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/lua/adapters/gemini.lua

@@ -68,6 +68,15 @@ function adapter.transform_response(raw_body)
         unified.token_usage.prompt = resp.usageMetadata.promptTokenCount or 0
         unified.token_usage.completion = resp.usageMetadata.candidatesTokenCount or 0
         unified.token_usage.total = resp.usageMetadata.totalTokenCount or 0
+        -- Gemini reports context-cache reads as cachedContentTokenCount;
+        -- normalize into OpenAI-standard prompt_tokens_details.cached_tokens
+        -- so clients and the audit trail see the hit count. Emitted even when
+        -- 0 so a reported miss stays distinguishable from "not reported".
+        if resp.usageMetadata.cachedContentTokenCount ~= nil then
+            unified.token_usage.prompt_tokens_details = {
+                cached_tokens = resp.usageMetadata.cachedContentTokenCount
+            }
+        end
     end
 
     if resp.candidates and #resp.candidates > 0 then
@@ -100,9 +109,12 @@ function adapter.transform_stream_chunk(raw_chunk)
         local t = chunk.usageMetadata.totalTokenCount or 0
         if p > 0 or c > 0 or t > 0 then
             uses = { prompt = p, completion = c, total = t }
-            local cacheRead = chunk.usageMetadata.cachedContentTokenCount or 0
-            if cacheRead > 0 then
-                uses.prompt_tokens_details = { cached_tokens = cacheRead }
+            -- Emit details whenever the field is present, even at 0, so a
+            -- reported cache miss stays distinguishable from "not reported".
+            if chunk.usageMetadata.cachedContentTokenCount ~= nil then
+                uses.prompt_tokens_details = {
+                    cached_tokens = chunk.usageMetadata.cachedContentTokenCount
+                }
             end
         end
     end