fix(sync): yield while reconciling large directories

[norgan]

Fixing both lag and visibility of very large file stores

Implementation	Detail
Directory reconciliation	At most 128 entries or 8 ms per turn; one reusable nonzero timer
Scheduling	Directory lookup and scheduling slot retained; child discovery after entry reconciliation
Cancellation	Pending continuations stopped; direct-receiver cancellation and nested event-loop regression coverage
Independent branch	Base 5bf7bdf42c30d93f9896d8fcbe1c44bbc58640af; no changes from #11087

Local Windows validation	Result
Release client, RemoteDiscoveryTest and DiscoveryBench build	Passed
RemoteDiscoveryTest	45 passed, 0 failed, 0 skipped
Original-base regression reproduction	4 expected failures: two large-directory yield rows and two cancellation rows
Changed-line clang-tidy and clang-format	Passed; no source changes
Human review and DCO	Reviewed and signed off

DiscoveryBench metric, milliseconds	Baseline median (range)	PR median (range)
Queued root-health reply delivery	1119 (1104–1544)	9 (9–10)
Maximum heartbeat gap	1345 (1324–1813)	269 (269–412)
Discovery	1467 (1444–1943)	1732 (1504–3194)
Complete fake sync	79292 (79180–86350)	81922 (79896–98303)

Benchmark conditions and limits	Detail
Workload	1 directory, 5000 regular files, 5 parallel jobs; 3 interleaved baseline/PR pairs
Transport	Queued fake replies; no socket or server latency
Correctness	All 6 complete fake syncs successful; 5000 files each
Throughput improvement	Not established
Complete large live sync	Still running; completion unverified
macOS and Linux validation	Not run in this Windows environment

AI assistance: Assisted-by: Codex:GPT-6 (implementation, tests, source analysis and validation).