[Bugfix][SM70] Restore DFlash2 prefill closure

[yangzhuxinyzx]

Public migration note (2026-08-27): this Draft is the public continuation of private PR #22, mirrored at exact head 702722b after public main resumed through #342. The original implementation notes and evidence are preserved below.

Purpose

Restore the accepted SM70 D256 prefill operators when a source overlay shadows
the bundled FA2 extension, then close the remaining NVFP4 large-M prefill gap
without changing DFlash2 M<=8 verification. This is stacked on private
quality-audit PR #18.

Test Plan

• Exercise bundled, missing, partially loaded, and explicit-sidecar loader paths.
• Require real dense and paged D256 operators before retained benchmarks.
• Run same-card cold 32K/64K DFlash2 prefill pairs with prefix cache reset.
• Separate FP32-SSM Q3296/Q6592 from FP16-SSM Q7920 geometry.
• Microbenchmark TurboMind versus bounded QPN4 large-M NVFP4 projections.
• Gate the QPN2-packed prefill candidate on layout/numerical parity, 32K/64K
  speed, unchanged M<=8 validation cost, PPL, and scored coding quality.

Test Result

• Focused D256 loader/status tests: 6 passed; partially loaded runtime
  capability regression: 2 passed.
• QPN2/QPN4 policy and dispatch tests: 3 passed.
• Ruff check/format, Python compile, shell syntax, and git diff check: passed.
• Missing-extension cold baseline: 32K 10.485347 s (3125.12 tok/s), 64K
  25.317796 s (2588.54 tok/s).
• Stable D256 sidecar with FP32 SSM/Q3296: 32K 9.425490 s (3476.53 tok/s,
  +11.24%), 64K 21.120102 s (3103.02 tok/s, +19.87%).
• FP32 SSM/Q6592 is neutral: 3481.83/3077.03 tok/s at 32K/64K.
• FP16 SSM/Q7920 is rejected on speed: 3271.49/2827.03 tok/s.
• Every measured request above retains the same first-token hash.
• Single-V100 M3296 projection microbenchmark:
  • fused gate/up: 5.351 -> 3.896 ms (-27.19%);
  • down: 2.646 -> 1.866 ms (-29.47%).
• The staged large-M candidate reuses already resident QPN2 codes/scales plus
  one shared 85-MiB FP16 workspace. It is default-off; M<=8 remains QPN2.
• GPU layout, full-model, validation-cost, and quality gates are pending behind
  active GPU owners. No pending candidate is enabled by default.

[yangzhuxinyzx]

Projected-main source audit (current main plus repaired #346) is complete for head 702722bab51731812e561011cadc139f8b8148d1:

• The [Bugfix][SM70] Restore DFlash2 prefill closure #350-only 9-file delta merges without conflict over the repaired [Bugfix][SM70] Audit DFlash2 output quality #346 tree.
• Reviewed the explicit D256 sidecar loader and capability validation. It only loads a user-selected library, verifies the real dense/paged operators, catches import/load/runtime failures, and retains the existing fallback otherwise.
• Reviewed the QPN2-packed large-M route. It reuses the existing per-device bounded FP16 workspace and zero-copy QPN2 code/scale views; M<=8 behavior is unchanged. Because its full-model quality/layout gates are still pending, the new large-M route correctly remains explicit opt-in, while the measured D256 sidecar recovery remains available.
• git diff --check, Python compile, benchmark CLI parsing, and all changed-file pre-commit hooks passed.
• Focused CPU/source-contract tests passed: benchmark loader status 3/3, Flash-V100 routing/policy 113/113, QPN2 preparation/dispatch 3/3 (119 total).
• Recorded V100 D256 evidence is positive with identical first-token hashes: 32K +11.24%, 64K +19.87%. Rejected FP16/Q7920 and neutral Q6592 variants remain rejected/not selected.

No source blocker found. After #346 reaches main I will retarget this PR, recompute the exact merge tree, and merge if the head remains locked.