[diffusion][AMD] kernels: Fix fused QK-norm JIT kernels on ROCm and enable them

.agents/skills/ci-test-audit/SKILL.md

@@ -22,6 +22,31 @@ An audit is only useful if it is specific. "This suite could be trimmed" is not
 finding; "`TestFooLargePage` is a strict subset of `TestFooRetractLargePage`, drop it"
 is. Every finding needs the file, the pattern it matches, and what to do.
 
+## Auditing at scale
+
+An audit over hundreds of files (a whole tier, every CPU test) splits the reading from
+the judgment:
+
+- **Inventory first.** Collect each file's registrations, cost (live estimate x GPUs,
+  from the `sglang-ci-stats` model), and the exact items to check (CI-skipped classes,
+  overlapping configs) with a script, before any agent reads a file.
+- **Fan out file reads to the cheapest model** with a fixed brief and a TSV answer, in
+  chunks of about 15 files. The brief lists the exact items to verify; "review every
+  case in the file" makes the pass slow and adds nothing.
+- **Prosecutor stance.** Each file in the brief is deleted or demoted by default; keeping
+  it needs a reason: the change that would turn it red, and the grep that shows no other
+  test catches that change.
+- **The main agent re-judges** every verdict, most expensive first, before reporting.
+  Treat cheap-model "covered by X" or "superseded by Y" verdicts as leads: in one audit
+  none of 20 proposed file deletions held on verification.
+- **Re-tier before deleting.** For cost, moving a test down a tier (base, extra,
+  nightly, weekly) saves far more than deleting: the tier policy is in
+  [write-sglang-test, Choosing a Tier](../write-sglang-test/SKILL.md#choosing-a-tier).
+  Delete only what never runs in CI or is verified redundant.
+- **Failure evidence comes from failed runs' job logs.** `sglang-ci-stats` records only
+  successful jobs, so whether a tier catches regressions the tier above misses is read
+  from the failed scheduled runs on main.
+
 ## Reporting
 
 Group findings by confidence in the claim, not by pattern id. Lead with the ones where

.agents/skills/ci-test-audit/action-items.md

@@ -70,7 +70,7 @@ un-skip them; skips older than the issue they were waiting on; registrations nam
 retired `runner_config`; registered files whose TestCase classes never execute. Not
 this pattern: a hardware-conditional skip (`skipUnless` on compute capability), which is
 a gate; or a `disabled=` whose linked issue is still open, which is doing its job.
-Examples: #32324, #39368, #38585, #34377, #33772, #34779, #33654, #34070.
+Examples: #47603, #32324, #39368, #38585, #34377, #33772, #34779, #33654, #34070.
 
 ---
 
@@ -86,9 +86,12 @@ Examples: #34074, #33654, #33605, #34913, #35220.
 Spot: a multi-GPU accuracy suite registered per-commit whose failures have never been
 PR-specific; a nightly job that has been green for weeks, occupies several GPUs per run,
 and whose coverage a weekly run would hold just as well; a retired platform still
-running daily. Demoting is not
-deleting -- the nightly and weekly grids exist to hold this coverage at the right cost.
-Examples: #38011, #37990, #33809, #36814, #37532, #34204, #37586, #31749.
+running daily; a model test on a per-commit or extra tier for a model that is old, little
+used (check its downloads), or superseded by a newer model in its family on the same model
+file. Demoting is not deleting -- the nightly and weekly grids exist to hold this coverage
+at the right cost. The tier policy is in
+[write-sglang-test, Choosing a Tier](../write-sglang-test/SKILL.md#choosing-a-tier).
+Examples: #47619, #38011, #37990, #33809, #36814, #37532, #34204, #37586, #31749.
 
 **B3. Declared cost matches measured cost; shard counts and timeouts derive from it.**
 Spot: an `est_time` off from the measured run by an order of magnitude (the partitioner
@@ -221,7 +224,9 @@ copy of the package under test taking precedence over the working tree. Examples
 **F3. Repeated work inside the test process is cut.**
 Spot: a checker that re-parses the whole package on every scan; a fixture that reloads a
 tokenizer or model per case; independent subprocesses run serially; lint environments
-rebuilt on every run. Examples: #36240, #36235, #36241, #37203, #34309.
+rebuilt on every run; a suite of short files where each file pays the same import
+(measure the common test import with `python -X importtime` before reading test content).
+Examples: #47592, #36240, #36235, #36241, #37203, #34309.
 
 **F4. Work whose precondition already holds is skipped: installs, disk reclamation, cache refreshes.**
 Spot: unconditional installs; a skip check whose matching is too strict to fire on real