Eval caches include the GPU count and type

[renmengye]

The dispatched-eval determinant (DispatchedMeasurer._det) omitted the GPU count, although the baseline cache checks it and its docstring said the determinant covers it. A benchmark whose gpus changed, with the same image, command, tree and environment, could reuse a result measured with a different GPU count. The pre-budget lookup could also discount the gate charge on that stale entry. Found while reviewing the accelerators design (#454, phase 0).

• The determinant now includes the GPU count and the resolved GPU type from the lane, in the dispatched-eval slot, the baseline cache, and the pre-budget lookup that reduces the charge to one main eval.
• Cross-run baseline cache entries written without these fields are misses (re-measured once), never hits.

Compatibility (RELEASING.md; measurement caches):

• Surfaces: the dispatched-eval slot and scheduler names (via the determinant), and baseline cache entries.
• In-flight runs: a run parked with an eval written or running under the old slot adopts it. When the new slot is absent, the measurer looks for the legacy slot in the same run directory and uses it, so nothing is re-dispatched and nothing is charged twice. A current slot always takes precedence. A run's GPU count is fixed within the run, so this does not reopen the cross-run bug.
• Fixtures: tests/fixtures/dispatched_pre_gpu_identity/, written by the previous kernel, with a completed result and an in-flight job. Each is exercised through the new code: first pass, an idempotent second pass, a retry after an interruption, and a missing submit marker.
• Rollback: safe. The old kernel ignores the extra baseline fields and finds its own slots for evals written before the upgrade. Evals written by the new kernel are re-dispatched by an old kernel.
• Upgrading: line in the CHANGELOG.

Built by Codex. I reviewed it and asked for the in-flight adoption and fixtures. Gate: pytest, ruff check, ruff format --check, mypy.

[github-actions]

Round 1 — reviewed head 9d8df6a1 — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: 1 blocking, 0 advisory.

1 finding attached to the lines below.

Merged verdict: blocking legacy slot adoption can reuse an in-flight evaluation across a GPU lane type change. No findings rejected; the credentials, deployment, and lifecycle reports are one duplicate finding consolidated at the sharpest location and severity.

[github-actions]

Round 1 — reviewed head e25c1194 — reviewer hermes/gpt-5.6-terra.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: no defects found.