Cache model responses and survive partial failures

[tabkram]

Why

src/llm.ts:106 is a single generateText call with no caching. Two consequences:

1. Regenerating an unchanged range costs full price. During template or config work you re-run the same range repeatedly and pay every time.
2. Partial failures may discard completed work. A multi-language or multi-release run makes several sequential model calls. If the last one fails, it needs confirming that the successful ones are not lost — that is real money and a bad CI experience.

Scope

Response cache

• Key on a hash of (changelog + context, resolved prompt, model, provider, language, temperature) — anything that changes the output changes the key.
• Store under the existing .ai-release-notes/ directory; add a cache directory to .gitignore.
• --no-cache to bypass, and a way to clear it.

Provider prompt caching

• The system prompt, scope guard, and instructions are identical across every call in a run. Mark them as cacheable where the provider supports it (Anthropic and OpenAI both do) — this is a straight discount on multi-language and multi-release runs.

Partial-failure tolerance

• Audit the multi-call paths in src/generator.ts and src/release-document.ts.
• A failure on call N must preserve calls 1..N-1, report exactly what succeeded, and let a re-run resume rather than restart.

Acceptance criteria

• Re-running an identical range makes zero model calls and reports a cache hit
• Any change to prompts, model, provider, or context misses the cache
• --no-cache bypasses; cache directory is gitignored and documented
• Prompt caching is enabled for providers that support it, and the token report reflects it
• A forced failure mid multi-language run preserves completed languages
• The run report distinguishes generated / cached / failed per output

Files

src/llm.ts, src/generator.ts, src/release-document.ts, src/config.ts, docs/configuration.md

Notes

The AI SDK retries transient errors by default, so this issue is about spend and resumability, not retry logic.