fix(claude): reconcile max_tokens vs thinking budget and lift per-model ceiling (#2381)

On the translated OpenAI->Claude path, adjustMaxTokens capped max_tokens
before applyThinking set thinking.budget_tokens, so max-effort budget
(128000) could exceed a 64k-clamped max_tokens -> Anthropic 400.
prepareClaudeRequest now reconciles after the budget is known: prefer
raising max_tokens, only shrink budget when it meets/exceeds the ceiling.

Also lift the global 64000 cap: the ceiling is now the model's real
maxOutput, so high-output models (fable/mythos, opus-4.8/sonnet-4.6) get
their full budget. adjustMaxTokens gains an optional ceiling arg (default
unchanged, callers untouched); openai-to-claude passes the model maxOutput.

Native Claude Code passthrough is unaffected.

Co-Authored-By: Claude <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

open-sse/translator/request/openai-to-claude.js

@@ -6,6 +6,7 @@ import { safeParseJSON } from "../concerns/json.js";
 import { parseDataUri } from "../concerns/image.js";
 import { extractTextContent } from "../formats/gemini.js";
 import { ROLE, OPENAI_BLOCK, CLAUDE_BLOCK } from "../schema/index.js";
+import { getCapabilitiesForModel } from "../../providers/capabilities.js";
 
 // Empty prefix matches real Claude Code behavior (no tool name prefix).
 // Previously "proxy_" was used but this is a detectable fingerprint difference.
@@ -15,9 +16,13 @@ const CLAUDE_OAUTH_TOOL_PREFIX = "";
 export function openaiToClaudeRequest(model, body, stream) {
   // Tool name mapping for Claude OAuth (capitalizedName → originalName)
   const toolNameMap = new Map();
+  // Cap max_tokens at the model's real output ceiling (e.g. Opus 4.8 = 128000),
+  // not the conservative 64000 default — otherwise a high-output model is
+  // pre-clamped here before prepareClaudeRequest's model-aware step runs.
+  const modelCeiling = getCapabilitiesForModel(null, model).maxOutput || undefined;
   const result = {
     model: model,
-    max_tokens: adjustMaxTokens(body),
+    max_tokens: adjustMaxTokens(body, modelCeiling),
     stream: stream
   };