fix(ci): key the metric-history baseline on the executing stage

[yueming-yuan]

Chain 3/5 — stacked on b200-ci-hardware-axis.

RunIdentity took its suite from registry.suite — the stage a test declares in register_*_ci(suite=...) — rather than the stage the run actually executed on. The two are the same today, because a test only ever runs in the job whose name equals its suite, so the gate has been correct by coincidence.

That coincidence ends as soon as a test can be dispatched to a second GPU generation (4/5). A test homed on stage-c-8-gpu-h200 and executed on B200 would append Blackwell numbers to the Hopper series; mxfp8/nvfp4 values sit nowhere near their BF16/FP8 counterparts, so the gate would compare across populations and either fire constantly or widen until it detects nothing — losing the regression detection that already works on Hopper, as a side effect of adding Blackwell.

The write path is narrower but reachable: a PR carrying both nightly and a dispatch label resolves to nightly cadence, which writes, and would poison the stored baseline for good.

evaluate_gate now takes a required executing_suite. Required rather than defaulted to registry.suite: a silent fallback restores exactly the bug, and its symptom is slow data corruption rather than a crash. That costs ~40 updated call sites in the gate tests, all mechanical — the production change is 4 lines.

Two new tests pin the point: an absorbed run writes under the executing stage, and the two stages keep independent series (0.42 recorded against Blackwell does not regress against Hopper's 0.81).

No behaviour change — run_suite passes args.suite, which filter_tests already guarantees equals every selected test's declared suite.

🤖 Generated with Claude Code

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}