[yogthos]

sample-never-derefs-one-input-twice-at-once times out intermittently on CI with jolt 0.8.21 and nightly (#13). Dispatch runs of ci-nightly-jolt: 0.8.2, 0.8.10, 0.8.16 and 0.8.20 passed every run (9/9 on 0.8.20), 0.8.21 failed 2 of 8, nightly failed 3 of 8.

Root cause is in sample. Sample.java's monitor blocks every other notifier while a transfer reads an input, so its do/while only re-reads for notifications the read itself raised, and a surplus notification is discarded by the thread that raised it. Ebb reads outside the lock, so a notification from another thread that lands mid-read made the transfer read again, and a surplus was left for whoever held the slot. An input notified about as fast as it can be read keeps one transfer reading for as long as that lasts.

The test's input sits right on that line since jolt 0.8.21. jolt 45bfe3df rounds deadlines up, so (Thread/sleep 0 n) went from returning at once to ~1.25ms (the nanos are ignored), which made the notifier's period and the read's length the same. The transfer then ends at a random point: the four-sample probe took up to 2.5s locally (8.1s under load) on nightly versus 65ms on 0.8.20, and on a 2-core runner it passes the 15s deadline. With the jolt sleep fixed to honour nanos (separate jolt PR) the notifier is 5x faster than the read and main's sample hangs every time.

Fix:

• A read marks itself in a carried context-local, so notifications from inside it, including an ap or cp input's owner fiber serving the deref, still count as the read's own. Anyone else's notification during a transfer is recorded as late and becomes the mark for the next transfer.
• An outside surplus waits for the slot and discards for itself instead of charging the holder.
• A transfer waiting for a slot is handed it on release. Without that, an outside notifier draining its own surplus retook the slot before the transfer woke, which failed every run on the 2-core runner.

Tests: two new stress tests. One has an in-protocol input re-notified from another thread during every read; before the fix it does 440k reads for 12 samples and times out. The other has a notifier that always outruns the reads; it timed out before too. Both pass now. The existing test is unchanged. Under load locally the probe now stays under 100ms per run. bin/test passes on nightly, 0.8.21 and 0.8.2.