fix(honcho): restore cache-stable system prompt for turn-varying recall

[erosika]

Summary

• Moves Honcho turn-varying recall out of the system prompt into a separate injected user message
• Keeps the system prompt prefix stable across turns for Anthropic/OpenRouter/OpenAI prompt caching
• Continuing turns still receive fresh Honcho recall via a <honcho-context> message positioned after system + prefill

Detail

When _honcho_turn_context is appended directly to effective_system, the outbound system message changes on every turn. All three major caching mechanisms (Anthropic prompt caching, OpenRouter sticky routing, OpenAI prefix matching) depend on an exact-match system prefix to avoid cache invalidation.

This change injects the turn-varying recall as a standalone user-role message instead, preserving cache-hit eligibility while still surfacing Honcho context to the model. First-turn behavior is unchanged (context is baked into the cached system prompt once at session start).

Test plan

• Verify prompt cache hit rates on multi-turn conversations (system prefix should be identical across turns)
• Verify Honcho recall still reaches the model on continuing turns (<honcho-context> visible in debug logs)
• Existing test suite passes

[erosika]

Closing in favor of #1201 which takes the better approach of appending recall to the current-turn user message rather than injecting a synthetic user message. Also handles multimodal content and includes regression tests.