[ismaeelakin]

What changed

Adds stage timings for the UpgradableRwLock investigation in #404:

• uncontended acquisition and promotion measured separately
• acquisition while an ordinary reader is held
• promotion while 1, 4, or 16 ordinary read guards remain held
• promotion while a writer is queued behind the existing reservation
• a later upgradable request queued behind a writer
• --test support for the repository benchmark smoke run

This PR adds the measurement step requested in #404; it does not change the lock implementation or establish a speedup. Per-operation clock reads add overhead. Reader-promotion timings include dropping the readers and the wake/handoff, and queued acquisition includes the writer's handoff. The benchmark manually polls futures with a non-scheduling waker, so its timings exclude executor scheduling overhead. Treat the results as diagnostics.

Verification

Passed locally on Windows 11 (Intel Core i5-6300U):

• cargo x test and benchmark runs with stable rustc 1.99.1
• cargo x lint with nightly rustc 1.101.0
• cargo x bench -- --test
• cargo x bench --bench upgradable_stages -- --rounds 1000000 --repeat 5

The local timings and their run-to-run spread are recorded in issue #404.

Refs #404

[orthur2]

cargo x bench -- --list worked before and now fails, because the default run includes this target and it rejects --list. I'd leave it out of the default run and smoke-test it separately in CI.

[orthur2]

drop(readers) also frees the Vec inside the timed window. Reusing one buffer across rounds and calling clear() here avoids that.

[orthur2]

An empty timed interval costs about 15 ns on my M5, roughly half the reported acquisition and promotion times. Could we print that baseline alongside the results?

[orthur2]

Thanks for adding the queued cases.

[ismaeelakin]

Thanks as well for the careful review. I kept upgradable_stages out of the default cargo x bench run so cargo x bench -- --list works, and added a separate CI smoke run for the target. The reader buffer is now reused, so its backing allocation is not freed inside the timed interval. I also report an empty interval with the same clock calls, without subtracting it from the measurements. The default listing, stage smoke run, and target-selection test each passed, cargo x test passed as well.