Add revisable diffusion decoding experiment

[tutao0123]

Summary

• add default-off low-confidence MDLM remasking with late decay and protected structural tokens
• add a resumable two-stage evaluation workflow and bounded public evidence
• document the relationship to LLaDA2.2 while keeping the repository focused on the controlled AR–diffusion comparison

Result

The 100-prompt gate selected 10% remasking with EOT protection and final-quarter decay. On the formal 2,000-continuation evaluation, the candidate reduced repeated 4-grams from 0.0951 to 0.0672 and raised distinct-3 from 0.832 to 0.867. However, blinded preference became worse:

• remask vs original MDLM: Flash utility 0.3565 over 2,000 pairs; Pro utility 0.300 over 100 audited pairs
• remask vs AR: Flash utility 0.0363 over 2,000 pairs; Pro utility 0.060 over 100 audited pairs
• batch-1 latency: 0.410 s for remask MDLM vs 0.386 s for original MDLM and 0.678 s for cached AR

This is recorded as a negative result. remask_fraction=0 remains the public default.

Scope

This is a sampler-only experiment, not a LLaDA2.2 reproduction. It does not implement Block Diffusion, KEEP/SUBSTITUTE/DELETE/INSERT, LCS-derived labels, MoE routing, or L-EBPO.

Validation

• ruff check .
• pytest (38 passed)
• bash -n scripts/*.sh
• Markdown local-link checks included in pytest
• 18/18 performance rows, 2,000 remask generations, 4,000 Flash judgments, and 200 Pro judgments validated against the bounded summaries
• all four A40 GPUs released after the run

Raw generations and API responses remain under out/ and are not included in Git.