence, and it looked like evidence. "18 reports name it on borg and zero name it anywhere else" is vacuous: borg is the only host producing tstate reports at all in that window (31 reports, all borg; 19 native, all borg). A census over a population that cannot contain the counterexample. So "host- specific" rests on ONE hand-run on plexus against borg's tier, not on a comparison across hosts. Recorded because it nearly went in as corroboration.

The one concrete difference anyone has

toolchain_fp: fbd303b097c6. Whether either matters is unknown — this test is pure Pascal and does not invoke gcc, so the gcc row is probably noise. It is here because it is the only measured difference between the two boxes.

First thing to do, and it is not debugging

Get the program's actual output on borg. Every reading so far is of a report that does not contain it. Run the tier narrowed to this one job on borg (the single---job form the auto-filed regression tickets print) and capture stdout — one line of output would replace this entire ticket's inference section.

Nobody on the fleet can reach borg, which is why this has sat for a day being re-read rather than resolved.

Do not

Do not chase the ok:/codeseg= spelling. That was checked and excluded above, and it is where two seats have already looked.