Implement prompt caching across LLM providers

[jrswab]

Problem

Axe re-transmits the full system prompt (skill + files + memory), tool definitions, and message history on every turn of a multi-turn agent run. No provider currently sends cache hints, so users pay full input-token costs for static content on every API call. For agents with large SKILL.md files and many tool calls, this is a significant cost multiplier.

Goal

Add provider-side prompt caching support so that repeated static context (system prompt, tool schemas) is cached by the LLM provider after the first turn, reducing input-token costs on subsequent turns. Cache usage must be observable in verbose mode and JSON output.

Provider Support

Provider	Mechanism	Request Change	Response Change
Anthropic	Explicit cache_control: {type: "ephemeral"}	Add cache hints to system blocks and tool defs	Parse cache_creation_input_tokens / cache_read_input_tokens
OpenAI	Automatic prefix caching	None (automatic)	Parse usage.prompt_tokens_details.cached_tokens
Bedrock (Claude)	cachePoint blocks	Add cache points to system/tool config	Parse cacheReadInputTokens / cacheWriteInputTokens
Gemini	Context caching API	Out of scope (requires stateful two-phase API)	Out of scope
Ollama	N/A	None	None
OpenCode	Pass-through	No gateway-level control assumed	No changes

Spec

See docs/plans/049_prompt_caching_spec.md for full requirements and implementation plan.

Acceptance Criteria

• Cache annotations added to provider.Request and cache metrics to provider.Response
• Anthropic provider sends cache_control on system prompt and parses cache usage
• OpenAI provider reports cached_tokens from API responses
• Bedrock provider sends cachePoint blocks and parses cache usage
• Runner automatically enables caching on every request without agent TOML changes
• Cache read/write tokens tracked cumulatively across conversation turns
• Cache token counts visible in verbose output and JSON result envelope
• Graceful degradation: unsupported providers/models continue to work unchanged
• Tests added for each provider's cache request construction and response parsing

[hiSandog]

The shared request model should represent cache boundaries semantically, then let each provider enforce its limits; a boolean “cache everything” cannot express minimum token thresholds or Bedrock's limited cache points. Cache metrics also need per-turn read/write counts and pricing semantics rather than only cumulative totals, because providers report them differently. Request-shape fixtures plus an opt-in two-turn live test should verify stable prefix placement, unsupported-model fallback, and that changing tools/system text invalidates the expected segment.