Add Qwen3.6 validation and benchmark coverage

tools/bench/quality.py

@@ -14,6 +14,7 @@ extra moving parts aren't worth it for the first iteration.
 from __future__ import annotations
 
 import asyncio
+import random
 import statistics
 import time
 from dataclasses import asdict, dataclass
@@ -38,6 +39,7 @@ class QualityRow:
     n_total: int
     n_correct: int
     n_errors: int
+    n_length: int
     accuracy: float
     mean_completion_tokens: float
     mean_ttft_ms: float
@@ -58,7 +60,9 @@ class QualityCase:
     tpot_ms: float
     e2e_s: float
     error: str | None
+    finish_reason: str | None
     response_preview: str
+    response_text: str
 
 
 async def _run_one_task(
@@ -66,7 +70,10 @@ async def _run_one_task(
 ) -> tuple[QualityRow, list[QualityCase]]:
     problems = task_mod.load()
     if cfg.quality_limit is not None:
-        problems = problems[: cfg.quality_limit]
+        if cfg.quality_seed is not None and cfg.quality_limit < len(problems):
+            problems = random.Random(cfg.quality_seed).sample(problems, cfg.quality_limit)
+        else:
+            problems = problems[: cfg.quality_limit]
     print(f"[quality] {ep.name} / {task_name}: {len(problems)} problems "
           f"(max_tokens={max_tokens})")
 
@@ -75,13 +82,20 @@ async def _run_one_task(
     async with httpx.AsyncClient(timeout=cfg.request_timeout_s) as client:
         for prob in problems:
             messages = task_mod.make_messages(prob["problem"])
+            extra_body = dict(ep.extra_body or {})
+            extra_body.update({
+                "top_k": cfg.quality_top_k,
+                "top_p": cfg.quality_top_p,
+                "presence_penalty": cfg.quality_presence_penalty,
+                "repetition_penalty": cfg.quality_repetition_penalty,
+            })
             r = await chat_stream(
                 client, ep.base_url, ep.model_id, messages,
                 max_tokens=max_tokens,
                 temperature=cfg.quality_temperature,
                 api_key=ep.api_key,
                 timeout=cfg.request_timeout_s,
-                extra_body=ep.extra_body,
+                extra_body=extra_body,
             )
             pred = task_mod.extract_answer(r.text) if r.error is None else None
             correct = task_mod.score(pred, prob["answer"]) if r.error is None else False
@@ -91,8 +105,9 @@ async def _run_one_task(
                 correct=correct, completion_tokens=r.completion_tokens,
                 ttft_ms=r.ttft_s * 1000 if r.ttft_s > 0 else -1.0,
                 tpot_ms=r.tpot_s * 1000 if r.tpot_s > 0 else -1.0,
-                e2e_s=r.e2e_s, error=r.error,
+                e2e_s=r.e2e_s, error=r.error, finish_reason=r.finish_reason,
                 response_preview=(r.text or "")[:240].replace("\n", " "),
+                response_text=r.text or "",
             ))
             mark = "✓" if correct else ("E" if r.error else "✗")
             print(f"  [{mark}] {prob['id']:>4s}  gold={prob['answer']:>6s}  "
@@ -109,7 +124,9 @@ async def _run_one_task(
         n_total=len(cases),
         n_correct=correct,
         n_errors=errors,
-        accuracy=correct / max(len(cases) - errors, 1),
+        n_length=sum(1 for c in cases if c.finish_reason == "length"),
+        # Transport/runtime failures are incorrect attempts, not exclusions.
+        accuracy=correct / max(len(cases), 1),
         mean_completion_tokens=statistics.mean(c.completion_tokens for c in ok) if ok else 0.0,
         mean_ttft_ms=statistics.mean(c.ttft_ms for c in ok if c.ttft_ms > 0) if ok else -1.0,
         mean_tpot_ms=statistics.mean(c.tpot_ms for c in ok if c.tpot_ms > 0) if ok else -1.0,