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>

chain.go

@@ -91,10 +91,17 @@ func chainDo[T any](ctx context.Context, c *chain, req llm.Request, attempt func
 	var zero T
 	var failures []error
 
+	observe := func(ev FailoverEvent) {
+		if c.cfg.Observer != nil {
+			c.cfg.Observer(ev)
+		}
+	}
+
 	for _, t := range c.targets {
 		if !c.tracker.Available(t.key) {
 			until := c.tracker.BackedOffUntil(t.key)
 			failures = append(failures, fmt.Errorf("%s: skipped (backed off until %s)", t.key, until.Format("15:04:05.000")))
+			observe(FailoverEvent{Target: t.key, Skipped: true})
 			continue
 		}
 
@@ -119,6 +126,7 @@ func chainDo[T any](ctx context.Context, c *chain, req llm.Request, attempt func
 
 			class := c.cfg.classify(err)
 			if class == llm.ClassPermanent {
+				observe(FailoverEvent{Target: t.key, Err: err, Class: class, Attempt: attemptN})
 				if errors.Is(err, llm.ErrModelNotFound) || errors.Is(err, llm.ErrUnsupported) || c.cfg.AdvanceOnPermanent {
 					// Not a health problem (or policy says keep going):
 					// advance without penalizing the target.
@@ -134,6 +142,7 @@ func chainDo[T any](ctx context.Context, c *chain, req llm.Request, attempt func
 			// attempts remain — but advance as soon as the tracker benches
 			// it (a freshly backed-off target is not worth more retries).
 			benched := c.tracker.ReportFailure(t.key)
+			observe(FailoverEvent{Target: t.key, Err: err, Class: class, Attempt: attemptN, Benched: benched})
 			if !benched && attemptN < retries {
 				continue
 			}