[noahgift]

Refs #4215. This covers (a), (b) and (c) from alt0693 REPORT.md §A. It does not close the 0.58× gap to ollama; measured below.

What changed:

• (a) Persistent embedding buffer: Qwen35CudaModel.hidden_buf is allocated once, and each token copies its embedding row into it.
• (b) Greedy GPU argmax: qwen35_gpu_decode at temperature 0 or top_k == 1 goes to qwen35_gpu_decode_greedy → forward_single_greedy → executor.gpu_argmax. The only download is the 4-byte token id; before, it was the 248,320-entry logits. The GPU argmax tie-break fix (GPU argmax kernel keeps the later index on an exact tie; CPU argmax keeps the first — breaks greedy token identity once #4215(b) moves argmax to GPU #4259, first index of an exact max, like the CPU) is merged in. It is in the batch(B1): emergency fold — serve/perf/runtime (9 sources, #4302 parked) #4311 fold too, and must land first or with this PR.
• (c) No format!/Vec on the launch path:
  • A module_key! cache interns (fmt, dims) → Arc<str> module keys.
  • The GDN launches pack their arguments into stack arrays (gdn_pack_args) and build arg_data only while recording a graph.
  • Converted wrappers: gdn_*, rmsnorm/per-head rmsnorm, residual, swiglu, the GEMV variants (q4k hw/mwv/dp4a, q6k, q5k, q4_0/q4_1/q5_0/q8_0/f32), q8_quantize, and gpu_argmax.
• New test counters: trueno_gpu::driver gains device_allocs_total() and device_to_host_bytes_total().

Red first, 0.8B Q4_K_M on the RTX 4090 via gpu-q. A counting global allocator measures steady-state tokens after warm-up:

per token	origin/main	this PR
host heap allocs	1713	0
device allocs	1	0
device→host bytes	993,280	4

• greedy_on_the_device_is_token_identical_to_the_cpu_argmax_of_the_logits: 38 tokens identical at temperature 0.
• The last 36 host allocs/token were found with a backtrace-tallying allocator, not by reading code: q5k_gemv_into's format! key on two DeltaNet projections per layer.

Decode tok/s at 850 prompt tokens (bench_greedy_decode_tok_s_at_850, ignored, release, 128 tokens, base and new interleaved on the same queue slot):

model	base (2 runs)	this PR (2 runs)
4B Q4_K_M	78.3, 78.3	80.2, 79.6 (+2%)
0.8B Q4_K_M	157.0, 166.8	172.6, 161.1 (noise-dominated)

Host overhead was not the bottleneck at this size; the remaining gap is in the kernels themselves.

Merge note: this touches the same executor files as #4258 (parked in #4311 over release/0.69.3 conflicts). I will merge origin/main into this branch after #4311 lands, with no rebase or force-push.

🤖 Generated with Claude Code

[noahgift]

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

{
 "ticket": "PMAT-4215",
 "head": "f80b3f82e2fe862eea432a8f477d529f7d4d7933",
 "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": 4
  }
 ]
}

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=4332 head=f80b3f82e2fe862eea432a8f477d529f7d4d7933 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]

Replaced by #4439: a fresh branch from current main with these commits cherry-picked (-x), per the operator ruling at 16:15 (no main-merges); the cop OK'd this close. #4439 passed a 3/3 quorum plus a 3/3 delta review; GPU tests green.