feat: conversion-driven extensions — resolvers, DefineTool, hooks, ops controls

Phase 9a (ADR-0014): Registry.RegisterResolver for dynamic tiers;
DefineTool[Args] typed tools; Usage cache/reasoning detail fields wired
through anthropic/openai/google; WithPromptCaching (Anthropic
cache_control); agent supervision hooks (WithMaxStepsFunc, WithSteer,
WithCompactor, WithToolErrorLimits + ErrToolLoop); health
Bench/Unbench/Snapshot; ChainConfig.Observer failover events.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

llm/response.go

@@ -18,10 +18,22 @@ const (
 	FinishOther FinishReason = "other"
 )
 
-// Usage reports token accounting for one request.
+// Usage reports token accounting for one request. InputTokens and
+// OutputTokens are always totals; the detail fields break out portions of
+// those totals where the provider reports them (0 = not reported).
 type Usage struct {
 	InputTokens  int
 	OutputTokens int
+
+	// CacheReadTokens is the portion of InputTokens served from the
+	// provider's prompt cache.
+	CacheReadTokens int
+	// CacheWriteTokens is the portion of InputTokens written to the
+	// provider's prompt cache.
+	CacheWriteTokens int
+	// ReasoningTokens is the portion of OutputTokens spent on
+	// thinking/reasoning.
+	ReasoningTokens int
 }
 
 // Total returns input plus output tokens.
@@ -31,6 +43,9 @@ func (u Usage) Total() int { return u.InputTokens + u.OutputTokens }
 func (u *Usage) Add(o Usage) {
 	u.InputTokens += o.InputTokens
 	u.OutputTokens += o.OutputTokens
+	u.CacheReadTokens += o.CacheReadTokens
+	u.CacheWriteTokens += o.CacheWriteTokens
+	u.ReasoningTokens += o.ReasoningTokens
 }
 
 // Response is the canonical generation result.