[elementalvoid]

Summary

crates/app/tests/                     crates/app/tests/it/
  dashboard.rs  ┐                       main.rs   (one binary)
  diagnostics.rs├─ 19 binaries          ├─ dashboard.rs
  …            ├─ ×230 MB GPUI link     ├─ diagnostics.rs
  shell.rs     ┘                       └─ common/   (compiled once)

Five changes, each measured before it was made:

Change	Local	CI
51 integration binaries → 4	app test build 24 s → 5.8 s	19 fewer GPUI-stack links on macOS/Windows, ~4.5 GB less to cache
cargo test → cargo nextest	gate 2 m 51 → 47 s	macOS 8 m 03 test step → ~4 m (predicted)
stall guard on an injected window	immich-client 60.3 s → 8.4 s	—
cache prefix drops the manifests	—	release-version-bump runs stop being cold builds
CARGO_PROFILE_TEST_DEBUG=0	—	smaller test binaries to link and to cache
Linux GUI smoke job dropped	—	one fewer job, distro package list and shell script; it was never on the critical path

Evidence

before   cargo test --workspace               real 2m51s   911 tests, one binary at a time
         56 test executables; 60.2 s inside one real-time bandwidth test

after    cargo nextest run --workspace        real 47s     912 tests, 2 skipped
         10 test executables (4 integration + 6 unit)

worst single test   60.22s → 0.65s      immich-client 60.3s → 8.4s
app edit → all test targets   24 s → 5.8 s

CI, warm cache, run 41266533223: Linux 3 m 49 (test 2 m 52), Windows 7 m 37 (test 5 m 45), macOS 9 m 31 (test 8 m 03, the critical path).
CI, cold, run 41269865243: "No cache found" on Linux and Windows — a version bump in every manifest changed the cache prefix. Windows clippy 4 m 24 against 57 s warm.
CI on this branch: run 42179209035 (four jobs, cold by design, prefix v0 → v1) and run 42335270218 (five commits) both green on every job.

Two bugs found on the way, each in its own commit:

• A latent flake. bandwidth::tests::a_spent_bucket_reports_the_time_the_budget_needs read Instant::now() twice, so a thread descheduled for 80 ms refilled the bucket: an empty bucket at 1 kB/s needs 100 ms: 20.026619ms. The clock is now read once.
• Half a minute of wall clock per timing test. Four limits.rs tests spent 60.2 + 37.9 + 30.1 + 44.9 s proving four properties about a 30 s window. ImmichClient now carries its window (REQUEST_WINDOW, 30 s) with with_window for tests, replacing a #[cfg(test)] static that two parallel tests could overwrite. Real sockets, real partial-body byte counts, a 200 ms window, and tighter assertions because the arithmetic is visible. The 30 s production value is asserted once, by the_production_window_is_the_specified_thirty_seconds.

Tried and rejected: tokio::time::pause(). It turned the 60 s test into 0.156 s with no production change, but auto-advance cannot tell a runtime parked on a socket from one parked on a timer, so the next real round-trip raced its own deadline and failed as TimedOut on the login handshake.

cargo machete is clean; the dead serde and three redundant dev-dependencies in mock-immich are gone. The 86 crates that resolve to two versions all belong to the pinned gpui stack, a platform binding, or a build-time host — the dependency audit is in that commit message.

Merge Danger

Door: two-way — revert restores the old layout; the caches need one cold run either way.

Blast Radius: test layout, test timing and CI; one new public method (ImmichClient::with_window) and no change to any production timing behaviour.

• --test <file> filters are now cargo nextest run -p <crate> -E 'test(<module>)'. CONTRIBUTING.md and the source doc comments are updated; any agent prompt or script still using the old form silently selects nothing.
• The first run after this is a cold cache on all three OSes by design (prefix v0 → v1).
• CARGO_PROFILE_TEST_DEBUG=0 drops line numbers from test backtraces (panic locations survive). One line in ci.yml to revert if a CI failure needs them.
• with_window is public API. The alternative is a test-support feature or moving the tests in-crate; both were considered and are worse than one documented knob.
• The GUI smoke job is gone. It was the only automated launch of the real binary on any OS (startup, window creation, Vulkan init, a non-blank frame, the single-instance handoff) and it was never on the critical path, so CI wall time is unchanged; what it cost was a second distro package list, a 60-line inner shell script and a cache key that had to stay restore-only. Those behaviours are now the capture recipe in CONTRIBUTING.md, run by hand when a ticket touches startup, rendering, the tray or the handoff. If a real-GPU regression ever slips through, this is the job to bring back.
• Known interaction: engine_model::a_click_on_a_running_engine_starts_a_scan is a wall-clock budget test (64 MiB hashed to outlive a 250 ms status poll) with no deterministic barrier. nextest's parallelism widens the starvation window: it failed once at 25.4 s on a loaded 12-core box, passed alone in 1.5 s, and passed on the next full run and on CI. CONTRIBUTING.md already lists it. The fix is a barrier on EngineModel's surface, not a wider bound.
• implement-orchestrator-rules.md is untracked in the operator's tree; the rule changes it needs (touch narrowed to the uplift path, mold dropped) are applied there but not part of this PR.