W9 implementation report

Branch: zdp/8-train-fast
Worktree: /home/pablo/.paseo/worktrees/26sr02x6/dirty-swan

APPLIED — runtime accelerator flags. dbfe5716 adds --compile-mode, --channels-last, --fused-adamw, and --amp-dtype. Defaults preserve current behavior. Production CUDA isolation passed one compiled, channels-last step with fused AdamW.
APPLIED — schedule controls. 9a19baef adds OneCycle overrides and constant-cooldown; 559b10fe fixes its optimizer-step boundary. Tests prove default scheduler-state equality and a [0.7, 0.4, 0.1] cooldown tail.
APPLIED — progressive resolution. 1e552eaa parses fractional resolution stages, rebuilds the loader at boundaries, and persists training_stage. zipdepth/catalog/dataset.py was unchanged.
APPLIED — named presets. 8620cba1 adds --preset {v4,fast} and persists the resolved recipe. fast currently aliases bit-compatible v4; winning fast values remain DEFERRED until the hill-climb.
APPLIED — tests/static gates; live gate externally blocked. ca49e726 strengthens bit-exact model-state and runtime override coverage. Final results: lint passed, typecheck passed with 0 errors, deadcode passed, 103 tests passed, 1 skipped. Tyro help exposes every requested flag.

The live command was attempted twice:

pixi run -e zipdepth-catalog-dev --frozen zipdepth-catalog-train-smoke --compile-mode default --channels-last --fused-adamw

Both attempts failed before training because the shared catalog returned ScanSegmentTable failed: http2 error (Unavailable) from 127.0.0.1:51235. The server was not killed or restarted. Rerun the command once the catalog is healthy; expect the loss, resume, torchrun, checkpoint-inference, and final catalog smoke passed gates.

Safeguards: clean worktree; no push or branch switch; no vendored, dataset, loss, architecture, or evaluation changes.

Full report: W9-REPORT.md