# Readthrough before the production-workbench integration

Inspected initial HEAD `61e484368c89034d9cbb5f19b08c32d50eb246bd` and refreshed against `f04dd5f2838465064cd6fb8efa7bc72bcc81bb80`, including the new hanger identity, 29-lot rack and 18-active/7-planned test registry, as well as README, pipeline, agent contracts, page builder, tracer, subject queue, shot state and test implementation on 20 October 2026.

## What is strong

The repo has a concrete bridge between two experiments. The selected-drawings batch preserves 40 reviewed still keys, prompts, rejected revisions and independent critiques. The motion study represents a walking giraffe using a shared rig, timed poses, camera displacement and six depth planes. The tracer converts selected raster keys to editable strokes; committed recall is at least 0.984 across the 40 keys. That is an ink-retention measure, not proof of animation or stable topology.

The shot state, exposure sheet, source identifiers, lineage and explicit failures make repairs inspectable. The agent contract correctly separates still acceptance from playback acceptance and rejects vacuous tests. The page builder keeps the existing operator pages reproducible and usable offline.

## What remains open

Background/character separation is unresolved. The trace layer heuristic marks about 19.8% of the dog batch's total strokes as persistent. This indicates instability; because the denominator includes character strokes and the split is heuristic, it is not a standalone measurement of background retention. Keep one background plate and draw the character separately, then measure the actual background layer.

The walking giraffe's committed test report has 11 true statuses, two failures and one unrun test. Several statuses are historical assertions in `tests.py`: M1 explicitly predates the head revision; S4 records an earlier browser render; S5 is asserted from construction despite an occlusion concern. J2 records a prior blind preference. These must not be described as fresh checks on the current render. The current study still lacks playback approval; style typicality and action readability remain unresolved in the saved report.

The scripts that produced browser model outputs are not committed, so those inference stages cannot yet be reproduced from this checkout. No model training is established. The new production pool must not convert retrieval scores, source-sample acceptance or queue completion into motion approval.

## Integration boundary

Add the 100-candidate workbench, timing references and coordinator beside the existing work. Preserve all current rig code, poses, traced strokes, test values, quality thresholds, wave choices and critic verdicts. Reuse the canonical selected-drawings assets. Keep the two giraffe clip IDs distinct. Add navigation through the page-builder source so regeneration preserves the integration.