fix: anthropic tool-call round-trip, cache zero-hit parity, round-robin load balancing

anthropic.lua v3.0.0:
- Issue 1: tool_result/tool_use round-trip
- Issue 3: thinking default OFF (opt-in via extra_body.thinking)
- Issue 4: tool_choice mapping
- Issue 5: collect_blocks preserves unknown part types
- message_stop no longer emits done=true (was overwriting tool_calls finish_reason)
- cache_read_input_tokens normalized even at 0

gemini.lua:
- transform_response was missing cachedContentTokenCount

openai.lua (Issue 6):
- transform_error handles flat envelopes, nginx HTML, bare text

chat.go mergeUsage:
- Keep PromptTokensDetails even when CachedTokens=0

scheduler.go:
- Remove sort.SliceStable by Pref; round-robin cursor is the only LB mechanism

provider.go ModelAvailable:
- Also check Pref() > prefMin, persistently failing slots exit cands

presets.go:
- 17 built-in source templates

Tests: 6 new test functions, 2 updated for new semantics

internal/scheduler/scheduler.go

@@ -248,10 +248,9 @@ func (s *Scheduler) chainDrive(ctx context.Context, chain *Chain, req *types.Cha
 			ce.Skipped = append(ce.Skipped, fmt.Sprintf("tier %d: no schedulable slot (cooling or quota exhausted)", tn.Tier))
 			continue
 		}
-		// preference orders a same-tier run; stable so equal prefs keep order
-		sort.SliceStable(cands, func(i, j int) bool {
-			return cands[i].Prov.Pref(cands[i].Model) > cands[j].Prov.Pref(cands[j].Model)
-		})
+		// No Pref sort: load balancing is done by round-robin cursor.
+		// Persistently failing slots are excluded by ModelAvailable
+		// (which checks Pref > prefMin).
 		base := tn.NextStart()
 		res := runTier(ctx, tn, cands, base, req, stream)
 		if res.resp != nil || res.chunks != nil {
@@ -350,9 +349,7 @@ func (s *Scheduler) ChainImage(ctx context.Context, chain *Chain, req *types.Ima
 			ce.Skipped = append(ce.Skipped, fmt.Sprintf("tier %d: no schedulable slot (cooling)", tn.Tier))
 			continue
 		}
-		sort.SliceStable(cands, func(i, j int) bool {
-			return cands[i].Prov.Pref(cands[i].Model) > cands[j].Prov.Pref(cands[j].Model)
-		})
+		// No Pref sort: load balancing is done by round-robin cursor.
 		base := tn.NextStart()
 		var hard []TierError
 		for i := 0; i < len(cands); i++ {

internal/scheduler/scheduler_test.go

@@ -177,14 +177,21 @@ func TestChainPreferenceSinksButStaysReachable(t *testing.T) {
 	if err != nil {
 		t.Fatalf("chain: %v", err)
 	}
-	// higher pref (good) is tried first and hard-fails; the pass moves on to
-	// the negative-pref slot, which sinks but stays reachable: with the real
-	// provider its success would RecordSuccess (+1 pref, self-heal)
+	// Round-robin starts at index 0 (neg). neg is available (pref=-5 > -20)
+	// and succeeds. good is never tried because neg already won.
 	if src != "neg" || model != "n" || resp.Content != "neg" {
 		t.Fatalf("served src=%q model=%q content=%q", src, model, resp.Content)
 	}
-	if good.chatHits.Load() == 0 {
-		t.Fatal("higher-pref slot must be tried first")
+	if good.chatHits.Load() != 0 {
+		t.Fatal("neg was tried first and succeeded; good must not be attempted")
+	}
+	// Second request: cursor advances. neg wins again (good hard-fails).
+	resp2, src2, _, err2 := s.ChainChat(context.Background(), ch, chatReq(), nil)
+	if err2 != nil {
+		t.Fatalf("second chain: %v", err2)
+	}
+	if src2 != "neg" || resp2.Content != "neg" {
+		t.Fatalf("second req src=%q content=%q", src2, resp2.Content)
 	}
 }