feat(cache): rolling message-prefix cache breakpoint + hit-rate telem…

[wusijian007]

…etry (M3.1a/b)

First v3 milestone (§2 Cache Aligning). Adds the third prompt-cache breakpoint and makes cache effectiveness measurable -- the prerequisite for tuning §1 compaction without flying blind.

M3.1a -- message-prefix cache breakpoint:

• ModelRequest gains cacheConversation?: boolean. The agent loop (query.ts -> collectModelTurnWithRetry) sets it true on every stream request; the plain chat path (which doesn't go through query()) leaves it off.
• toAnthropicMessages(messages, { cacheLastMessage }) attaches cache_control: ephemeral to the LAST block of the LAST message, normalizing a trailing string body into a single text block first (cache_control can't ride a bare string). Both adapter call sites (create + stream) thread request.cacheConversation through.
• Result: 3 of Anthropic's 4 breakpoints in use (system + tools + conversation prefix). Each turn is a strict extension of the prior request, so the prefix hits the incremental cache. The breakpoint rolls forward to the new last message every turn.

M3.1b -- hit-rate telemetry:

• myagent usage <id> adds a cache hit ratio line: cache_read / (cache_read + input), with the raw fraction shown.
• myagent eval run totals line + REPORT.md gain the same ratio (on the deterministic fixture it's a stable 43.6%).
• Pure render layer; the tokens were already collected in M1.5a.

M3.1c (main-loop fork-trace attribution -- "why did this turn miss") is deferred to when §1 compaction lands and actually needs the alarm, per the roadmap.

Tests: prompt-caching.test.ts gains 5 cases pinning breakpoint placement (default off / last-block-only / block-form last message / empty list); query.test.ts asserts cacheConversation reaches every stream request; cli.test.ts asserts the usage hit-ratio line. Roadmap

• CLAUDE.md updated.

Local: 192 tests, 3/3 green.