[codex] Improve local Gemma summary prompts

[r3dbars]

Summary

• tighten the local Gemma meeting-summary prompt for more useful, concise output
• align the autoeval prompt with the shipped app prompt
• normalize Gemma title values that accidentally include markdown heading markers
• add prompt and title-normalizer regression coverage

Why

The previous summaries could pass structural checks while still sounding bland or meta, especially on test/demo transcripts. The new prompt pushes Gemma toward substance, explicit decisions, owner-oriented action items, and cleaner titles.

Validation

• Local Gemma baseline sample: YELLOW
• Local Gemma candidate 2 sample: GREEN
• Longer meeting direct sample: GREEN
• bash build-deps.sh --force
• bash build.sh --no-open
• bash run-tests.sh: 4519 passed
• bash run-integration-smoke.sh
• python3 -m py_compile scripts/ops/local-gemma-summary-autoeval.py

Notes

Raw transcript-derived prompts and outputs stayed local under ignored .autoeval/ paths.