Qwen3.5 dogfood lane: blessed model, parity receipts at the serving shape, thinking-on closure #5317

Scope: The first-party Qwen3.5 dogfood: name the blessed size with its selection rule, parity receipts at the serving shape (including the decide-lane cell and the llama.cpp logit-compare oracle), the 0.8B thinking-on closure rate, the GB10 F2 guard rejection, FP8 prefill parity, the comparator cell, and classifying feature refusals in the ladder.

Why it matters: We should run the model we tell users to run, with the same receipts we ask of others.

Done when: One Qwen3.5 size is named with its rule, its parity receipt at the serving shape is committed, and the thinking-on leg meets its closure rate over N seeds.

Epic: #4381. Folded under triage rule 7b (2026-10-14): each row below was its own issue, closed as not planned and kept here as a sub-ticket. Reopen an issue when its work starts or its PR opens.

Sub-tickets

• Name the Qwen model 0.69 blesses and commit its parity receipt at the SERVING shape — our gates are all single-stream #3555 Name the Qwen model 0.69 blesses and commit its parity receipt at the SERVING shape — our gates are all single-stream (milestone 0.71.0); branch PMAT-3555-apr-bin-full-sha
  • accept: the outcome in the title, with a receipt
• First-party dogfood: Qwen3.5 at a Pareto-optimal size — publish the selection rule and the evidence a consumer re-derives it from #3558 First-party dogfood: Qwen3.5 at a Pareto-optimal size — publish the selection rule and the evidence a consumer re-derives it from (milestone 0.71.0); branch 76/3558-pareto-select
  • accept: the outcome in the title, with a receipt
• ARB-APR-3: parity receipt for the decide lane's cell on the agent's host (Qwen3.5-9B-Q4_K_M @ lambda-labs) #3574 ARB-APR-3: parity receipt for the decide lane's cell on the agent's host (Qwen3.5-9B-Q4_K_M @ lambda-labs) (milestone 0.71.0); branch c1/3574-parity-receipt-v2
  • accept: the outcome in the title, with a receipt
• PROMETHEUS (PRM-001 v3) epic: Qwen 3.5 4B local lane — fast enough and good enough to break quorum ties #4354 PROMETHEUS (PRM-001 v3) epic: Qwen 3.5 4B local lane — fast enough and good enough to break quorum ties (milestone 0.71.0)
  • accept: the outcome in the title, with a receipt
• apr: llama.cpp logit-compare verb as the PRM parity.oracle (PRM-001 v2 §4.1) #4444 apr: llama.cpp logit-compare verb as the PRM parity.oracle (PRM-001 v2 §4.1) (milestone 0.71.0); branch feat/4444-llamacpp-logit-oracle
  • accept: Acceptance:
• 0.70.2 P0: qwen35-0.8b-q4km golden_output thinking_on: think block unclosed within 2048 tokens (known failure in 0.70.1) #4666 0.70.2 P0: qwen35-0.8b-q4km golden_output thinking_on: think block unclosed within 2048 tokens (known failure in 0.70.1) (milestone 0.71.0)
  • accept: Done when: the ladder row is OK on lambda and gx10, and the known_red entry for this rung is deleted.
• Qwen3.5-0.8B thinking-ON leg: Qwen sampling over N seeds + a required closure rate instead of greedy (GATE CHANGE) #4696 Qwen3.5-0.8B thinking-ON leg: Qwen sampling over N seeds + a required closure rate instead of greedy (GATE CHANGE) (milestone 0.72.0); branch build-kaizen/4696-sampled-think-on-r2
  • accept: the outcome in the title, with a receipt
• Qwen3.5-4B Q4_K_M: F2 guard rejects the CUDA path on GB10 at the first decode position (cosine 0.8675); 0.70.3 too, sm_89 passes #4958 Qwen3.5-4B Q4_K_M: F2 guard rejects the CUDA path on GB10 at the first decode position (cosine 0.8675); 0.70.3 too, sm_89 passes (milestone 0.71.1); branch fix/4958-gb10-prefill
  • accept: the outcome in the title, with a receipt
• FP8 batched prefill fails CPU parity at the post-prompt decode step — catastrophic on per-head-QK-norm models (Qwen3-8B cos −0.10), marginal on Qwen2.5 (0.92) #3483 FP8 batched prefill fails CPU parity at the post-prompt decode step — catastrophic on per-head-QK-norm models (Qwen3-8B cos −0.10), marginal on Qwen2.5 (0.92) (milestone 0.72.0); branch la-71/3483-fp8-gate-delete
  • accept: the outcome in the title, with a receipt
• BLOCKER: no W1 cell is producible on the comparator lane in either mode — streaming reports prompt_tokens=0, non-streaming reports 534 against a 512 ± 8 band #2817 BLOCKER: no W1 cell is producible on the comparator lane in either mode — streaming reports prompt_tokens=0, non-streaming reports 534 against a 512 ± 8 band (milestone 0.72.0); branch fix/2817-stream-usage
  • accept: the outcome in the title, with a receipt
• G3.EX: 38 of 44 non-pass rows were classifier gaps, not defects — feature refusals need a re-run, not a skip #3182 G3.EX: 38 of 44 non-pass rows were classifier gaps, not defects — feature refusals need a re-run, not a skip (milestone backlog)
  • accept: the outcome in the title, with a receipt