Twinned from https://github.com/cfsh/ninefold/issues/598 by tools/beads/import_github.py's reconcile pass.
The first probe sweep ever run on the Linux box — merge_pr.py's post-merge gate for #558 at f06da59 — reported 28 passed, 6 failed. Five are the documented Linux-platform set. The sixth, explore-walk, is not documented, and does not reproduce.
| | |
|---|---|
| in the sweep, f06da59, Linux | FAIL explore-walk |
| standalone, same sha, same box, immediately after | PASS ×3 — exit 0, all 5 assertions OK, three consecutive runs |
⚠️ It is not from #558. That commit touches notes/, tools/office/ and one test file — no game code, no scene, nothing a probe observes. The gate's generated title says *"f06da59 regressed explore-walk …"*; that wording is wrong for all six, and this issue exists so the claim does not stand unexamined in the log.
Why it failed in the sweep. The per-probe detail was truncated in my capture of the merge output, and re-running the whole sweep to recover it is a 3–4 minute job I am not the right owner for. So the honest statement is *fails in a 34-probe sweep, passes 3/3 alone*, cause unknown — flake, ordering, or resource contention under the sweep are all live.
One artifact worth passing on, though it is present in the passing runs too and therefore not the cause:
``` ERROR: Unable to open file: res://.godot/imported/char2.png-1a287912085519e147e80773bd272dee.ctex. ```
A missing Godot import artifact for char2.png (the sprite Explore adopted in #434). It does not fail the probe — the assertions pass with it logged — but a missing import on a cold .godot is the kind of thing that behaves differently under a sweep than alone, so it is the first place I would look.
The five documented failures were measured and written down (probe_gate.py:85-110) precisely so a Linux sweep would not be read as a regression. A sixth, non-reproducing failure defeats that: the operator sees six names, knows five are expected, and has no way to classify the sixth without doing what I just did. Either explore-walk joins the documented set, or it is a flake worth fixing — but "six failed, five are fine" is not a state anyone can act on at a glance.
Related, and the reason it surfaced now: #566 made the gate runnable on Linux and Godot was installed here today, so this is the first time these probes have ever run on this platform. More of this class is likely.
Filed unclaimed by nf-dev-office-spawn while wrapping up #557. Not on my chunk — I surfaced it running the merge gate. Related: #566, #567, #584, #434.
None.
No comments.