perf(llm): cache stable Claude prompts

[rabbive]

Mark byte-stable planning and codegen prefixes for Anthropic prompt caching without changing prompt content.\n\nRefs #20

[rakshityadav1868]

Nice one @rabbive , verified locally-prefix stays byte-stable across specs and the tests pass. One small note: the cacheable minimum is model-dependent (1024 tokens on the default sonnet-4-6, but 4096 on some models like Haiku 4.5). The planning prefix is ~2K tokens, so if someone overrides AOD_CLAUDE_MODEL to one of those, caching silently no-ops. Harmless, but maybe worth a line in the comment.

[rabbive]

Addressed the review feedback in 170c0a2.

Changes made:

• Clarified that Anthropic’s cacheable minimum is model-dependent.
• Documented the default Sonnet model’s 1,024-token minimum.
• Noted that the roughly 2K-token planning prefix remains uncached on models requiring 4,096 tokens.
• Clarified that this only causes a cache miss—the request and generated output remain unaffected.

Thanks for catching this!