feat(anthropic): prompt caching via inline cache_control markers

[hb-cam]

What

Adds Anthropic prompt caching to AnthropicLLM via inline cache_control markers — the "inline-marker provider" strategy that LLMInterface.get_or_create_cached_prefix's docstring already describes for Anthropic. No engine changes, no new config.

Today the provider sends no cache_control at all, so every call pays full input price on content the engine resends verbatim:

• fact extraction reuses the same system prompt across every chunk;
• the reflect agent loop resends the entire growing conversation on each of its iterations (up to HINDSIGHT_API_REFLECT_MAX_ITERATIONS), re-paying the full prefix each time.

Anthropic cache reads bill at ~10% of the base input price, so on reflect-heavy workloads (mental-model refresh) this is a large input-cost reduction with no latency penalty.

How

Two breakpoints (of the four Anthropic allows):

1. System prompt, both entry points — rendered as a block list with cache_control: {type: "ephemeral"}. Caching is a prefix match, so this caches tools + system together. Schema text-injection (the non-strict response_format path) happens before marking and lands inside the cached block.
2. Final message content block, call_with_tools() only — each agent-loop request's end-marker becomes the next iteration's cache read point. String content converts to a text block (skipped for empty strings, which the API rejects); block-list content (tool results) gets the marker on its last block.

Marking is safe unconditionally: below the model's minimum cacheable prefix the marker is silently ignored — no write premium, no error. cache_read_input_tokens already flows through _usage_from_anthropic_response into metrics/tracing, so cache effectiveness is observable immediately via cached_tokens.

Testing

• 5 new tests in tests/test_anthropic_prompt_caching.py (written first, watched fail): marker placement on both entry points, schema-injection landing inside the cached block, tool-result block-list handling, and the no-system passthrough.
• One existing assertion updated for the representation change: test_non_strict_keeps_text_injection_fallback checked a substring on system as a plain string; the schema-in-prompt behavior itself is unchanged and still covered.
• test_anthropic_structured_tool_use.py, test_llm_extra_body.py, and test_mental_model_delta.py pass; ruff and ty clean.

🤖 Generated with Claude Code