fix(ai): gemini default model is dead for new API keys

[shossain786]

Summary

• Live-checked ai.runSummary against a real Gemini key (the one thing feat(ai): opt-in AI run summary on the HTML dashboard #168 left unverified) and found the default model, gemini-2.5-flash, returns HTTP 404 — Google retired it for new API keys, pointing callers at gemini-3.8-flash instead.
• Any ai.provider: gemini config that doesn't set an explicit ai.model is broken out of the box today. Switched the default and the docs table to gemini-3.8-flash.

Verification

• curl directly against the Gemini endpoint: gemini-2.5-flash → 404 before, gemini-3.8-flash → 200 after.
• End-to-end through the actual framework code path (AiRunSummarizer.buildPrompt → AiFailureAnalyzer.callApi → response parsing) against the real API with a realistic two-failure run summary prompt — returned a correct plain-English summary.
• Note for anyone tuning ai.timeoutSeconds with Gemini: observed latency for this model was highly variable in testing (as low as ~5s, as high as ~95s for the same prompt), driven by the model's internal "thinking" tokens. The default ai.timeoutSeconds: 20 may be too tight for Gemini in practice — not changed here since it's a latency/cost tradeoff for the user to tune, not a correctness bug, but worth knowing.
• mvn test: 740/740 green.

Test plan

• mvn test passes
• Live API call confirms the old default 404s and the new default works
• Live end-to-end call through the actual AiRunSummarizer prompt-building and AiFailureAnalyzer HTTP/parsing code

🤖 Generated with Claude Code