Follow-up issues to async results posting #113

[mihow]

Follow-up items from code review of #113.

1. Benchmark total_images counter uses configured batch size instead of actual count

trapdata/antenna/benchmark.py:121 increments total_images += batch_size (the configured parameter) rather than the actual number of items in the batch. The last batch is typically partial, so this overcounts and corrupts the reported success rate.

Fix: total_images += batch_successful + batch_failed

ami-data-companion/trapdata/antenna/benchmark.py

Lines 116 to 123 in 50996ea

	# Count images in this batch
	batch_failed = len(batch["failed_items"])
	# Successful items are those with reply_subjects that are not in failed_items
	batch_successful = len(batch["reply_subjects"])
	total_images += batch_size
	total_successful_images += batch_successful
	total_failed_images += batch_failed

2. ResultPoster.shutdown() can mask exceptions

If an exception propagates out of the batch loop in _process_job, the code jumps to the cleanup path without calling wait_for_all_posts() first. If pending_futures is non-empty at that point, shutdown() behavior may be unexpected. Consider wrapping the loop in try/finally that calls wait_for_all_posts() before shutdown().

ami-data-companion/trapdata/antenna/worker.py

Lines 165 to 167 in 50996ea

	)
	continue
	crop = image_tensor[:, y1:y2, x1:x2]

3. Removed pre-batch progress logs reduce debuggability

The PR replaced two per-batch logger.info calls ("Processing worker batch N (M images)" and "Total items processed so far: N") with a single end-of-batch summary log. If a crash occurs mid-batch during detection or classification, no progress is logged before the error, making it harder to diagnose which batch failed.

ami-data-companion/trapdata/antenna/worker.py

Lines 343 to 345 in 50996ea

	detections_for_terminal_classifier = detector_results
	detections_to_return = []

4. load_time metric includes post-submission overhead

The timer resets after classification. The next iteration's load_time measurement therefore includes result-object construction, post_async() call, and the logger.info call, in addition to actual dataloader I/O. This inflates reported load times for batches after the first.

ami-data-companion/trapdata/antenna/worker.py

Lines 167 to 171 in 50996ea

	crop = image_tensor[:, y1:y2, x1:x2]
	crop_pil = to_pil(crop)
	crop_transformed = binary_transforms(crop_pil)
	binary_crops.append(crop_transformed)
	binary_valid_indices.append(idx)