[Bugfix][SM70] Audit DFlash2 output quality

[yangzhuxinyzx]

Public migration note (2026-08-27): this Draft is the public continuation of private PR #18, mirrored at exact head f5a93e2 after public main resumed through #342. The original implementation notes and evidence are preserved below.

Purpose

• Keep stale selector-QPN8 deployment overrides on the scored dense-order contract unless a separate research-only opt-in is present.
• Backport the minimal structured-output/spec-decode state fixes needed for DFlash2 reasoning, JSON schema, and tool calls.
• Add reproducible multi-seed quality evidence without replacing Eagle, MTP, DDTree, or the retained DFlash2 acceleration stack.

Test Plan

• Focused DFlash2, structured-output, XGrammar, comparator, and benchmark tests.
• TP4 V100 Graph API tests at B1/B4 with reasoning, tools, JSON schema, prefix cache, FP8 KV, and current Flash-V100.
• Three-seed 16K natural-EOS MBPP, HumanEval, and LiveCodeBench scoring.
• Long repeated-prefix state isolation, same-seed temperature 0.6 coding follow-up, and current-source target-only/DFlash2 Graph PPL pair.

Test Result

• Pre-commit: passed.
• CPU-focused gate: 106 passed, 12 CUDA-only skipped; separate V100 compact-rejection temperature 0.6/1.0 equivalence: 8 passed.
• Structured API: B1/B4 24/24; long alternating-prefix state 5/5; 32,960 prefix-cache hit tokens; no FSM, HTTP 500, traceback, or engine-error signature.
• MBPP three-seed EvalPlus: Base 89/93, Plus 80/93.
• HumanEval three-seed EvalPlus: Base 94/96, Plus 92/96.
• LiveCodeBench three-seed: 33/48, exactly 11/16 per seed.
• Same-seed historical no-DFlash comparison: MBPP+HumanEval both Base 62/63 and Plus 59/63; LiveCodeBench both 11/16.
• Temperature 0.6 weak-seed follow-up: MBPP Base/Plus 29/27 versus 27/24 at temperature 1.0; HumanEval 31/31 unchanged; LiveCodeBench 11/16 unchanged. It is an optional precise-coding profile because natural stops move 72/80 to 70/80.
• Current 256K/4096/prefix/Mamba-align Graph PPL: target-only 5.4993116, DFlash2 5.4993622 across 16,376 WikiText tokens; maximum segment difference 0.0062143.
• Quality-safe D2 dense-order selector keeps the bounded aggregate score and improves same-card mean steady decode 3.88% versus D1.

Full evidence and decisions are recorded in docs/design/sm70_dflash2_quality_audit.md.

This PR remains private and Draft; no public branch was updated.

[yangzhuxinyzx]

Audit result on latest public main c29cf3410b47ab14fefabbe69447cc536fd3402e and locked repaired head 5adbb4a5eecf90a67aa9f9005332348d56f26ec7:

• Reviewed all runtime deltas: dense-order selector guard remains generic and keeps the measured +3.88% quality-safe route; candidate ordering requires an explicit research opt-in.
• Reviewed structured/spec-decode transitions, reasoning-boundary trimming, XGrammar termination/reset behavior, and Qwen3Coder XML/JSON streaming parsing. The changes preserve the grammar boundary and add focused regression coverage for partial draft rejection, terminal tokens, literal parameter tags, alternate JSON calls, and final-brace emission.
• DCO is present on every implementation and repair commit.
• Exact merge-tree replay onto current main is conflict-free. Its result tree ce50409a5c9f7237a492109086ed9cb071f3d53f exactly matches the isolated audit tree.
• git diff --check, Python compile, all changed-file pre-commit hooks, and the three new benchmark CLI parse checks passed locally.
• Found and repaired the two forbidden stdlib re imports through [CI][SM70] Fix #346 benchmark regex imports #357; [CI][SM70] Fix #346 benchmark regex imports #357 local and remote pre-commit both passed.
• Recorded V100 evidence remains acceptable: structured API 24/24, focused CPU 106 passed plus 12 CUDA skips, compact rejection 8 passed, matched multi-seed coding parity, 256K target/DFlash PPL 5.4993116/5.4993622, and +3.88% steady decode for dense ordering.

Decision: source audit passes after #357. I will merge once the latest-head remote pre-commit completes.