fix(format): ModelConstraints dropped the gated-DeltaNet shape, so 18 of every 24 Qwen3.5 layers were counted as if their tensors did not exist

[noahgift]

#3346 asked which was wrong, the range or the descriptor. The answer is neither — it was the type that reads the descriptor. Stacked on #3348.

contracts/model-families/qwen3_5.yaml declares inner_size, state_size, conv_kernel, group_count and full_attention_interval. ModelConstraints carried none of them, so every consumer counted a hybrid model with dense/GQA accounting and silently missed the conv, the gates, the state norm and the mixer projections of 18 of every 24 layers.

Ground truth first, and it now matches exactly

The measurement is against a real file, not against the formula's own assumptions: ~/models/Qwen3.5-0.8B-Q4_K_M.gguf (sha256 bd258782…dc517), GGUF header parsed directly — 320 tensors, 752,393,024 parameters.

measured GGUF tensor sum   752,393,024
model_parameter_count      752,393,024      delta 0
  gated-DeltaNet layer      21,555,360  ==  measured blk.0
  full-attention layer      18,352,640  ==  measured

Two shapes in that file contradict dense accounting and are now modelled, both read off the tensors rather than assumed:

• attn_q is [1024, 4096] = 2·n_h·d_k — the q projection emits the output gate, and attn_output [2048,1024] proves only q is doubled.
• the file is tied: it has no output.weight.

The mutation: drop the attn_gate term and qwen35_0_8b_config_derived_count_equals_the_measured_gguf_inventory goes RED, short by exactly d·inner_size = 2,097,152.

QE2E-INV-001 is still NOT asserted, and the range was not widened

P(9B) moves 8,208,519,168 → 8,344,907,136, still 0.655 B below [9.0B, 9.2B]. The residual is not arithmetic — the 9b descriptor contradicts itself: it keeps inner_size: 2048 (the value the 0.8B uses at hidden_dim 1024) while quadrupling hidden_dim, and its group_count: 8 fails group_count · state_size == inner_size (8·128 ≠ 2048), a consistency the measured 0.8B satisfies at 16·128.

Only a real Qwen3.5-9B file settles that, and none is on this box, so the obligation stays unproved and the number is pinned by a test instead of asserted. DeltaNetShape::heads_span_the_mixer() now catches the descriptor's self-inconsistency directly.

Two disclosures

• Outside scope, unavoidable: five one-line deltanet: None, fixture literals in apr-cli tests/oracles. Zero logic change. The enum-payload alternative breaks the same crate, because apr-cli matches AttentionType::HybridGatedDeltaNet in two files.
• qk_norm is left alone deliberately. The descriptor says the family has none, the real file has attn_q_norm/attn_k_norm on its attention layers, and tensor_expectation.rs asserts the opposite. Both are true of different layer kinds — the repo should settle whether qk_norm is per-family or per-layer-kind. It is worth 4,096 parameters at 9B, so it does not touch the verdict.

The roadmap write pmat work add produced was reverted: #3297 moved entries to fragments, and a monolithic write is exactly the contention that PR removed.

Gate: aprender-core --lib model_arithmetic 28 passed, aprender-contracts --lib 1526 passed, clippy -D warnings clean, fmt clean.

Refs #3346, #3347, #3091, #3114

no-close: #3346 stays open — which side of the 9b descriptor is wrong is undecided until a real Qwen3.5-9B GGUF is measured.

ont-delta: shape ModelConstraints gains inner_size, state_size, conv_kernel, group_count and full_attention_interval, and model_parameter_count gains the gated-DeltaNet layer term; no obligation is added or removed, and QE2E-INV-001 is not discharged.

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3350 head=abf8b56acd24be766f3eb806d08216e321926715 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]

stale — APR-RELEASE-001 §6, applied mechanically 2026-09-20.

No activity since 2026-09-17 13:44Z; 2 release trains have been cut since (v0.68.2, v0.68.1). Age is the only input to this sweep — it is not a judgement on the work, and a train label such as pp-066/carry-* is not an exemption.

What happens next: if one more train passes while this is still labelled stale, §6 closes it. The branch is kept — closing a PR here discards a review thread, never commits.

To clear it: push, rebase, or say on the thread what it is waiting for. Any of the three removes the label at the next sweep. If it is blocked on something external, name that here — a blocker with an owner is not sprawl, and it stops the clock.

[noahgift]

Adopting this PR (2026-09-21 sweep; author session is not live — branch cut 09-16, no session on the box predates it). Plan, before touching anything: the DIRTY state is the squash-of-the-stack-base class — c5b7bce70 on this branch is byte-identical (git patch-id) to main's dfd3d5f96 (#3348's squash), so every hunk built on it conflicts. Fix is git rebase --onto origin/main c5b7bce70: replay only the three commits that are this PR's own, no content change, then run the model_family / model_arithmetic tests and the roadmap guards on the rebased head before a --force-with-lease push. If a real conflict appears in the replay I stop and report it here rather than resolve it by guessing. Refs #3346.

[noahgift]

Adopted and rebased — and the rebase found a real defect in the branch, now fixed.

What was done. git rebase --onto origin/main c5b7bce70 with the by-id roadmap driver bound: the first branch commit was byte-identical (git patch-id) to main's squash of #3348, so only this PR's own three commits were replayed. No conflicts. New head da1ee1506.

What the rebase found. Regenerating contracts/contracts.nt on the rebased head (the R-18 derivative rule) produced a graph 356 triples smaller than main — and the missing nodes were prune::run, distill::run, harness_ir::*, symbols this PR never touches. Cause: d3cc76f1f rewrote the QE2E-INV-001 note in contracts/binding.yaml and dropped the closing ", so the file stopped being valid YAML at line 764 (found unexpected end of stream). Every bound symbol after that entry vanished from the extraction. pv extract did not refuse; it emitted the smaller graph, and --check would have agreed with itself. One character, fixed in da1ee1506; binding.yaml parses again (156 entries, same as main) and the extraction is now byte-identical to main's committed graph, so this PR owes no contracts.nt change after all.

Verified on the new head: format::model_family + format::model_arithmetic 114 passed; apr-cli oracle 218 passed; ont4b_shapes_gate 11 passed; roadmap additive deleted=0 and fragment guard PASS; fmt and clippy -p aprender-core -D warnings clean; census unchanged. Measured with a pv built from this tree under a pinned target dir, and main's own graph re-derived with the same binary as the control.

Not armed: per the board ruling every PR needs a quorum receipt first; that runs next. Refs #3346.

[noahgift]

quorum-review (AD-04): NOT agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-3346",
 "head": "1ffdccf2500116a4dd39af6a66a697f7bb841136",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "FAIL",
   "findings": 2
  },
  {
   "lane": 2,
   "verdict": "NO-VERDICT",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 4
  }
 ]
}

[noahgift]

Quorum round 0: 1 FAIL / 1 no-verdict / 1 PASS — and the FAIL was a real finding, now fixed in e060c7e2c. Lane 1 (gemini-3.1-pro-high, cited model_arithmetic.rs:144): the projection term used q_out for a gated family's q matrix (2·n_h·d_k, measured in Qwen3.5-0.8B) while the bias term still used q_dim — a bias vector narrower than the projection it biases. One token (q_dim → q_out in the bias sum). It changed no shipped model's count — Qwen3.5 has no attention bias, Qwen2.5 has biases but is not gated — which is exactly how it survived a delta-0 measurement. The test that pinned 12 now asserts 16 and says why; a second test holds the other polarity (a non-gated family with biases stays at 12). 115 tests pass, oracle 218, clippy clean. Re-running the quorum on the new head. Refs #3346.

[noahgift]

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

{
 "ticket": "PMAT-3346",
 "head": "e060c7e2c22bc221b5e454a9b677d09d2b27c2e2",
 "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": 5
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 2
  }
 ]
}

[noahgift]

Cross-inspection of docs/audits/quorum-PMAT-3346.json lanes (non-author: aprender-04, 23:46Z)

lane	conversation	status	verdict	findings	refs to sibling lanes / receipts / $WORK	duration
lane 1	f3ea1f1c	SUCCESS	PASS (structured_output)	5 (2 cited)	0	644 s
lane 2	3f01ebe3	SUCCESS	PASS (structured_output)	0 (0 cited)	0	55 s
lane 3	81760200	SUCCESS	PASS (structured_output)	2 (0 cited)	0	394 s

Distinct agy conversation ids: 3/3; no lane references a sibling lane, another receipt or $WORK. Read from /tmp/claude-1000/-home-noah-src-aprender/51506a7a-265a-4e5c-bc53-50274de4a757/scratchpad/lanes-3346-round1 on this box. Note: round 0's pro lane found the gated-family bias-term defect (model_arithmetic.rs:144), fixed in e060c7e; this is round 1 on the new head; d4 armed via pmat-merge on the parent rule and this inspection is confirmatory.

Verdict line: 3/3 PASS, independent. Arm confirmed.

[noahgift]

Folded into the 0.69 release batch #3669 by the cop (08:05Z, operator: "most PRs can be batched"). One CI run and one queue slot for all of them, and the generated files (roadmap.yaml, census, graph, shapes, README count) regenerated once. This PR's receipt is in the batch tree unchanged, and its closing keywords are carried in #3669's body. Disarmed here so the queue doesn't take it twice. It closes as landed-in-#3669 when the batch merges. Don't push here; changes go to release/0.69-batch.

[noahgift]

Landed in #3669 (squash a877fa056, merged 2026-09-21T10:50:29Z). This PR's receipt stands as the constituent review; the batch folded its commits verbatim. — cop