nightly: pi5 is red

[github-actions]

The scheduled workflow pi5 failed. This issue is kept by nightly-failures.yml: every further failure comments here, the first green run closes it.

First failure: pi5 run — c3f42a3, 2026-10-04

[geisten]

Looked at the failed run. Only one cell is over the limit: seq 32 prefill −6.3 % (limit −5 %). The other prefill cells are −4.5 % and −3.9 %, and decode is flat. The board reported throttled=0xe0000 after the run, which means a soft temperature limit, throttling and an ARM frequency cap all happened at some point since boot (the bits are sticky, so this doesn't say when).

ebeaf32..c3f42a3 is mostly cpu_x86/Metal/Vulkan. The shared changes that reach cpu_neon are the DeltaNet probe in exec_plan.c, transformer_gather_rows (PLE/lookup staging, Gemma only) and the Hadamard rotation plumbing. None of these runs on the BitNet 2B bench path per token, so I can't find a code cause.

I couldn't re-run the job: workflow_dispatch returns 403 for my integration. Question: could you run pi5 once by hand (without reset_baseline)? If it's green, tonight's run closes this anyway; if seq 32 is red again on a cool board, I'll bisect the shared commits above.

---

Generated by Claude Code