perf: read each package once per track, not once per attempt

Vision took a crop from 5 ms to 25 ms, so re-reading the same package on every
attempt was the measured cause of the live coverage collapse: 789 -> 60 crops/s
at 8 workers, and 12/8/2/6 labels on the same clip at four pacings.

The reading is now cached on the track and later attempts re-run only the local
lane. Measured through the real server on add/image6.jpg:

  saturated pool   15.9 -> 49.4 crops/s (8 workers, 64 crops)
  per crop         142 -> 92 ms on a 300x400 crop
  3 s clip, send   19.7 mean labels (15-24, unstable) -> 24.0 (23-25, stable)
  12 s clip, send  48/131 -> 53/131; result pacing 53/134 -> 50/135

The 12 s numbers are within run-to-run noise; the short clip is where the
pacing sensitivity lived and it is now flat across pacings.

An empty read is not cached - a crop too small or too blurry to read says
nothing about the next one - and the cache is cleared with the rest of the
evidence on a method change or a drift relabel, so a reading never outlives the
product it described. The native crop lane skips OCR on the same rule.

Test: OCR runs once per track while the local lane keeps running. Disabling the
skip makes it fail 2 != 1. Suite 280 -> 281.

recognization/code/live.py

@@ -147,6 +147,13 @@ class TrackState:
     crop_attempts: int = 0
     # Which evidence produced the newest observation: frame or crop.
     last_source: str = SOURCE_FRAME
+    # The newest text reading for this object, cached so OCR runs once per
+    # track rather than once per attempt. Vision takes a crop from 5 ms to
+    # 25 ms, which is what made streamed coverage pacing-dependent: at 8
+    # workers throughput fell from 789 to 60 crops/s. Both fields are cleared
+    # with the rest of the evidence when the object generation changes.
+    text_rank: List[Tuple[str, float]] = field(default_factory=list)
+    text_read: str = ""
 
     def area(self) -> int:
         if not self.raw_box:
@@ -421,6 +428,8 @@ class LiveSession:
                     track.candidates = []
                     track.score_sums = {}
                     track.observations = []
+                    track.text_rank = []
+                    track.text_read = ""
                     track.attempts = 0
                     track.in_flight = 0
                     track.last_attempt_seq = 0
@@ -741,6 +750,10 @@ class LiveSession:
         track.candidates = []
         track.score_sums = {}
         track.observations = []
+        # The reading described the product the box has left, so it goes with
+        # the label rather than naming whatever moved into the box.
+        track.text_rank = []
+        track.text_read = ""
         track.attempts = 0
         track.in_flight = 0
         track.last_attempt_seq = 0
@@ -813,8 +826,15 @@ class LiveSession:
             # drift check later compares against.
             submitted_box = list(track.box)
             track_id = track.track_id
+            # OCR once per object: a track that already has a reading re-runs
+            # only the local lane, which is the difference between 5 ms and
+            # 25 ms per crop. A native crop is preferred for the one reading
+            # because it is the sharpest view the phone will send.
+            skip_ocr = bool(track.text_read)
             try:
-                future = self._executor.submit(self.engine.score_crop_lanes, crop, method)
+                future = self._executor.submit(
+                    self.engine.score_crop_lanes, crop, method, skip_ocr
+                )
             except RuntimeError:  # executor already shut down
                 with self._lock:
                     track.in_flight = max(0, track.in_flight - 1)
@@ -988,8 +1008,14 @@ class LiveSession:
                 self._release_crop_slot(record)
                 self._crops_rejected += 1
                 return _crop_error("CROP_STALE", stale)
+            # A native crop is the sharpest view the phone sends, so it is the
+            # one worth reading - but only once per object, like the frame path.
+            track = self._tracks.get(track_id_out)
+            skip_ocr = bool(track is not None and track.text_read)
             try:
-                future = self._executor.submit(self.engine.score_crop_lanes, crop, method)
+                future = self._executor.submit(
+                    self.engine.score_crop_lanes, crop, method, skip_ocr
+                )
             except RuntimeError:  # executor already shut down
                 self._release_crop_slot(record)
                 self._crops_rejected += 1
@@ -1134,6 +1160,15 @@ class LiveSession:
                 track.score_n += 1
             self._classify_times.append(finished_at - submitted_at)
 
+            # The reading is cached on the track so later attempts skip OCR.
+            # An empty read is not cached: a crop that was too small or too
+            # blurry to read says nothing about the next one.
+            if text_read:
+                track.text_rank = list(text_rank or [])
+                track.text_read = text_read
+            elif track.text_read:
+                text_rank, text_read = track.text_rank, track.text_read
+
             if not any(track.observations) and not text_rank:
                 if track.attempts >= self.settings.max_attempts:
                     track.state = STATE_UNRECOGNIZED