feat(ai): prompt-cache correctness fixes for Kimi, OpenRouter, and Bedrock

[code-yeongyu]

Summary

This PR implements PR-A from the approved cache-hit maximization plan:

• Approved plan: cache-hit-maximization.md
• Scope: T1, T5, T14, T2, T3, T4, and T15 only.

Behavior changes

• Parse Kimi's flat usage.cached_tokens as cache-read usage without changing existing precedence.
• Use one shared Bedrock capability helper so only Claude Opus/Sonnet/Haiku 4.5 receive a one-hour cache TTL; resolver estimates now match the wire TTL.
• Report a 300-second prompt-cache TTL for the actual Claude SDK OAuth model shape.
• Add supportsPromptCacheKey compatibility detection and send stable, clamped prompt_cache_key values for Moonshot/Kimi requests.
• Default OpenRouter session affinity and send the session ID in both x-session-id and body session_id.
• Apply explicit OpenRouter cache markers only to anthropic/, qwen/, and google/ model prefixes after stripping one optional leading ~; runtime and generator logic are identical.
• Regenerate and validate the Moonshot/OpenRouter catalog metadata and add package changelog/source change notes.

TDD evidence

Each implementation todo has captured failing and passing output under:

• .omo/evidence/cache-hit-maximization/task-1-{red,green}.txt
• .omo/evidence/cache-hit-maximization/task-5-{red,green}.txt
• .omo/evidence/cache-hit-maximization/task-14-{red,green}.txt
• .omo/evidence/cache-hit-maximization/task-2-{red,green}.txt
• .omo/evidence/cache-hit-maximization/task-3-{red,green}.txt
• .omo/evidence/cache-hit-maximization/task-4-{red,green}.txt
• .omo/evidence/cache-hit-maximization/task-15-{red,green}.txt
• T4 hydration receipts: .omo/evidence/cache-hit-maximization/task-4-hydrate.txt and task-4-model-data-check.txt
• T15 prose receipt: .omo/evidence/cache-hit-maximization/task-15-changes.md.txt

These evidence files are task-local and intentionally ignored by git.

Verification

• cd packages/ai && npm test -- test/ - 191 files passed, 1816 tests passed, 25 files / 806 tests skipped by existing suite conditions.
• npm run check - passed.
• npm run build - passed (required to materialize workspace package artifacts before the full AI suite).
• npm run check:model-data - passed.
• node scripts/check-pr-changelog.mjs --base origin/main - passed.
• git diff --check origin/main...HEAD - passed.
• Runtime and generator OpenRouter cache-control blocks were compared byte-for-byte and are identical.

---

Summary by cubic

Improves prompt cache correctness across OpenAI-compatible and Bedrock lanes to boost cache hits and accurate usage/TTL reporting. Aligns Kimi usage accounting, OpenRouter cache-control and affinity, and Bedrock TTL behavior.

• New Features

  • Added supportsPromptCacheKey; sends a clamped prompt_cache_key for Moonshot/Kimi when caching is on.
  • Defaulted OpenRouter session affinity on; sends the same session ID in x-session-id and body session_id.
  • Applied explicit OpenRouter cache markers to anthropic/, qwen/, and google/ prefixes (handles one leading ~).
  • Regenerated and validated Moonshot/OpenRouter catalog metadata.

• Bug Fixes

  • Parsed Kimi’s flat usage.cached_tokens as cache-read tokens without changing precedence.
  • Restricted Bedrock 1h prompt-cache TTL to Claude Opus/Sonnet/Haiku 4.5 only; others use 5m so resolver matches the wire TTL.
  • Reported a 300s prompt-cache TTL for the actual claude-sdk-oauth model shape.

^{Written for commit 1014c26. Summary will update on new commits.}