feat: add cache_control breakpoints to compiler for Anthropic prompt caching

[pitimon]

Problem

OpenKB compiler reuses a "base context A" (system + document) across N+M+2 LLM calls per document (summary → concepts-plan → N create + M update concept pages). Without cache_control markers, every call re-bills the full document content as input tokens.

For Anthropic Sonnet 4.5 (via OpenRouter or direct), prompt caching can reduce input cost by ~90% on cached prefix and reduce TTFT. Minimum cacheable size is 1,024 tokens, easily exceeded by typical document content.

Proposal

Add cache_control: {"type": "ephemeral"} markers at two breakpoints in openkb/agent/compiler.py:

1. End of doc_msg in compile_short_doc + compile_long_doc — caches system + doc for all downstream calls (summary, plan, every concept).
2. End of assistant summary message in _compile_concepts (3 call sites: plan, create, update) — caches system + doc + summary for all concept generation calls.

Two breakpoints, well within Anthropic's max-4 limit.

Compatibility

• Anthropic / OpenRouter→Anthropic: cache_control honored.
• OpenAI: list-of-blocks content format is valid (Vision API uses it); cache_control silently ignored.
• Other providers: LiteLLM normalizes/strips unknown fields.

Side fix

_llm_call_async currently does not forward **kwargs while _llm_call does (asymmetry noted in memory #82886). Add **kwargs for parity.

Out of scope

• OpenRouter Response Caching (X-OpenRouter-Cache: true) — different mechanism, evaluated separately.
• Refactoring messages into a dedicated builder module — keep patch surgical.

Test plan

• Existing pytest suite passes (mocks accept *args, **kwargs).
• New assertion: completion payload contains cache_control block on the doc_msg.
• Manual smoke against a real Anthropic key: observe cached_tokens in prompt_tokens_details on calls 2..N.

References

• Memory observation S11144 (3-5 line patch feasibility, OpenKB compiler audit).
• CLAUDE.md compiler architecture: "Designed around prompt-cache reuse: a single base context A reused across summary → concept-plan → concept-page calls."