Additional pipeline tests

[evanjt]

Tests for the features added in #25: device selection, orchestrator lifecycle (cancel, pause, viewer protocol), mapping progress callbacks, LoGeR progress counting, and resume with uncompressed npz.

[HuguesSib]

Review

Ran the suite locally on this branch: 139 passed. Good coverage of the #25 surface, and the byte-identical / reference-comparison style is the right call. One test is broken, though.

test_failed_run_stops_the_sampler_thread is vacuous 🔴

It counts threads named "drm-resource-sampler", which exists nowhere in the repo except that test — the only threads the package ever creates are viser-render and simple-viser-render. So _sampler_threads() returns [] before and after, and the assertion is 0 == 0.

Verified rather than inferred — asserting the counted set is empty on both sides passes, with only MainThread alive through three failing runs:

LIVE THREADS: ['MainThread']
1 passed

It would stay green through any real thread leak. The docstring describes a sampler that presumably lives in the GUI repo; the assumption seems to have leaked in.

Fix: drop it, or retarget it at a thread that actually exists (viser-render) with a real viser app in play.

The rest — solid ✅

• test_device.py — nice short-circuit check: capability 7.5 → float16 without running the flash probe, enforced via an AssertionError side-effect. Dynamo globals restored in finally.
• test_resume_npz.py — the strongest one here: saves the npz exactly as the orchestrator does, reloads, proves the cloud is byte-identical fresh-vs-resumed. Locks the Add support for GUI, type fixes, dependency updates #25 local_points change down properly.
• test_orchestrator_lifecycle.py (rest) — owned-vs-injected viewer semantics are exactly right, and the signature-position guard protects GUI callers from enable_viser drift.
• test_loger_progress.py — _count_windows expectations check out against the implementation; the block-progress test also re-asserts the transform against a single-shot reference, so it can't silently accept a broken re-anchor.

Minor

• _single_shot_reanchor_world is duplicated here and in test_loger_reanchor.py — a shared helper would avoid drift.
• _run_expecting_failure uses a broad pytest.raises(Exception); tightening it would catch a test failing for the wrong reason.

Verdict: approve once the sampler test is dropped or retargeted.