integration — fleet merges to main, 2026-10-02

Owner: integration lane (“исправь всё, добиваем оборванное”). Scope: inventory and merge of the branch wave that grouped on origin/main after the docs-heavy over-commit. Everything below is measured today on this box unless named otherwise. GPU discipline (§1.5) held: CUDA gates ran in one slot, in sequence.

Inventory, verdict, action

branch	base	ahead	verdict	action
wt/graphfix	2ccff38	2 commits	ГОТОВО	merged f2e9ab3; follow-up fix 054337d (see below)
wt/fidelity-fix	2db2aa6	2 commits	ГОТОВО	merged e790242
wt/kdafix2	f4f7a0b	2 commits	ГОТОВО (недописанный док)	merged d65ed16; its own findings doc was left uncommitted in the branch worktree — taken to main at 9a46ac2
wt/byteflow2	2ccff38	3 commits	ГОТОВО	merged (burn-byteflow crate + paper + findings); A/B byte-vs-byteflow is a queue item, not a merge gate
wt/fixverify	—	0 ahead	already in main	none
wt/carryfix	—	0 ahead (ancestor of main)	already in main	none
wt/cuda-gate	—	0 ahead (ancestor of main)	already in main	none

Related: wt/kdaci’s CI-proof doc (kda-oracle-ci-2026-10-02.md) is cited by the kdafix findings doc; taken to main at 07ceef0. The kdaci #[ignore] workaround itself stays superseded by the kdafix fix, per its own findings doc.

Gates

Run on the MERGED tree, not the branches (the branch-side claims were verified by re-running what is cheap, and the CUDA side on the live card):

• graphfix: dormouse-train lib 60/60 (the two new CPU half-tests fail the merge as-committed — see the follow-up fix below); CUDA target graph_seam_cuda with --test-threads=1: 6/7 green, all five differential/mechanism gates included the_noengram_key_contract_captures_ and_replays. the_stale_pointer_trap_is_reproduced_without_the_pin fails identically on pre-merge main (proved by running it at 08cd648) — pre- existing red, the graph-trainer lane’s subject, NOT introduced by this wave. The default parallel test run poisons one CUDA context and cascades — run this target single-threaded.
• fidelity-fix: dormouse-core lib 89/89; burn-mhc lib 9/9 (the new b_res_starts_at_zero_not_the_10i_stiffness pin green); burn-attnres lib 11/11 (the gain fixture green); inverted init gates (the two > 0.05 / > 0.3 asserts) measure the FIXED init, so a drift back onto the 10I stiffness re-reddens them.
• kdafix2: burn-kda lib 12/12; --test kda_oracle 9/9 (the designed reds are green: kimi_linear_softplus_decay_matches_fla_reference, read_scale_matches_fla_reference); burn-gdn2 ops_batched_autodiff (autodiff feature) 1/1; CUDA --test fused_cuda 4/4 on hardware — and the gate did not even compile off the shelf (see the fix commit). --test bench_cuda 1 ignored (by design).
• byteflow2: burn-byteflow crate tests 14/14 + facade test + oracle targets green (CPU).
• doc-gate (tools/check_doc_refs.py): was RED on 9 citations of the missing docs/reviews/kdafix-2026-10-02.md + its cited evidence doc; green after 9a46ac2 / 07ceef0 (0 missing paths).
• 1-step smoke on the release binary of the fully merged tree (real corpus, small/batch 8/s512): green — step 0 ce=5.571 bpb=8.038, aux=0.0289, retr_arm batched:0/factor:1, checkpoint saved, no NaN. (Step-0 numbers are autotune-cold, not measurements; §3.1.)
• preset_exec queued under the build lock at the end of the day’s queue (agents stage-gate-b, maxopt2 held it) — see the follow-up note.

The follow-up fixes (each its own commit)

1. 054337d fix(graph): the keys gates precede the pin copies. The branch’s own doc comment promised the key-less quiet path does not touch the copy arms, but the gate sat AFTER copy_into_int(x)/copy_into_int(y) — on any non-CUDA build the copies refuse loudly and both new CPU half-tests failed on the pin refusal instead of testing the keys gate. Hoisted the consumption gate and the None-keys refusal above the copies; CUDA ordering for a real keyed feed unchanged.
2. fe69e5f fix(kda): fused_cuda gate compiles after the read-scale merge. The kdafix2 branch read q.shape() at the chunk_wy_forward call site, but main’s state of the test had already moved q bare into the fused call — E0382 under cuda,autodiff features ONLY, invisible to every CPU gate. The head dim is captured before the move.
3. 9a46ac2/07ceef0 docs(kda): the kdafix findings doc and its cited kdaci evidence doc — both existed only in other lanes’ worktrees uncommitted, while 9 committed files + the doc-gate pointed at them.

Pre-existing reds found on the way (NOT this wave’s defects, carried to their owners)

• the_stale_pointer_trap_is_reproduced_without_the_pin (graph_seam_cuda) — fails on pre-merge main identically; same crash class that killed the graph-trainer lane’s gbench2_graph run ~18:48 today (server.rs:144, illegal address). Owner: graph-trainer / cuda-graph lane.
• gated_delta_chunk_path_runs_at_the_production_shape (vendor/dormouse-fused, burn-fused gpu_production_shape, cuda+autodiff, release) — matmul inner-dim mismatch [10,12,16,64] × [10,12,512,64], fails identically with burn-kda/burn-gdn2 checked out at the kdafix2 base f4f7a0b. Pre-existing red → tools/gpu-gate.sh is red on main for reasons outside this wave. Owner: burn-gdn2 / chunk-path lane.
• tools/lib_gate.sh:95 — the rc-overwriting red-mask (rc=$rc2) reported by the wgpu and dispatch-guard lanes: still unfixed, still a shared tool.
• fuse-library CI on main: the burn-kda designed-reds are now GREEN on the merged tree (this wave), burn-spectral’s known-red ×3 remains (its own lane).

Notes on the merged content itself

• burn-mhc block.rs:27 carries pub(crate) const B_RES_INIT: f32 = 10.0; // FALSIFY — but it multiplies Tensor::<2>::zeros, so the init is 0 at ANY value of the constant: the multiplier is dead arithmetic wearing a mutation marker (a falsify mutant reverting it to 10 cannot flip the init, and no gate can see that mutant). Functional fix is correct (b_res == 0 is pinned by a burn-mhc test); the dead multiplier should be dropped or actually wired the next time someone touches the init — flagged, not changed, it is the mhc lane’s marker.
• The dirty files found in the main checkout (fusion-flip WIP: Cargo.lock, core/Cargo.toml burn-situ cuda feature, the research/reviews/graph-trainer-2026-10-01.md continuation, and the burn-fused facade wgpu feature) were preserved exactly (backup in /tmp/opencode/dirty-backup/, stash bookkeeping fully unwound). The facade wgpu feature travelled through the byteflow2 merge already (both entries present, builds clean).
• No push: 17 commits ahead of origin/main locally, awaiting the owner’s OK (§1.6).

Queued next (per the lane’s order)

• (b) byte-vs-byteflow A/B — 3 seeds, 2k steps, pure CE, ONE batch size (§1.2/§2.6); costs are the A/B-protocol’s unknown-per-arm problem, unchanged.
• (c) next wave of idle branches (the inventory shows ~40 more wt/*; the digest’s priority order governs).
• CUDA queue: burn-attnres gain fixtures on CUDA are covered only through the smoke today; a targeted burn-attnres cuda gate rides with the attnres A/B wave.