Document verified HRX MLA decode accuracy

[Copilot]

The reported GLM-4.7-Flash MLA decode degradation past 512 KV cells is not reproducible on the current pin; masked-V nondeterminism is addressed by the pinned flash-attention fix.

• Current measurements: Document HRX (-ub 1) and CPU perplexity at 4096 and 8192 tokens, including the CPU reference caveats.
• Scope: Link the pinned MLA strided-V and masked-V fixes; note that the historical A/B and op-level drift bisect remain unverified.

• Fixes HRX MLA decode path loses accuracy past ~512 KV cells (GLM-4.7-Flash, deepseek2) #310

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 310 - Partially compliant Compliant requirements: Document that the reported GLM-4.7-Flash MLA decode degradation is not reproducible on the current pin Link the pinned MLA strided-V and masked-V fixes Include current measurements of HRX and CPU perplexity at 4096 and 8192 tokens Note that the historical A/B and op-level drift bisect remain unverified Address the masked-V nondeterminism issue Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Clarity The documentation states that the historical 11.79 and 10.64 results do not have a runnable old-path A/B build, so the op-level drift bisect remains unverified. This is a significant limitation in the documentation as it leaves the root cause of the issue unresolved and unverified. The documentation should either provide a clear explanation of why the bisect cannot be performed or suggest a way to perform it in the future. invalid for this model. The historical 11.79 and 10.64 results do not have a runnable old-path A/B build, so the op-level drift bisect remains unverified. The masked-V kernel