fix(e2e): normalize agent smoke status to lowercase "pass"/"fail"

[zhaoan12-prc]

Summary

Extraction gates check smoke === 'pass', but the schema accepted any string. Agents often returned text like "PASS (exit 0) ...", so a passing unittest was treated as a failed extraction and the kernel was skipped for optimization.

Changes

• Restrict smoke / unittest_smoke to "pass" or "fail" via schema enum; other values are rejected and the agent must re-answer.
• Add a smoke_detail field for free-text details.
• Document the rule in kernel_extractor.md.
• Add test_smoke_status_contract.js to CI L0.

Test

• New test passes on this branch and fails on main.
• Existing JS tests pass.

[chao-xu-spec]

Verdict: the fix is correct and complete, with no regression. I'd like two small changes before merge.

What I validated (MI355X ×8, GLM-5.2-MXFP4, sglang TP8)

Check	main (cfd5b58)	this PR (695c895)
Probe: extract agents on real GLM-5.2 kernel tasks, native Claude Workflow runtime	21 samples, all exactly pass/fail, 0 prose	21 samples, all exactly pass/fail
Probe: same tasks, standalone runtime (codex backend)	23 samples, all exactly pass/fail, 0 prose	23 samples, all exactly pass/fail
Full e2e, phases=head, fast mode, same carried setup state	dense GEMM head pass → bakeoff author_recommended → author; MoE head fail (real infra blocker) → dropped	identical behavior; agents also filled smoke_detail
Schema rejections seen in e2e transcripts	n/a	0

• The enum works as intended. In a probe where I pushed the agent to answer in prose, its first StructuredOutput was "PASS (exit 0) on GPU 4". The runtime rejected it with /unittest_smoke: must be equal to one of the allowed values: ["pass","fail"], and the agent re-answered "pass" in the same session. The gate then let it through.
• Gate coverage is complete. Smoke status only comes from EXTRACT_OP_SCHEMA.smoke and EXTRACT_SCHEMA.unittest_smoke, and every consumer compares against 'pass' (L1545, L3695, L4060, L4268, L4586). A schema-valid "pass" can no longer be read as a failure.
• I could not reproduce the original bug. On main, none of the 44 probe samples and none of the extraction results in the full e2e run were in prose form. The commit message cites the 0924/0925 runs. Could you link the log lines or journal entries where smoke was prose such as "PASS (exit 0) ..."? That would show how often this actually happens. It doesn't block the PR, since the change is a cheap and correct hardening either way.

---

1. Failure reasons move into smoke_detail, which nothing reads

On main, every extraction-failure log line builds its reason as notes || smoke. This PR keeps that unchanged:

• L3696: const why = ext ? ext.notes || ext.smoke : 'none';
• L4082: const why = p.ext ? p.ext.notes || p.ext.smoke : 'none';
• L4269: const why = ext ? ext.notes || ext.smoke : 'none';
• L4587: reason: `extraction failed/non-editable (${ext ? ext.notes || ext.unittest_smoke : 'none'})`

The PR changes where the agent writes the reason, though:

• The status field can now only hold "fail". The fallback ext.smoke used to carry text like "fail: gpu_lock refused ...", and now it yields just "fail". notes is not in required, so it can be empty.
• The new contract (kernel_extractor.md L82-83) tells the agent to put the command, exit code and caveats in smoke_detail. smoke_detail is defined in the schema (L845, L867) but nothing reads it.

Observed on real hardware. The task was glm52_dsa_indexer_topk, the standalone runtime used the codex backend, and in both runs gpu_lock refused to run because GPU 2 was busy:

# main: the cause is in notes, so it appears in the skip reason
unittest_smoke: "fail"
notes: "The prescribed smoke command exited 1 before unittest.py ran because gpu_lock
        reported that GPU 2 had foreign work running or VRAM above the permitted threshold. ..."

# this PR: the cause is only in smoke_detail; notes is unrelated, so the skip reason loses it
unittest_smoke: "fail"
notes: "The task uses synthesized fixed-seed inputs and has no reference_io.pt. ...
        No files under TASK were created, regenerated, captured, or edited."
smoke_detail: "Ran `... gpu_lock.sh 2 python3 unittest.py`. Exit code 1: GPU 2 had foreign work
               running above the lock's busy/VRAM threshold, so the lock refused measurement ..."

In the full e2e run the MoE notes happened to include the blocker, so the reason survived there. That was luck, not something the contract guarantees.

Suggested fix: apply this to all four sites, or factor it into a small helper.

-        const why = ext ? ext.notes || ext.smoke : 'none';
+        const why = ext ? [ext.smoke_detail, ext.notes].filter(Boolean).join(' | ') || ext.smoke : 'none';

A test assertion that each extraction-failure reason includes smoke_detail would keep this from regressing.

2. The new contract contradicts the exit-3 rule

The new section (kernel_extractor.md L81-82) says:

"pass" means the smoke run exited 0 AND correctness passed; anything else is "fail".

The existing rule (L491-496) says:

On smoke exit 3 ... REGENERATE the UT ... Do NOT record unittest_smoke:"fail" or drop the head for exit 3 ... Only after 3 failed regenerations set unittest_smoke:"fail"

An agent that reads "anything else is fail" first may report "fail" on exit 3 and skip the regeneration. Suggested wording:

-passed; anything else is `"fail"`.
+passed; a genuine failure is `"fail"` (exit 3 / `UT_HARNESS_INCOMPLETE` is NOT a failure yet:
+regenerate the UT first, see PHASE=extract step 6).

---

Summary: the change is good to go once these two items are addressed.

[zhaoan12-prc]

Addressed both review comments in 8b276e2:

• Failure reasons now preserve smoke_detail at all four extraction-failure sites.
• The exit-3 contract now requires UT regeneration before reporting failure.
• Added regression coverage, and the test passes locally.