Fused rewrite: two rungs, then death

ADR-0003 measured fused and it lost (1.25x slower at flagship; the 1.7-2.0x smoke number was pre-fix and is retracted). The 2026-09-25 analysis (docs/archive/research/2026-09-25-fused-rewrite-plan.md) quantified three causes: the arms_inner_adjoint re-run per loop iteration (~1-1.5 s of the 2.16 s deficit), a full weight-grad D2H round trip every step (~30 blocking transactions, ~60 MB PCIe), and hand-rolled fp32 FFMA matmuls against burn’s autotuned tensor-core path. The “direct adjoints” already in kernels.rs relocate the round trip rather than remove it and stay unwired.

Decision: one rewrite window with two ranked rungs and a hard kill switch. Rung 1 — seed KDA/MSA/Engram adjoint kernels from the per-iteration buffers the forward already saves (only KDA’s recurrent state S needs adding to the workspace), delete the inner-adjoint re-run, one device sync per step, grad registration on device buffers, junk-kernel launches removed. Rung 2 — delete the hand matmuls, materialize + burn’s matmul_autotune, persistent workspace. Rung 1 must buy >= 1 s/step at flagship; Rung 2 must reach parity-or-better vs the fusion-on burn path (pre.4). Either rung failing its number deletes all ~7.4k LOC of fused/ with this ADR as the pre-justification. CUDA graph capture (available since pre.3) is the single bounded last resort and nothing more.

The fusion backend flag is part of the same window: it unlocks on pre.4 (#5673 extension-crate metadata callbacks) and is benched as its own arm — the bar fused must beat is fusion-on burn, not yesterday’s burn.