fix(ai): select OpenRouter auto cache markers by model - #52119

Summary

• auto is selected by the OpenRouter route using model prefixes: anthropic/ places tool/system/tail breakpoints; qwen/ places system/tail content breakpoints (Alibaba ignores tool-definition markers); all other prefixes add none and retain implicit caching.
• The shared cache-policy pass just places the selected hints. Explicit cache policies and per-part hints always bypass the route default; prompt_cache_key still flows normally.
• One focused wire-body test table covers the model families, and one test verifies explicit overrides.

Why

Live OpenRouter streamed usage showed multi-turn Qwen Plus saving ~36% with explicit markers and a stable prompt_cache_key; the previous auto tail marker made Gemini 2.5 Flash cost more than its unmarked implicit caching. A Qwen snapshot that does not support explicit caching accepted the markers and reported the same billed cost as the unmarked request in a compatibility probe. This replaces the blanket OpenRouter enablement in #52110; the two PRs overlap and should not be merged independently.

Verification

• bun test test/provider/openrouter.test.ts test/cache-policy.test.ts in packages/ai: 47 pass.
• bun typecheck in packages/ai: pass.
• Prettier and diff whitespace check: pass.
• Full bun run check: lint passes; monorepo typecheck stops in unrelated @opencode/posts because this host's Node 18.19.1 is below Astro's required >=22.12.0.

Requested by: @rekram1-node (Aiden via Slack)