[ashhart]

Two source pins had gone stale, and CI never ran them, so they protected nothing and only blocked #2095.

• test-qwen-serving-host: 11 of its 14 file hashes no longer matched main. They're re-baselined to main's bytes.
• test-gradient-serving: 102 retained hashes are re-baselined, and ck_experts.cu leaves the manifest, since its own note says Flash Next's learning doesn't read those kernels. Approvals and HELD obligations stay.

No Living Weights device re-qualification is claimed. The dense learner's re-qualification stays owed, as the notes already say. From here, the next change to those kernels fails the check and needs a receipt again, which was the point of the pins.

Both steps pass on the host, and they failed before this. No engine, kernel or test logic changes.

[ashhart]

A note for anyone reading this later. 13 HELD rows get new hashes here, ck_lane.cu among them, so each HELD note now sits beside bytes that were never qualified: those are the device gates still owed. 63 of the 102 re-baselined entries aren't kernels (58 host Zig files, 5 Flash Next test files), so an edit to those tests will fail test-gradient-serving. Shrinking the manifest to what a learner reads, dropping the qmm_group.cu prefix pin that now cuts mid-word, and clearing the stale approval rows are follow-ups, not this PR.