[AlxcBTW]

Follow-up to #124.

real_window::the_worker_saves_a_clip_of_a_real_window (share-worker access.rs; 3 failures, all ended: no_window). The test waited a fixed 3 s after Calculator's process appeared; on a cold start (this test runs first) the window is not listed by then, and the worker only looks for a few seconds. The test now polls sck::find_window until the system lists the window (up to 60 s) before it starts the worker. What the test proves is unchanged. One failure (02:31 UTC) was on a head that already had #101, so #101 did not cover this.

share_play.rs:103 "ring cannot be created: File exists" (2 failures): no change. Both runs (01:40 and 01:44 UTC) predate #102 (02:04 UTC), which added the per-process counter to ring_name; no failure since.

tunnel.rs:451 without_an_answer_to_the_probes_the_relay_still_makes_it_type_2 (2 failures). After relay.stop() the endpoint must notice the loss within 10 s. Measured on an idle Windows machine: 61-64 ms (the relay's close arrives). I could not reproduce the failure (the build laptop was unreachable, no macOS here), so the cause is inferred, not observed: when that close does not reach the endpoint, the endpoint only learns from its QUIC connection's idle timeout, 35 s in iroh-relay 1.4.1 (quic.rs). The bound is now 45 s, derived from that constant. The test still fails if the endpoint never notices. Unverified on macOS until CI runs it repeatedly.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}