perf(providers): cache conversation history on Anthropic requests

[kamushadenes]

Summary

Adds a rolling cache_control breakpoint on the last message in every Anthropic request. Without this, conversation history is sent uncached every turn — the dominant cost line on long agent sessions.

System prompt and the last tool definition were already cached; messages were not, contradicting the description in #1042. Anthropic permits up to 4 breakpoints, so we have room.

Refs #1042.

Evidence (28h on a deployment routed via provider_type=anthropic_native)

Agent	Fresh input	cache_read	cache_create	Real hit ratio	Calls
chloe	3.84M	2.86M	1.18M	36% (2.86/7.88)	110
bones	239k	586k	253k	71%	32

Chloe's 36% comes purely from system + tools. Each turn sends 21–57k tokens of conversation history at full input rate. With this PR, that history rolls forward as cache_read on every subsequent turn within the 5-min TTL.

Estimated impact for Chloe: ~$443/mo → ~$130–180/mo (60–70% reduction).

Changes

internal/providers/anthropic_request.go:

• New applyCacheControlToLastMessage() helper called once in buildRequestBody before the body map is assembled.
• Handles all three content shapes used in the function:
  • string (typical text-only user message): converted to a single text block to attach cache_control. Anthropic accepts both shapes.
  • []map[string]any (multi-modal user, tool_result, assistant text+tool_use): cache_control on the last block.
  • []json.RawMessage (assistant raw blocks preserving thinking signatures): re-marshals the last block; decode failures skip silently to avoid corrupting the request.

internal/providers/anthropic_message_cache_test.go:

• 6 unit tests covering all three content shapes, empty messages, tool_result blocks, and the invalid-JSON edge case for raw assistant content.

Trade-off considered

cache_create is 1.25x input on the new tail each turn, while cache_read is 0.10x on everything before. Net positive for any session with > 1 turn of reuse within the 5-minute TTL. For ephemeral one-shot delegations, the loss is bounded by the size of a single user message (a few hundred to a few thousand tokens) — well below the savings on the cached prefix.

A second rotating breakpoint (last 2 user-role messages) would extend cache survival across more concurrent contexts. Held off here to keep the change minimal; happy to add if reviewers prefer.

Test plan

• go test ./internal/providers/ — 562 passed
• go vet ./internal/providers/ — clean
• go build ./... — PG build passes
• go build -tags sqliteonly ./... — SQLite build passes
• 6 new unit tests covering string content, block array, tool_result, raw assistant blocks, invalid JSON, empty slice
• Production validation pending: monitor usage_snapshots.cache_read_tokens for 24–48h after deploy

Out of scope

This does not address the OpenAI-compat path issue described in #1042 (openai_config.go:99 hardcoded CacheControl: false). That still requires Option 1 or Option 2 from the issue body. The fix here only helps agents bound to a native Anthropic adapter (provider_type=anthropic or anthropic_native via OpenRouter).