feat(reflect): step-by-step context caching for the Gemini tool loop

[nicoloboschi]

Summary

Reflect's LLM cost is dominated by the tool-result context re-sent on every turn, not the static system prefix. Caching only the fixed system+tools prefix (Gemini CachedContent) recovered ~9% of input on a short reflect; Gemini implicit caching gave 0% on short loops (and is unreliable in general).

This PR rolls a step-by-step context cache forward through the agent's tool loop: after each turn the CachedContent is extended to cover that turn's full input, so the next auto turn reuses the entire prior conversation at the cached-input rate and sends only its own new tool results as the delta. Each retrieved payload is billed at full price exactly once, then cached thereafter.

Measured (gemini-2.5-flash-lite)

Loop	Implicit only	Old static prefix	This PR
Short (3 calls)	0%	9%	~29%
Deep (8-11 calls)	~70% (unreliable)	~10%	~74-81%

On deep loops the per-call ratio ramps to ~99% on the deepest turns (each turn reuses nearly the whole prior context). Unlike implicit caching, this is deterministic and works on short loops too.

How it works

• GeminiCacheManager: caches contents (the growing conversation), with session-tracked create/delete for ephemeral per-reflect caches.
• call_with_tools gains cached_prefix_message_count so only the un-cached tail is sent. Forced (tool_config) retrieval turns stay uncached — Gemini rejects cached_content + tool_config — but the cache still advances underneath them, so the first auto turn inherits a cache covering the whole forced phase.
• Latency hidden: the next-turn cache create is scheduled to run concurrently with tool execution and only awaited right before the next auto call.
• Ephemeral: the reflect wrapper deletes the whole cache session on every exit path (answer / error / cancellation); a short TTL is the backstop. Verified no cache leak.
• Flag: HINDSIGHT_API_REFLECT_PROMPT_CACHE_ENABLED (default true) disables reflect caching independently of the global prompt cache.

Tests

• Deterministic step-by-step contract test (each auto turn reuses the previous turn's full input; caches torn down once) — no real LLM.
• GeminiCacheManager unit tests (create/track/delete/soft-fallback/error-swallow), no network.
• Real-Gemini cached-ratio measurement test updated with a real floor (gated behind HINDSIGHT_RUN_GEMINI_EVALS).

Not an LLM-behaviour change (no prompt/classification changes), so no judge test needed.