🎯 Phase 6 — Evaluation, v0.1.0

WP-071 — packed uint8 batch transport

The transport format and its error bound are A33. What the register does not carry is the shape of the win: flattening the ragged list[Targets] into eight dense tensors takes the DataLoader IPC hop from roughly 600 pickled segments per batch to 9, with bytes down 4x from the uint8 quantization. Segment count, not byte count, is what makes a containerized shm ceiling bite.

WP-073 — compact annotation store

CocoDetectionDataset held the raw JSON annotation dicts after building its per-image Targets. Millions of small Python objects, each with a refcount touched on every access — and refcount traffic materializes copy-on-write pages in every persistent DataLoader worker, so a fork-shared dataset is only shared until Python reads it.

Observed on a Colab 176 GB host, killed at validation boundaries: epoch 1 at 48 workers, epoch 24 at 32 workers. The worker count sets how fast the leak accumulates, not whether it does.

Dropping the dicts fixed the growth; the val-loader worker cap (min(num_workers, 4)) addressed the second half, a persistent val pool doubling the worker population at exactly the moment the training pool is still resident. That cap is what WP-102 later narrowed and WP-103 restored for inherited counts.

WP-076 — epoch-recycled loader workers

WP-073's compact store did not end the OOMs: Colab still killed the run at epoch 24 and epoch 38. The residue is not one leak but a class of them — allocator high-water growth, arena fragmentation, and whatever copy-on-write pages survive — all of which are per-worker and none of which a dataset change reaches.

Respawning the pool costs seconds against a 9-minute epoch and hard-resets every one of them, so persistent_workers defaults off. This is a case where the cheap blunt fix dominates: no diagnosis of which allocator behaviour dominated was needed, because recycling ends all of them and the measurement said the cost was noise.

WP-079 — per-worker, per-epoch augmentation RNG: what the collision cost

Measured exposure: 1/3200 of the intended draws at 32 workers over 100 epochs. Two DataLoader defects compose multiplicatively and neither is visible in a loss curve — the run trains, it simply trains on far less augmentation variety than the config describes. The mechanism is a _TrainPipeline bug fixed the same day; the full story is at wp-079.md#wp-079-engineering.

WP-082 — objective invariant gate

Written after WP-078 and WP-079 survived 474 tests and three tier runs, which is the fact that justifies the package: the suite was large and did not contain the shape of test that would have caught either.

The useful negative result is the test not written. An obvious guard for WP-078's defect is "the L1 term must not exceed some share of the objective". On a converged fixture the broken pixel frame reads 10.7% against the correct frame's 1.5% — both comfortably under any threshold anyone would have picked. Such a test would look like protection and catch nothing. The gate pins the coordinate frame directly instead (1/stride against the pixel frame, per level) and its scale on a near-converged state (0.75 stride-units against 5.99 broken).

WP-083 — synthetic-shapes generalization golden

Frozen at NMS mAP50-95 0.8823, E2E 0.8583, mAP50 0.9842 — three MPS runs identical to four decimals, 127–169 s wall clock. The spread in wall clock against the exactness of the metric is the point: it is a cheap gate whose number is stable even though its timing is not. The harness this measures — 2000 generated scenes, held-out scoring, mosaic held on for five of six epochs specifically so a collapsed augmentation RNG (WP-079) would be detectable — is described at wp-083.md#wp-083-engineering.