Happy-path E2E visual test (full UI flow) + Quantize page + Predict compare

[mrjunos]

What

One happy-path E2E visual test that launches the real Streamlit UI and drives the whole pipeline end-to-end, recording the run to a .webm:

Tray capture → crop/segment → Train → Evaluate → Quantize/Export → Predict (compare float vs INT8)

It runs on a tiny committed real-crop dataset (≈320 KB), so it needs no DVC, and doubles as the CI gate: a new e2e job installs Chromium, runs the test, and uploads the recording as an artifact. The existing fast job now runs pytest -m "not e2e".

Product changes the flow needed

• New Quantize/Export page (page_quantize.py), separate from Train as requested: exports float ONNX + INT8, shows sizes / size-reduction / parity, with a dynamic | static | none mode selector.
• Predict compare mode: run a model's float and INT8 ONNX side by side on the same bean (top-1 agreement + per-model latency).

Bugs the E2E surfaced (and fixed)

• Tray Save never worked: segmentation results and the Save button were gated behind a transient Process button, so clicking Save reran with Process unclicked → early return → the save never executed. Now the result is persisted in session_state.
• Train could hang on "Training…": the "Done" status was keyed on OS-process liveness while auto-refresh stopped on the metrics done event — a race that left the page stuck. Now keyed on the done event.

Harness / fixtures

• tests/e2e/build_fixture.py — samples the committed mini-dataset from local data/processed/ (run once locally; reuses canonical taxonomy indices).
• tests/e2e/synth_tray.py — composites real crops into an ArUco tray photo (segmentation verified 8/8 wells).
• tests/e2e/harness.py — builds an isolated sandbox (ALMENDRA_ROOT + cwd) with fast/offline config tweaks (pretrained off, cpu, num_workers 0, dynamic INT8) and launches almendra ui.
• e2e extra (Playwright), pytest e2e marker, make e2e / make test targets.

Verification

• make test → 68 passed (e2e deselected); ruff check . + format clean.
• make e2e → passes in ~28 s locally; .webm written under tests/e2e/recordings/ (gitignored).
• Quantize mode: the gate drives dynamic INT8 (reliable, calibration-free → deterministic green gate); the page still offers static for real use.

🤖 Generated with Claude Code