[autotuner][cute] repair partial FFI candidates away from the seed, not onto it

[calebmkim]

Stacked PRs:

• [autotuner][cute] add formula-based tcgen05 matmul seed heuristic #3152
• [autotuner][cute] rename ab_stages "three" symbols that are no longer 3-only #3151
• [autotuner][cute] widen the tcgen05 matmul search space: block_n=128 cluster_m=2 + ab=4 #3150
• ->[autotuner][cute] repair partial FFI candidates away from the seed, not onto it #3167
• [autotuner][cute] demote tcgen05_tvm_ffi_launch to a seed-only knob #3149
• cute: remove redundant subprocess-autotune DSL recompiles #3148

---

[autotuner][cute] repair partial FFI candidates away from the seed, not onto it

The parent commit made the search draw tcgen05_tvm_ffi_launch False-only, but that
flag is just ONE of the six clauses in target1_tvm_ffi_promotion_requested. The
other five -- tcgen05_flat_role_coordinates, tcgen05_layout_strategy='explicit_epi_tile',
and the three tcgen05_layout_overrides{epi_tile_m,epi_tile_n,d_store_box_n} -- are
independent search draws, so ~96% of candidates still tripped the gate and were
overwritten wholesale with the FFI seed (measured, S3 2048x4096x4096 bf16: 293/300
candidates -> 1 distinct config, 0 of 30 keys varying). That is not a search.

The projection was never needed for FFI coverage: the envelope is already in the
population as an eligibility-gated seed (CuteTcgen05ClusterM2FfiHeuristic plus the
formula FFI-alt), which _generate_best_available_population_flat additionally records
as a pinned finalist, so it is benchmarked and protected from pruning regardless.
Verified on 4 FFI-eligible shapes with the projection bypassed: the seed still reaches
the population every time, canonical-hash identical to the normalized seed.

_fix_target1_tvm_ffi_search_config becomes a three-way gate:

1. A config already carrying the COMPLETE envelope (_is_coherent_target1_tvm_ffi_config
   over the 9 keys in _TARGET1_TVM_FFI_ENVELOPE_KEYS) is left untouched -- it is a
   valid direct-entry kernel, and projecting would clobber out-of-envelope knobs it
   legitimately chose. Not reachable from a random draw (0/2000 pure draws, since
   tvm_ffi_launch is drawn False-only and is an envelope key); this branch exists for
   the seed and for user-supplied / cache-restored configs.
2. An EXPLICIT but partial tvm_ffi_launch=True (user config or cache entry; the search
   never draws True) is still projected onto the validated envelope -- unchanged
   behavior. An explicit request must not be silently downgraded, and a partial
   envelope cannot compile on its own.
3. Everything else -- a search candidate that tripped the gate via one of the five
   other keys -- is repaired AWAY from FFI via _strip_target1_tvm_ffi_promotion_surface,
   preserving the ~20 other keys it sampled. The four layout keys move as one unit
   because validate_tcgen05_strategy_invariants couples them.

fix_search_config now runs the FFI stage FIRST, ahead of fix_cluster_m2_search_config.
Its bn-snapping guard only fires when the layout is already DEFAULT, so under the old
order a repaired candidate had already been forced to bn=256 and could never reach the
un-seeded [256,128,] tile. Measured contribution, same seeded population: [256,128,]
71 -> 139 (S3) and 72 -> 140 (medM). The original position (added by #2518) predates any
such dependency -- the old projection wiped every tcgen05* key, so ordering was moot.

Effect (single B200, 300-draw populations, interleaved old/new control in one process):

shape distinct ffi=True cm2 [256,128,*]
S3 2048x4096x4096 bf16 18 -> 296 283 -> 1 7 -> 132
medM 512x4096x4096 bf16 18 -> 296 283 -> 1 8 -> 133
S3resid 2048x4096x4096 resid_add 19 -> 300 282 -> 1 9 -> 144
8192x1024x1024 bf16 11 -> 299 290 -> 1 4 -> 149

Correctness: 36 distinct post-repair configs compiled, run, and compared against
torch.matmul (atol=rtol=2e-2) -- 36/36 correct, zero wrong-number results. Repaired
candidates skip codegen slightly more often than naturally-non-FFI ones (20% vs ~13%,
all the same pre-existing non-whitelisted-fused-epilogue gate), which the autotuner
handles by skipping; net usable configs still rise by more than an order of magnitude.
Branch 1 verified directly: a coherent config with a mutated out-of-envelope knob
(l2_swizzle_size, l2_groupings, acc_stages, num_stages) keeps that knob and compiles
correct, while breaking one envelope key sends it back through branch 2.

Known adjacent issue, pre-existing and NOT addressed here: on residual_add shapes the
normalized FFI seed does not compile (flatten/unflatten rewrites ab_stages 3->2 and
c_stages 2->4 via _fix_aux_tma_full_tile_search_config, after which the flat-role
codegen guard rejects it). Confirmed identical on the unpatched parent.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com