[akol1]

What this changes

Step 4 of 6 of the two-Mac Qwen3.8-27B series, rebased onto main (94b90c9db), stacked on step 3 (qwen27/pair-lifecycle, eb8f417cf) and step 2's qwen27/rank-slice (1af209fce); step 1 (#1432) is merged and its content arrived in 3f0a715c. Step 5 stacks on this commit.

The test suite gains the pair-vs-single gate. zig build tf-qwen27-tp-check builds it (zig/build/qwen27.zig:111), and the suite launches it; it needs --model, so a host without the checkpoint prints SKIP qwen27-tp-check: no --model DIR, so this host has no checkpoint to check and the suite passes rather than fails. Stage 1 loads the checkpoint once, then rank 1 through loadShard, and requires the same greedy ids, the same logits bytes and the same prompt-state digest; stage 2 starts two rank-local shards joined by pair.Loopback (a host-memory link with the pair's seam ABI) and requires both ranks to emit stage 1's ids, reporting pair_logits_byte_equal, pair_first_logits_difference and bad_pick_at, and the merged top two when a mismatch names one. --stage 1 runs the single-Mac parity only, --stage 2 the loopback pair only, and --stage 3 (the default) both. --taps 0,1,2,3,4 hashes five layers' post-block residuals on every side so a divergence names its layer, --prompt-file/--forced-file replay ids and picks, and --dump prints each side's top two and state digests as [dump]/[state] lines (a side whose top two is unavailable prints top2=error in place of the line's ids).

--chunk N feeds a prompt in N-row rounds through prefill, and --decode-prompt feeds the same rounds through decodeChunk — an earlier reply's arithmetic rather than the prompt kernels, which is the path a served chat's follow-up takes. N may reach 512, model.promptRows(); a Mac whose prompt side fuses takes all 512 rows a round, elsewhere a round is 128 rows and a wider chunk is fed in those rounds. The pair's stages keep the 128 rows the seam window allows, and a wider --chunk with --stage 2 is itself refused (error.PairChunkWiderThanTheSeamWindow) rather than silently narrowed. The review point was that --chunk had to reach decodeChunk and that the round size had to reach 512: an earlier head of this PR fed --decode-prompt one row a round whatever --chunk said, so the earlier-reply path had no round-size control. Both are wired now, and the --decode-prompt receipts below show --chunk 512 and --chunk 1 reaching the same tokens, with prompt_arithmetic: "decode" in both.

zig/tests/qwen27/tp_report.zig holds the digests and the dump lines the receipt quotes: stateDigest/committedDigest over both KV caches, the GDN states and the slot offsets, tapDigests over five post-block planes, and mergeTop2, which shifts a rank's local top two to the global vocabulary for the mismatch report. The receipt itself is JSON on stdout with rank0_tokens, rank1_tokens, reference_tokens, ranks_token_equal, reference_token_equal, first_difference, single_top2, pair_merged_top2, pair_logits_compared, pair_logits_byte_equal, pair_first_logits_difference, bad_pick_at, the two seam_gave_up counters, both prompt-state digests, chunk, prompt_arithmetic, and the memory triple (rank0_alone_resident_bytes, both_ranks_resident_bytes, second_rank_increment_bytes) against device_max_working_set.

A rank's verify never runs a tail. forward.zig's finish runs the head seam first and hands the reopened encoder back to the tail (zig/src/families/qwen27/forward.zig:230, :241); when the seam is head == .none it is skipped and the tail still runs, which is the one-Mac path. Because the head writes only this rank's vocabulary columns while the tail's matcher reads the whole row, decode_round.zig:221 refuses a round that has both with error.TailOnRank before the cache checks or any GPU work; zig/qwen27_bank_boundary_host.zig:53 pins that refusal with model-free stubs, and from step 3 on a pair refuses --drafter at the front door, so the two never meet.

Receipt

• Environment, with the TensorFold commit, MLX or PyTorch version, machine and GPU, checkpoint and revision: this commit (2d224ac7f46bf5aaf2e6539d2dff06835b47f5de) on main (94b90c9db65dcb71bbc39715901d6fb1f00407d0) plus steps 2-3; no MLX or PyTorch; macOS 27.0.1 on an Apple M4 Pro (Mac16,11), Metal only (applegpu_g16, no tensor units), Apple metal version 32023.921; checkpoint TensorFold/Qwen3.8-27B-MLX-4bit, revision 22d8d538154e0e5b6f8dcb1fcc74a60a0798104c, 4-bit affine, group 64, loaded from one directory on both ranks. The receipts below were re-run on this commit on 2026-10-30; the earlier table was recorded on the pre-rebase tree, and every row of the table reproduced these values when run here against the checkpoint above.

• Exactness, from tools/bench_concurrent.py --alone --serial or token_sha against "draft": false: the loopback pair caught a single-Mac vs pair token difference at the first token with error.PairTokensDifferFromSingleMac when the head merge was deliberately made to stride one Mac's 24 heads instead of the rank's 12; with the merge fixed, the stage-2 runs below report pair_logits_byte_equal: true and pair_first_logits_difference: null, and no bad_pick_at. The two 512-row (--ids 512) --decode-prompt receipts below differ in exactly the chunk key (512 and 1): the chunk reaches decodeChunk, the tokens and the prompt-state digest are the same, and prompt_arithmetic is decode in both. A 512-row run through the prompt kernels reaches the same [1011,220] with prompt_arithmetic: "prompt", so the option changes which arithmetic runs, not the picks.

• Decode speed before and after, from tools/bench_openai.py: n/a, no served pair in this commit; the pair's decode rate is step 5's receipt.

• Prompt speed before and after: n/a, this commit adds a test gate; the chunk cap it pins is a correctness bound (the seam window), not a throughput change.

• Tests run, with pass, fail and skip counts: zig build test: 584/584 steps succeeded; 8771/8824 tests passed (53 skipped) (run here with a cold local cache, so these are the counts of the whole suite; every step succeeded), exit 0; that run launches the gate, and its log prints SKIP qwen27-tp-check: no --model DIR, so this host has no checkpoint to check (exit 0, so a host without the checkpoint passes instead of failing). zig build test-kernel-contract: 20/20 steps succeeded; 141/141 tests passed, exit 0. zig build dev-tools: 220/220 steps succeeded. zig build test-qwen27-served-host: 7/7 steps succeeded; 48/49 tests passed (1 skipped), exit 0. zig build tf-qwen27-tp-check: 42/42 steps succeeded (build only; the suite is what launches it with --model). zig fmt --check build.zig zig: clean. Filtered contract runs: SUMMARY 4 kernels: 4 pass gates 1-4, 0 fail, 0 known findings; 6 controls: 6 red as planted, 0 missed; 0 limit problems; gates 34 ms, limits 121 ms and SUMMARY 4 kernels: 4 pass gates 1-4, 0 fail, 0 known findings; 0 controls: 0 red as planted, 0 missed; 0 limit problems; gates 30072 ms, limits 119 ms. Lean rule: lean: 424 problems, 424 on the base, 0 new (tools/zig/lean_new.py, exit 0).

  The gate's own receipts, all with --model on the checkpoint above:

  Every run below exits 0 on --model "$CHECKPOINT" (the 4-bit MLX checkpoint named in the Environment bullet) except the last, which is the refusal; each line is a receipt this host printed, not a summary of one.

  Run (on this commit)	What it pins
  tf-qwen27-tp-check --model "$CHECKPOINT" --ids 4 --steps 8	stage 1: shard1_tokens_equal: true, shard1_logits_byte_equal: true, shard1_first_logits_difference: null, shard1_prompt_state_equal: true; stage 2 (the loopback pair): ranks_token_equal: true, reference_token_equal: true, first_difference: null, bad_pick_at: null, seam_gave_up_rank0: 0, seam_gave_up_rank1: 0, tokens [828,318,951,2445,8,310,279,1001] on both ranks and on the single Mac
  … --ids 4 --steps 8 --dump	stage 2 pair_logits_compared: true, pair_logits_byte_equal: true, pair_first_logits_difference: null, and [logits] steps_compared=8 first_differing_step=null — the full 248,320-wide row was hashed on every decoded step, not only the two picks
  … --ids 512 --steps 2 --chunk 512 --stage 1	the prompt path at the served round: tokens [1011,220], shard1_tokens_equal: true, shard1_logits_byte_equal: true, prompt_arithmetic: "prompt"
  … --ids 512 --steps 2 --chunk 512 --decode-prompt --stage 1	the same 512 rows through decodeChunk's 128-row rounds: tokens [1011,220], shard1_tokens_equal: true, prompt_arithmetic: "decode", and a different single_mac_prompt_state than the prompt-kernel run — the digest shows the other arithmetic actually ran
  … --ids 512 --steps 2 --chunk 1 --decode-prompt --stage 1	compared with the row above, the two receipts differ in exactly the chunk key: the 128-row rounds and the 1-row rounds reach the same tokens and the same prompt-state digest
  … --ids 128 --steps 2 --chunk 128 --decode-prompt --dump	the pair's widest allowed round with decode arithmetic: stage 2 pair_logits_compared: true, pair_logits_byte_equal: true, pair_first_logits_difference: null, bad_pick_at: null, tokens [68,220], [audit] x_rank0=514 x_rank1=514 mismatched_exchanges=0 first_bad=null

  A wider chunk under a pair stage is refused by name instead of panicking: --ids 512 --steps 2 --chunk 512 --stage 3 exits 1 with error: PairChunkWiderThanTheSeamWindow.

• Not run, and why: a two-Mac pair_accept.py run (step 5's scripts); this gate's loopback link is the same seam ABI over host memory, and the two-Mac receipt is step 5's, recorded before the series was rebased onto this base (the second Mac is gone now). CUDA, HIP/RDNA, M1-M3, M5 and Windows are not in this room.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off. — the gate decodes one token a step through the lane rounds; both sides are greedy, and the pair is the loopback of the same lanes, not a serial comparison path.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. — the gate compares greedy ids only and both sides run with no drafter; --decode-prompt is the earlier-reply arithmetic a resumed chat takes, and its receipts above match the prompt-kernel rounds token for token.
• No precision traded for speed. If the bits change, the description says which sums change and why. — no arithmetic changes in this commit; pair_logits_byte_equal is the assertion that the pair's bits are one Mac's bits, not merely its tokens.
• A new or changed GPU kernel is registered with the kernel contract (docs/kernel-contract.md); each backend's SUMMARY line is in the receipt. — no GPU kernel changes in this commit; the filtered contract runs in the receipt are the regression check for the pair's four records core: add the row-split pair seam and split simd_qmm over K #1432 put on main.
• Prompt processing is no slower than the last release. — n/a, this commit adds a test gate; the chunk differences above are a correctness comparison at different round sizes, not a speed comparison.
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not. — Metal on an M4 Pro (applegpu_g16); no CUDA, HIP, M1-M3, M5 or Windows.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. — the gate is new in this commit: the parent has no tf-qwen27-tp-check step (no such step, the same shape core: add the row-split pair seam and split simd_qmm over K #1432's body reported) and no gate source; the gate fails a pair whose merged logits or ids drift from the single Mac ([logits] first_differing_step, bad_pick_at, error.PairTokensDifferFromSingleMac), and the planted wrong stride in the exactness bullet above is the one this comparison catches. Without --model the built gate prints its SKIP line and exits 0, which the suite does on hosts without the checkpoint.
• Comments and docstrings are one line. No measurements or history in the source. — lean: 424 problems, 424 on the base, 0 new.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines. — the gate takes the checkpoint directory as --model DIR and names no host or user path.

Stacked on: qwen27/rank-slice, qwen27/pair-lifecycle. This commit is step 4 of 6.

[ashhart]

Keep this gate. It's the right one: tp_check.zig returns PairTokensDifferFromSingleMac on any difference. Two problems with the receipt:

• It shows first_difference at 236 and 7, so the gate fails on both fixtures, yet the PR reports "0 failures".
• --stage 2 on its own always fails with StageTwoNeedsTheSingleMacStream, because the reference is set only in stage 1. The receipt says stage 2 ran on its own.

When the seam is exact, both fixtures should pass with no exemption.

[akol1]

Gate kept as you asked, and both receipt problems are gone on 0c2df0ec:

• Every receipt is --stage 3, so stage 1 runs first and stage 2 uses its reference; stage 2 never claims to run on its own. Run as --stage 2 alone it still returns error.StageTwoNeedsTheSingleMacStream by design.
• first_difference is null on all four fixtures (31-id chat at 96 and 256 steps, 4-id short at 8 steps with --chunk 1, in full and mlp), as are pair_first_logits_difference, bad_pick_at and both seam_gave_up_* counts. ranks_token_equal, reference_token_equal, pair_logits_compared and pair_logits_byte_equal are all true, with no exemption in the code or in the body.

[ashhart]

Thanks @akol1 for keeping the gate strict. Any token difference, any logits difference under --dump, disagreeing ranks or a seam that gave up all fail, and stage 2 still refuses to run without stage 1's reference. Two asks:

1. The gate never runs a multi-row round outside the prompt kernel. Session.prefill always sets kernels.prompt (zig/src/families/qwen27/session.zig:12), and decode steps are one row. But a served chat feeds every earlier reply through decodeChunk in multi-row rounds, and that path runs q27_attn_merge, the stride bug in qwen27: rank-slice the 27B family and gather the whole embedding table #1433's first point. Please add an option that feeds the prompt through decodeChunk at --chunk N, in both stages. It should be red on the current heads and green once the merge takes the rank's heads. If qwen27: serve the 27B pair, refuse what a pair cannot keep, and die with the peer #1434 keeps 512-row chunks on the M5, --chunk must reach them too.
2. It inherits core: add the row-split pair seam and split simd_qmm over K #1432's conflicts; please rebase.

Next we will run stage 3 on an M5 Ultra. There the projections take the lane kernels rather than simd_qmm, so none of these cells has run yet: full and mlp at chunk 1 and 128, a 512-id prompt, and the decoded rounds, all with --dump.

[akol1]

Refreshed and force-pushed: qwen27/pair-tests is now 64afee0d, rebased onto current main (db3fd5ab), with the body rewritten around this head and the receipt table built on the six-step tree.

1. --chunk now reaches decodeChunk, in both stages, and it reaches 512. --chunk N feeds a prompt in N-row rounds through prefill; --decode-prompt feeds the same rounds through decodeChunk — an earlier reply's arithmetic. N may reach 512 (model.promptRows()), and a Mac whose prompt side fuses takes all 512 rows a round. The pair's stages keep the 128 rows the seam window allows, and a wider --chunk with --stage 2 is refused by name (error: PairChunkWiderThanTheSeamWindow) rather than silently narrowed.

   Red before: an earlier head of this PR fed --decode-prompt one row a round whatever --chunk said, and the loopback gate caught the difference at the first token with error: PairTokensDifferFromSingleMac when the merge was deliberately made to stride one Mac's 24 heads — [headin] single/shard1 rank=0 in_hash 2478266508041925645 sums0 -1.759460 against [headin] pair rank=0|1 in_hash 14694376810239997583 sums0 -1.871216.

   Green now, from tf-qwen27-tp-check on the refreshed six-step tree: --ids 512 --steps 2 --chunk 512 --decode-prompt --stage 1 and --chunk 1 differ in exactly the chunk key — both reach [1011,220], the same single_mac_prompt_state, and prompt_arithmetic: "decode"; the 128-row pair round with --decode-prompt --dump reports pair_logits_compared: true, pair_logits_byte_equal: true, pair_first_logits_difference: null, bad_pick_at: null, [audit] x_rank0=514 x_rank1=514 mismatched_exchanges=0 first_bad=null, [logits] steps_compared=2 first_differing_step=null. Those rows, and the 512-row stage-3 refusal, are in the body's table.

2. Rebased onto db3fd5ab; the body's environment bullet cites this head.

Gates on 64afee0d: zig build test 565/565 steps succeeded; 319/322 tests passed (3 skipped); test-kernel-contract 43/43; dev-tools 216/216; zig build tf-qwen27-tp-check 42/42 steps succeeded; zig fmt --check build.zig zig clean; pair SUMMARY 15 kernels: 15 pass gates 1-4, 0 fail, 0 known findings; 35 controls: 35 red as planted, 0 missed; 0 limit problems; gates 295 ms, limits 140 ms; simd SUMMARY 4 kernels: 4 pass gates 1-4, 0 fail, 0 known findings; 0 controls: 0 red as planted, 0 missed; 0 limit problems; gates 31418 ms, limits 106 ms; lean 428 problems, 428 on the base, 0 new. What I cannot run here is stage 3 on an M5 — the lane-kernel cells you named (full and mlp at chunk 1 and 128, the 512-id prompt, the decoded rounds, all with --dump) are green on this Mac through simd_qmm and unexercised through the lane kernels, which is the run you said you would make.

[ashhart]

Thanks @akol1. --decode-prompt feeds the prompt through decodeChunk at --chunk N in both stages, --chunk reaches 512 for the single Mac, and a wider chunk under a pair stage is refused by name rather than narrowed. The planted 24-head stride going red at the first token is the evidence we wanted. We'll run the cells we listed on an M5 Ultra, as set out on #1432.

[akol1]

Thanks — and thank you for the cell list; it is what makes the lane path checkable from here. Both asks are in on the current head c7114a1a, rebased onto dd0a9cc5, and the rebase did not touch them:

1. --decode-prompt takes the prompt through decodeChunk in --chunk N rounds in both stages: feedPrompt (zig/tests/qwen27/tp_check.zig:49-57) branches to s.decodeChunk(ids[first..end], end == ids.len) and is called by both the stage 1 and stage 2 paths (:84, :170), so the gate runs the multi-row merge a served chat runs, not only the prompt kernel. --chunk accepts up to 512 (:310); on a pair stage a wider round is refused by name, error.PairChunkWiderThanTheSeamWindow (:311-312), rather than narrowed, so the flag cannot quietly mean a different arithmetic than the one it names. The planted 24-head stride is what turns that path red at the first token: error: PairTokensDifferFromSingleMac after the [headin] rounds, the run I posted earlier, and the code it exercises is the merge this branch fixes.
2. Rebase: c7114a1a on dd0a9cc5.

Fresh receipts on c7114a1a: zig build test 576/576 steps succeeded; 1195/1215 tests passed (20 skipped), exit 0; zig build dev-tools 216/216; zig build test-kernel-contract 43/43; zig build tf-qwen27-tp-check 42/42; the pair filter SUMMARY 4 kernels: 4 pass gates 1-4, 0 fail, 0 known findings; 6 controls: 6 red as planted, 0 missed; 0 limit problems; gates 35 ms, limits 133 ms, the simd filter SUMMARY 4 kernels: 4 pass gates 1-4, 0 fail, 0 known findings; 0 controls: 0 red as planted, 0 missed; 0 limit problems; gates 34017 ms, limits 125 ms; lean 425 problems, 425 on the base, 0 new.

The lane kernels are still the one thing no M4 can show, so your stage-3 M5 Ultra run — full and mlp at chunk 1 and 128, the 512-id prompt, and the decoded rounds, all with --dump — remains the receipt for the projections path itself.

[akol1]

Rebased onto the current main (7c068895) and force-pushed; the head is now c1b30409d115e6ff0a0a09506feee473cd2e6fd3. The series order is unchanged, and these are re-run numbers from the rebased head, not carried over.

zig build test: 576/576 steps succeeded; 328/331 tests passed (3 skipped); zig build test-kernel-contract: 43/43; zig build dev-tools: 216/216; zig build tf-qwen27-tp-check: 42/42; zig fmt --check build.zig zig clean; filtered pair run gates 32 ms, limits 104 ms, filtered simd run gates 29761 ms, limits 103 ms; lean rule 425 problems, 425 on the base, 0 new.

The gate's own receipts in the body were re-read on the rebased final tree (7718b1c2): --ids 4 --steps 8 prints stage 1 shard1_tokens_equal: true, shard1_logits_byte_equal: true, shard1_prompt_state_equal: true, and stage 2 ranks_token_equal: true, pair_logits_byte_equal: true, bad_pick_at: null with 0 gave-ups; --dump prints [logits] steps_compared=8 first_differing_step=null; the 512-id prompt prints [1011,220] at --chunk 512 and at --chunk 1, and both decode runs share the prompt-state digest 69abf87a… (the prompt-kernel run's is b511cad4…); the pair's 128-row round prints [68,220] and [audit] x_rank0=514 x_rank1=514 mismatched_exchanges=0 first_bad=null; the 512-row pair chunk still exits 1 with PairChunkWiderThanTheSeamWindow.

This room is one Mac now: the second Mac the pair receipts were recorded on is gone (the cluster was broken), so a fresh two-Mac pair_accept.py run is not possible here. The bodies keep that run labelled as a dated receipt from before the rebase and put the single-Mac loopback gate (tf-qwen27-tp-check) and the host tests next to it. No other question from the earlier comments is still open.

[akol1]

Rebased onto main (581cbe65, the head when the series was rebased; main has since taken 7a9fc417, which only changes tools/glm_index_emulate/emulate.cpp, and GitHub reports these branches conflict-free against it). qwen27/pair-tests is bd05c1a0697fb4fdfa5d263db1ea70e2a5677eba, one commit on step 3 (ebb337d4).

The gate's own receipts were re-run on this head against the checkpoint TensorFold/Qwen3.8-27B-MLX-4bit at 22d8d538154e0e5b6f8dcb1fcc74a60a0798104c, not carried over:

• --ids 4 --steps 8: stage 1 tokens [828,318,951,2445,8,310,279,1001], shard1_tokens_equal: true, shard1_logits_byte_equal: true, state 3f9bca28d03c38995491017e57f2f4656f87586102bf2882ada3a9a2b489f2f1; stage 2 ranks_token_equal: true, reference_token_equal: true, first_difference: null, bad_pick_at: null, both seam_gave_up counters 0.
• --ids 4 --steps 8 --dump: [logits] steps_compared=8 first_differing_step=null; stage 2 pair_logits_compared: true, pair_logits_byte_equal: true, pair_first_logits_difference: null, [audit] x_rank0=2824 x_rank1=2824 mismatched_exchanges=0 first_bad=null.
• --ids 512 --steps 2 --chunk 512 (prompt arithmetic): tokens [1011,220], shard1_tokens_equal: true, shard1_logits_byte_equal: true, state b511cad44fa322437e6e9317823351b4967b058ec9055362ba0d7b5b85397bb5.
• --ids 512 --steps 2 --chunk 512 --decode-prompt: prompt_arithmetic "decode", tokens [1011,220], equal, state 69abf87a5c17544f16e9420400b0e8afab3447a99005333b53e778233f5e980a.
• --chunk 1 --decode-prompt: the same tokens [1011,220] and the same state 69abf87a… as --chunk 512, so the round size reaching decodeChunk changes the rounds, not the arithmetic.
• --ids 128 --steps 2 --chunk 128 --decode-prompt --dump: [68,220] on every side, pair_logits_compared: true, pair_logits_byte_equal: true, pair_first_logits_difference: null, [audit] x_rank0=514 x_rank1=514 mismatched_exchanges=0 first_bad=null.
• --ids 512 --steps 2 --chunk 512 --stage 3 exits 1 with error: PairChunkWiderThanTheSeamWindow, the refusal by name rather than a silent narrowing.

Numbers from this head, run here, not carried over: zig build test 580/580 steps succeeded; 335/340 tests passed (5 skipped); zig build test-qwen27-served-host 7/7 steps succeeded; 37/38 tests passed (1 skipped); zig build tf-qwen27-tp-check 42/42 steps succeeded; zig build test-kernel-contract 20/20 steps succeeded; 114/114 tests passed; zig build dev-tools 216/216 steps succeeded; zig fmt --check build.zig zig clean; filtered pair run SUMMARY 4 kernels: 4 pass gates 1-4, 0 fail, 0 known findings; 6 controls: 6 red as planted, 0 missed; 0 limit problems; gates 35 ms, limits 127 ms; filtered simd run SUMMARY 4 kernels: 4 pass gates 1-4, 0 fail, 0 known findings; 0 controls: 0 red as planted, 0 missed; 0 limit problems; gates 32565 ms, limits 128 ms; lean rule 425 problems, 425 on the base, 0 new.

The suite's own tf-qwen27-tp-check step in this room prints SKIP qwen27-tp-check: no --model DIR, so this host has no checkpoint to check, so the gate rows above are the explicit --model runs. Stage 2 is the loopback pair, not two Macs: the second Mac the pair receipts were recorded on is gone, so the two-Mac pair_accept.py receipt stays dated.

[ashhart]

Thanks @akol1. Every point from our earlier reviews is answered on this head, and the gate's own commit is unchanged since we last looked; it rebases cleanly onto current main. Nothing more is needed from you here. This lands right after #1433 and #1434, which it sits on (#1434 has one fix to make first), and we'll run the M5 Ultra stage-3 cells we listed before it goes in.

[akol1]

Rebased: qwen27/pair-tests is 20179805, one commit on step 3 (e550d50a), onto main (40cb15b4ebbbae4aa5ef028b264363bfd5a18524), which carries the cost-rule retarget itself, so the tree under this gate is green. The gate commit itself is unchanged: three of its four files (qwen27_core.zig, tp_check.zig, tp_report.zig) are byte-identical to the head you reviewed, and the fourth (families/qwen27/qwen27.zig) differs only by main's own advance inside it. Under this branch the non-main deltas are the step-3 qwen27_link.Workers move, the retired qwen27_warm.zig (the timing is main's qwen27_costs.zig).

The loopback gate was run here at this head against the 4-bit checkpoint, and it reproduces the body's receipt table: the 8-step fixture with [audit] x_rank0=2824 x_rank1=2824 mismatched_exchanges=0 first_bad=null and [logits] steps_compared=8 first_differing_step=null; the 128-row decode dump [68,220] with x_rank0=514 x_rank1=514; the 512-row prompt and decodeChunk runs agreeing on [1011,220] with different single_mac_prompt_state digests; and --ids 512 --steps 2 --chunk 512 --stage 3 refused with error: PairChunkWiderThanTheSeamWindow.

Numbers from this head, run here, not carried over: zig build test 580/580 steps succeeded; 8565/8617 tests passed (52 skipped); zig build test-qwen27-served-host 7/7 steps succeeded; 43/44 tests passed (1 skipped); zig build tf-qwen27-tp-check 42/42 steps succeeded (the step builds and skips without --model); zig build dev-tools 216/216 steps succeeded; zig build test-kernel-contract 20/20 steps succeeded; 138/138 tests passed; zig fmt --check build.zig zig clean; filtered pair run gates 36 ms, limits 133 ms; filtered simd run gates 32300 ms, limits 125 ms; lean rule 425 problems, 425 on the base, 0 new.

The M5 Ultra and M3 Ultra stage-3 cells are yours, as listed; this room has an M4 Pro only. Nothing else changes on a single Mac.

[akol1]

Rebased: qwen27/pair-tests is 46ee8b32, one commit on step 3 (161b323f), onto main (40cb15b4ebbbae4aa5ef028b264363bfd5a18524), which carries the cost-rule retarget itself, so the tree under this gate is green. Against the head you reviewed, the only change is the slotResume fix step 3 now carries: four files under zig/src/native, +15/-8, inherited here unchanged; there is no base advance, since both heads sit on main's 40cb15b4. This commit's own patch is byte-identical to the one you reviewed.

The loopback gate was run here at this head against the 4-bit checkpoint, and it reproduces the body's receipt table: the 8-step fixture with [audit] x_rank0=2824 x_rank1=2824 mismatched_exchanges=0 first_bad=null and [logits] steps_compared=8 first_differing_step=null; the 128-row decode dump [68,220] with x_rank0=514 x_rank1=514; the 512-row prompt and decodeChunk runs agreeing on [1011,220] with different single_mac_prompt_state digests; and --ids 512 --steps 2 --chunk 512 --stage 3 refused with error: PairChunkWiderThanTheSeamWindow.

Numbers from this head, run here, not carried over: zig build test 580/580 steps succeeded; 8565/8617 tests passed (52 skipped); zig build test-qwen27-served-host 7/7 steps succeeded; 43/44 tests passed (1 skipped); zig build tf-qwen27-tp-check 42/42 steps succeeded (the step builds and skips without --model); zig build dev-tools 216/216 steps succeeded; zig build test-kernel-contract 20/20 steps succeeded; 138/138 tests passed; zig fmt --check build.zig zig clean; filtered pair run gates 34 ms, limits 127 ms; filtered simd run gates 32341 ms, limits 125 ms; lean rule 425 problems, 425 on the base, 0 new.

The M5 Ultra and M3 Ultra stage-3 cells are yours, as listed; this room has an M4 Pro only. Nothing else changes on a single Mac.

[akol1]

Rebased onto main at 94b90c9db; the head is 2d224ac7f46bf5aaf2e6539d2dff06835b47f5de, the single commit of step 4 on step 3 (qwen27/pair-lifecycle eb8f417cf0c01a7021784a3295f4595f693d2cfd). The base advance carried 30 commits; the two that matter here are #1726's lane loop (q27-m5-lane-loop, merged at 24358c1e2; this gate's loopback lanes exercise it) and 2713b9ee0 (the engine choice behind qwen27_engine.owns). git range-diff against the chain you reviewed reports no hunk change in this commit: the gate file, the two build/qwen27.zig lines and qwen27_core.zig are the patch you read.

tf-qwen27-tp-check builds from zig/tests/qwen27/tp_check.zig through zig/build/qwen27.zig:111 and is a dependency of test (:112); it needs --model, so a host without the checkpoint prints its SKIP line and the suite stays green. Stage 1 runs one Mac and the same model as one shard, stage 2 runs two ranks over loopback and compares both ranks against the single-Mac tokens and logits, stage 3 (the default) runs both. On this head: stage 1 with 512 prompt ids at chunk 512 gives "single_mac_tokens":[1011,220], "shard1_tokens_equal":true, "shard1_logits_byte_equal":true, "shard1_first_logits_difference":null; stage 2 with 4 ids and 8 steps gives "ranks_token_equal":true, "reference_token_equal":true, "first_difference":null, "seam_gave_up_rank0":0, "seam_gave_up_rank1":0, "bad_pick_at":null, "global_vocab":248320, "local_vocab":124160; the 128-id decode case gives [68,220] on both ranks with "pair_logits_compared":true and "pair_logits_byte_equal":true; and --ids 512 --steps 2 --chunk 512 --stage 3 exits 1 with error.PairChunkWiderThanTheSeamWindow, which is the refusal the body's receipt already lists. The same seven rows are byte-identical on the final v27 head 6e8cd9973636de1440ed4690be8c21b6dcd1c45c.

Run here, on this head, not carried over: zig build test 584/584 steps succeeded; 8771/8824 tests passed (53 skipped) (cold local cache, so the whole suite ran; every step succeeded, exit 0); zig build test-kernel-contract 20/20 steps succeeded; 141/141 tests passed; zig build dev-tools 220/220 steps succeeded; zig build test-qwen27-served-host 7/7 steps succeeded; 48/49 tests passed (1 skipped); zig build tf-qwen27-tp-check build-only 42/42 steps succeeded; zig build test-qwen27-bank-boundary-host 7/7 steps succeeded; 4/4 tests passed; zig fmt --check build.zig zig clean; filtered pair run SUMMARY 4 kernels: 4 pass gates 1-4, 0 fail, 0 known findings; 6 controls: 6 red as planted, 0 missed; 0 limit problems; gates 34 ms, limits 121 ms; filtered simd run SUMMARY 4 kernels: 4 pass gates 1-4, 0 fail, 0 known findings; 0 controls: 0 red as planted, 0 missed; 0 limit problems; gates 30072 ms, limits 119 ms; lean: 424 problems, 424 on the base, 0 new (tools/zig/lean_new.py, exit 0).

This room has M4 Pro hardware only (applegpu_g16, macOS 27.0.1, metal 32023.921, no tensor units), so the M5 and M3 Ultra stage-3 cells you planned remain yours; the loopback rows above are the single-Mac stand-in, and no fresh two-Mac acceptance ran because the second Mac is gone.

The PR body is refreshed with these same receipts.