[yogthos]

Stacked on #1340 (base is fix/sleep-overshoot; retarget to main once that merges).

#1340 made Thread/sleep exact. Every other timed wait still built its deadline as a whole millisecond rounded up from now, so a 1ms wait took ~2ms where the JDK takes ~1.25ms. This makes mono-deadline exact for all of them and fixes the places that compared or subtracted deadlines against the truncated mono-millis clock.

Sites: timed deref of promise/future, BlockingQueue poll/offer, CountDownLatch.await, ReentrantLock.tryLock, awaitTermination, Future.get, CompletableFuture.get, FutureTask.get, Thread.join, Object.wait, Process.waitFor, await-for, core.async timeout, scheduled executor delays/getDelay/keep-alive, and the io poller (socket SO_TIMEOUT on read/accept). jolt has no Semaphore or Condition, so those aren't covered.

Median of 101 x 1ms waits, macOS, in us:

site                 JDK 21   before   after
deref promise        1267     1995     1273
BlockingQueue.poll   1342     1997     1270
latch await          1300     1988     1271
tryLock              1299     3525     1270
awaitTermination     1289     1996     1274
Process.waitFor      1296     2528     1274
Object.wait          1284     1995     1277
read SO_TIMEOUT 1    1169     2373     1178
poll 500us            680        0      637

Also fixed:

• tu->ms truncated sub-millisecond units, so (.poll q 500 TimeUnit/MICROSECONDS) returned at once and a 1500us schedule delay was 1ms.
• A timed tryLock ignored the interrupt flag. The JDK throws InterruptedException even when the lock is free.
• tryLock kept polling for a whole millisecond past its deadline, with a 1ms nap on top.
• ms->duration floored fractional ms, which turned a sub-ms pause into a busy loop.

One-shot waits now pass a timeout that starts at the first actual wait (jolt-timeout in locks.ss), so a wait that answers at once doesn't read the clock. Building an exact deadline costs a rational division, and skipping it makes the non-blocking paths faster. A/B/A, ns/op min (median):

settled deref 1      128 (129) / 69 (70) / 131 (136)
latch(0) await 1ms   145 (145) / 83 (85) / 149 (154)
free tryLock+unlock  116 (116) / 74 (74) / 120 (124)
offer+poll 1ms       107 (107) / 107 (107) / 110 (115)
Future.get done      148 (149) / 93 (94) / 154 (156)

Tests: async-timer-test section 11 times each site's 1ms wait interleaved with a sleep(1) from the same run, and retries a round that noise puts behind. All 15 rows fail on #1340's code. There's a socket-test row for SO_TIMEOUT, and corpus rows for interrupted tryLock, negative timeouts and sub-ms schedule delay, all checked against the clojure CLI.