config: adopt benchmark-qualified Gemini default

[tom-sapletta-com]

Summary

• adopt benchmark-qualified google/gemini-3.1-pro-preview as the todo2code runtime fallback
• preserve all global and stage-specific environment overrides
• add focused regression coverage and retain deterministic/offline boundaries

Validation

• focused model tests — PASS 2/2
• clean-base npm run verify — PASS (337 tests, 336 passed, 1 JDK-dependent skip)
• GitHub verify/Docker smoke — PASS
• Java/JDK 17 fixture — PASS
• authorized OpenRouter live contract — PASS 6/6, no fallback/degradation, 219.602 s, 137,229 tokens, $0.578808

Live model evidence

Every semantic stage reported google/gemini-3.1-pro-preview with effectiveMode=llm: NL, Markdown, documentation, communication, task synthesis and summary. Task synthesis plus summary account for 109,081 tokens, identifying context compaction as the next optimization target.

Merge ordering

PR #4 should land first so PR #3 receives the benchmark-qualified Gemini Koru workflow. The automatically triggered stale-DeepSeek Koru run for this evidence-only push was canceled before review execution; the same source diff already has attested Gemini review evidence in dispatch 30714664770.

Remaining blockers

• inherited ticket-019 governance conflicts
• independent protected review/merge evidence
• PR ci: adopt benchmark-qualified Koru judge #4 Koru workflow ordering