Reusable smoke workflow: let a caller choose the runner

[Jammy2211]

Enables autolens_workspace_test#271 (phase 2 of the jax-compile-stall epic). Ledger: PyAutoMind/draft/bug/ci/jax_vmap_jit_compile_stall.md.

Why

smoke-tests.yml hardcoded python .github/scripts/run_smoke.py. So the only way to run a different workspace script under the same ceremony — chain checkout at the matching branch, python setup, install epilogue, cache dirs, profile validation — was to copy that ceremony into a second workflow.

run_smoke.py's own docstring records what that costs:

This file used to carry a copy, and the copy had already drifted (its load_env_config hardcoded the smoke profile path, so the PR gate was structurally unable to read the release profile — the seed incident's failure mode 4/7). One resolver means the PR gate and the release runner cannot disagree about what a script's environment is.

The same argument applies one level up: one ceremony, so a diagnostic sweep and the PR gate cannot disagree about what environment a script runs in.

What changed

Three optional workflow_call inputs, all defaulting to today's behaviour exactly:

Input	Default	Effect
runner	.github/scripts/run_smoke.py	Workspace-relative script to execute
runner-args	""	Extra arguments, word-split
script-timeout	""	Exported as BUILD_SCRIPT_TIMEOUT, only when non-empty

Every existing caller is byte-identical in behaviour: none of them pass these, so runner resolves to the same path, runner-args expands to nothing, and BUILD_SCRIPT_TIMEOUT stays unset.

That last point is deliberate rather than incidental. Setting BUILD_SCRIPT_TIMEOUT unconditionally would have handed run_smoke.py an empty string to int() on every default run, and would also have stated a process-wide cap where today there is none — which risks overriding a profile's per-script overrides value that build_util.timeout_for is supposed to resolve. Hence the if [ -n "$SCRIPT_TIMEOUT" ] guard.

All three reach the run step through env: rather than ${{ }} interpolation inside run:. Callers will supply them from workflow_dispatch, and splicing dispatch input text straight into a shell body is a script-injection surface.

The caller this unblocks

autolens_workspace_test#271 has to establish whether 21 SLOW-marked JAX entries are genuinely slow or are the #245 XLA compile stall mislabelled. A slow script has a tight timing distribution; a stalling one is bimodal. Telling them apart needs the same script run N times under the real cap — which is a different runner, not a different workflow.

The workspace side is a small retime.yml with workflow_dispatch that uses: this workflow and passes runner: .github/scripts/retime.py. Those PRs follow once this merges; they cannot pass a runner input that does not exist yet, so this one goes first.

Verification

The change is inert on every current caller by construction — the defaults reproduce the previous invocation exactly. YAML parses and the input set is chain, python-versions, runner, runner-args, script-timeout, slack-channel-id. Real exercise comes from the first workspace retime.yml dispatch, and from every existing workspace's next smoke run continuing to behave as before.

---

Generated by Claude Code