Route routine AI through free Gemini tier

[oosuhada]

Goal

Move routine PM and ordinary technical-comment interpretation off paid Vertex without weakening formal review/reasoning paths. No event batching is added.

Routing

• routine PM + ordinary GitHub technical-comment PM: Gemini Developer API Free Tier gemini-3.5-flash-lite first
• free API failure: Vertex Flash-Lite fallback only while the paid automatic budget remains available
• formal pull_request_review, chat and commit reasoning: Vertex gemini-3.7-flash
• project-memory refresh obeys the same paid fallback guard, so an automatic docs refresh cannot silently bypass the paid cap
• paid and free automatic budgets are tracked separately

Validation

• backend pytest: 19 passed
• Python compile smoke
• git diff --check
• production free API smoke: provider=gemini-developer-free, billingClass=free, model=gemini-3.5-flash-lite
• live production after rollout: paid counters stayed at 26 calls / 1,264,963 input while 2 automatic PM calls / 82,942 input were recorded separately as free-tier usage
• repeated service restart with unchanged semantic state did not reopen paid usage

Production

Deployed to the Mac mini service. /api/health reports simpleProvider=gemini-developer-free, freeTierEnabled=true, and separate paid/free budgets. The paid automatic budget remains capped while routine free-tier PM can continue.