Publish DeepSeek V4 Flash 0731 MMBT campaign

[Lightheartdevs]

What this publishes

This PR publishes the complete verified DeepSeek V4 Flash 0731 MMBT campaign:

• the optimized Tower2 deployment recipe for 2x RTX PRO 6000 (1M served context, FP8 KV cache, 500 W/GPU);
• canonical microbench results and corrected grader overlay;
• extended single-PR, finance, board-material, and frozen 75-PR results;
• compact audit JSON, hashes, completion audit, and verified-results synthesis;
• reproducible runners, telemetry, validators, and harness/grader hardening;
• root README, scorecard, claims registry, limitations, and findings-index integration.

Headline results

• Canonical microbench: 23/36 raw; 35/36 after audited grader correction across N=3. The remaining miss is a genuine word-limit failure.
• Single-PR audit: 3/3 complete artifacts, 2/3 expected verdicts, with one over-strict REVISE.
• Wall Street finance tasks: 0/2 substantive success.
• Board-material tasks: shipped outputs, but material deck-quality defects.
• Frozen 75-PR audit: 0/3 strict success. Two runs were scaffold-heavy/incomplete; one became an 815,279-token runaway. Earlier artificially capped v2/v3 attempts are preserved but excluded from valid outcomes.
• Production/performance validation: 1,048,576-token context, about 7,255 tok/s prefill, 272 tok/s single-stream decode, and 1,373 tok/s aggregate at concurrency 8, with restart/recall plus Sanctuary and Pixel validation.

Interpretation guardrails

• DeepSeek used its documented model-specific sampling, context, and FP8 KV operating point; this is the best validated configuration for this hardware, not a uniform leaderboard setting.
• DeepSeek canonical N=3 is less statistically mature than the older Qwen N=10 entries.
• Corrected scores use the hardened grader and retain the raw score for auditability.
• Cloud-model comparisons are categorical where deployment and telemetry are not comparable.
• Full raw run archives are intentionally externalized under MMBT repository-space policy (the V5 archive alone is 137 MB). Compact audits, manifests, hashes, and verified summaries are committed.

Validation

• Publication integrity validator: 10 JSON files parsed, 34 unique claims, 3 valid 75-PR outcomes.
• All 17 changed Python files compile.
• Targeted regression suite: 4 passed.
• Full PR diff passes git diff --check.
• No large GitHub blobs; largest blob in the branch is about 7.7 MB.