Publish Gemma 4 31B Q4 verified Tower2 results

[Lightheartdevs]

What changed

• Adds the complete Gemma 4 31B QAT Q4_0 Tower2 campaign: preregistration, pinned model/runtime/topology, canonical N=3 and N=10 scorecards, correction overlays, grader manifests, evidence audits, and extended-suite audits.
• Documents the selected two-GPU operating point: one independent four-slot llama.cpp replica per RTX PRO 6000, Q8 KV, native 262,144-token context per slot, and 500 W per-GPU caps.
• Publishes direct comparisons to Qwen3.6-27B, Qwen3-Coder-Next, Qwen3.5-397B-A17B Q3, Qwen3.6-35B-A3B, and DeepSeek V4 Flash 0731 without treating unlike operating points as a controlled leaderboard.
• Hardens the 75-PR and completion gates so missing artifacts fail closed and a qualifying completion tag must be annotated, point at HEAD, and belong to a clean candidate repository.
• Records the exact post-campaign restoration to the proven DeepSeek V4 Flash deployment, including fresh Sanctuary and Pixel tool-use checks with no fallback.

Key results

• Canonical N=3: 29/36 raw, 32/36 corrected.
• Canonical N=10: 89/120 raw, 99/120 corrected; 116/120 ordinary completions.
• The only correction is the project-management family (0/10 raw to 10/10 corrected), with every changed cell tied to immutable grader, report, archive, and correction-script hashes.
• Versus Qwen3.6-27B: Gemma leads the comparable N=3 bounded-quality cohort (29/36 raw versus 20/36), while short-context single-stream speed is effectively tied (70.3 versus 72.1 tok/s).
• Qwen3.6-27B/vLLM remains the high-concurrency serving winner (1,336.5 aggregate tok/s at C32 on one GPU versus Gemma's 290.3 aggregate tok/s at total C8 across two GPUs; operational comparison, not a controlled model-speed A/B).
• Extended suite: 0/12 strict substantive passes. The model often produced polished but incomplete finance/presentation work and failed every frozen 75-PR marathon attempt under strict audit.

Why

This supplies the missing Gemma 31B data needed to compare a strong dense Q4 model with the existing Qwen and DeepSeek Tower2 results. It also records the practical boundary between bounded task quality and unattended marathon reliability.

Validation

• python3 -m pytest -q tooling/test_*.py tooling/scripts/test_*.py tooling/deployments/gemma4-31b-q4-tower2/test_*.py — 62 passed.
• python3 tooling/validate_gemma4_comparison_sources.py — passed.
• python3 tooling/deployments/gemma4-31b-q4-tower2/validate_preregistration.py — valid.
• sha256sum -c benchmarks/gemma4-31b-q4/SHA256SUMS — all files OK.
• All published JSON and claims.yaml parse; Markdown local-link audit and git diff --cached --check pass.
• Post-restore: DeepSeek health 200, served model DeepSeek-V4-Flash-0731, 1,048,576-token context, 500 W caps, both portals healthy, and fresh Sanctuary/Pixel exec traces pass with fallbackUsed=false.