fold(gpu-perf): #4576 + #4586 + #4532 — GGUF header-only read, dense F2 guard, sm_121 logits budget

[noahgift]

PRCAP fold (cop order 14:02Z, 24/10 CRIT). Fold = MOVE: each source PR's head commit is an ancestor of this branch (merged with --no-ff, so the original commits are kept). Built on origin/main 15f1b2a496. Fold head: 25fdbc4a12d0c2b62b517ad628aa089b95723c11.

Source PR	Head	Ancestor of fold head
#4576 perf(inspect): GGUF header-only read with a capped growing prefix (#3761)	8d12866154	yes
#4586 fix(#3602): dense F2 guard behind its receipt — --gpu setup 24.7s → 1.5s	2a56299f80	yes
#4532 fix(qwen35): per-arch e2e logits budget from measurement + sm_121 nightly lane (Closes #4523)	ef44857b9f	yes

Clean merge, fmt clean. Touches .github/workflows/cuda-nightly.yml (via #4532), so it needs a 3/3 non-Claude quorum.

Carried quorum/receipt files move with their commits. A fold changes the head, so this PR needs a new quorum receipt signed with --ticket listing every source ticket. Not armed (workers don't arm).

Workflow gate proof (cop 20:16Z: pre-approval conditions)

• Diff: git diff origin/main...HEAD -- .github/ touches .github/workflows/cuda-nightly.yml only: +25 −0. It adds one step, qwen35 CUDA parity module on sm_121 (aprender#4523), to the gx10 job. No existing step, check or if: is changed or removed.
• Permissions: the diff has no permissions: lines at all. No token scope changes.
• Gates: tightened only. The new step fails on a non-zero cargo rc, on any SKIP/is absent line, on a missing model file, and on zero tests selected.
• GREEN (CI): gh workflow run cuda-nightly.yml --ref fold/gpu-perf -f silicon=blackwell → run 36477877442 → NOT MEASURED. The run concluded success, but the decide step yielded (GPU busy (0 MiB, 1 procs) — yielding to training), so the new step was skipped. That is not a GREEN. It needs a re-dispatch while gx10 is idle (or force=true if the cop authorizes overriding the training yield).
• RED (offline, exact step script): the step run: block was extracted verbatim from the YAML and run with a stub cargo over fixture logs. That gives 6/6 rows as expected: green-3-pass rc=0; red-model-absent, red-cargo-fails, red-skip-line, red-zero-selected and red-no-summary each rc=1. A RED CI run was not dispatched. perf-gx10 keeps a single pending run, so a second dispatch could cancel main's nightly.

Agent: aprender-59

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=4607 head=b3faa936083aad4270e9e7f3916bc3ac986edce3 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

quorum-review (AD-04): three PASS — agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-3761,PMAT-3602,PMAT-4523",
 "head": "25fdbc4a12d0c2b62b517ad628aa089b95723c11",
 "width": 3,
 "executor": "agy",
 "agreed": true,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 3
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}