Harden trace replay measurement integrity

src/aituner/http_client.py

@@ -315,7 +315,15 @@ def stream_chat_completion(
     ):
         tpot_ms = None
     else:
-        tpot_ms = ((last_token_at - first_token_at) / max(used_tokens - 1, 1)) * 1000.0
+        # A response that delivers all content-bearing chunks at one observed
+        # instant has no measurable inter-token interval.  Reporting 0 ms here
+        # would turn an unobservable TPOT into an artificial SLO pass.
+        elapsed_s = last_token_at - first_token_at
+        tpot_ms = (
+            (elapsed_s / max(used_tokens - 1, 1)) * 1000.0
+            if elapsed_s > 0
+            else None
+        )
     return StreamMetrics(
         ttft_ms=ttft_ms,
         tpot_ms=tpot_ms,

src/aituner/search.py

@@ -38,7 +38,10 @@ def binary_search_max_feasible(
     for _ in range(max_probes):
         if cur_high - cur_low <= tolerance and best_payload is not None:
             break
-        threshold = round((cur_low + cur_high) / 2.0, 12)
+        # Thresholds are also trace-selection boundaries.  Rounding a midpoint
+        # can cross a real sampling_u value and silently replay a different
+        # workload than the materialized window specifies.
+        threshold = (cur_low + cur_high) / 2.0
         probe = cache.get(threshold)
         if probe is None:
             probe = evaluator(threshold)