test(ublk): rapid property-based state-machine tests

[ValentaTomas]

Summary

Adds pgregory.net/rapid v1.3.0 as a test dependency and a new TestRapidStateMachine integration test in ublk/rapid_integration_test.go.

The state machine uses rapid's t.Repeat API to generate pseudo-random sequences of five actions:

• create — ublk.New a fresh device (capped at 2 live, 16 total per Run, ~10 ms each on a warm host)
• write — pwrite(O_DIRECT) random data at a block-aligned offset on a randomly chosen live device, then update the per-device shadow
• read — pread(O_DIRECT) and assert bytes equal the per-device shadow
• fsync — unix.Fsync on a live device
• close — full Close cycle: close the user fd first, run dev.Close() under a 5 s timer, then call dev.Close() a second time and assert it does not hang or error

A separate empty-key handler runs after every action and probes one block on every live device against the shadow.

Constraints (kept tight to make each Run cheap):

• block size 4096, device size 2 MiB
• write/read lengths drawn from {512, 4096, 8192}, offsets always block-aligned
• max 2 devices alive simultaneously, 16 creates per Run

Invariants checked

1. A Read returns bytes from the most recent Write at that range (per device).
2. Bytes written to device A never appear at the same offset on device B (cross-device isolation — falls out of (1) since each device has an independent shadow that the per-action invariant probes against).
3. Close terminates within 5 s (closeWithDeadline runs Close in a goroutine and times out via select — a hang in del_gendisk would otherwise deadlock rather than report a shrinkable failure).
4. Close is idempotent — a second call to Close must not panic or hang (mirrors TestCloseIdempotent).

fd-close-before-Close discipline

Per AGENTS.md: any test that opens /dev/ublkbN must close that fd before calling dev.Close(), otherwise del_gendisk blocks indefinitely. The state machine's closeDevice action does unix.Close(fd) first, and the per-Run cleanup (defer) does the same for any devices left alive at the end of a Run.

Why this is distinct from TestTortureRandomIO

TestTortureRandomIO is a long-running soak with fixed structure — N workers on disjoint regions of one device, doing random I/O within their region for the duration of the run. It does not exercise lifecycle transitions (only one device, no close mid-stream), and when it fails it gives you the failing op as-is — no minimization.

TestRapidStateMachine generates arbitrary command sequences including lifecycle transitions (create / close mid-stream, multiple devices), and rapid's automatic shrinking reduces any failing sequence to a minimal reproducer. That's the primary value: a 1000-action failing case shrinks to (typically) a handful of actions you can stare at.

Tooling

• make test-rapid — builds the integration binary and runs only TestRapid*. For local iteration on a shrunk failing case.
• The test runs as part of the existing test-integration CI job; no new CI job needed.
• Tunable via standard rapid flags / env vars: -rapid.checks=N or RAPID_CHECKS=N. See go test -args -h for the full list.

Test plan

• go vet ./... — clean
• golangci-lint run ./... — 0 issues.
• go test -count=1 -race ./ublk/uring/ ./ublk/ — pass
• go test -c -tags=integration -o /tmp/ublk.test ./ublk/ — compiles
• gofmt -l . — empty
• go mod tidy — clean
• Integration run on host with root + ublk_drv: deferred to CI's test-integration job (requires kernel module + privilege not available in the dev sandbox).

[cursor]

PR Summary

Medium Risk
Adds a randomized, property-based integration test that runs as root against real ublk devices; this can increase CI runtime/flakiness and exercises device create/close paths that can hang if assumptions change.

Overview
Adds pgregory.net/rapid as a dependency and introduces a new TestRapidStateMachine integration test that generates/shrinks random sequences of create/write/read/fsync/close operations while checking a per-device shadow model, cross-device isolation, bounded Close() completion, and Close() idempotency.

Extends developer tooling with make test-rapid to run only the new TestRapid* tests, and updates TODO.md by removing the now-completed rapid testing item.

^{Reviewed by Cursor Bugbot for commit ef6a719. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 02ac51e. Configure here.}