perf(runtime): reduce timed RPC emission overhead and restore batching

[SunSi12138]

Timed Unary requests currently create and supervise a send-completion task for every request. They also force a transport flush per timed frame, even when more requests are already queued. The release ABA against published 1.1.1 exposed higher client allocation and lower high-concurrency throughput on main b3078000d80a7400373ea5ae67a1e3ffd8e5dbcc.

The existing send pump now reports queued request failures directly to the pending-call table. It captures the request ID before returning the buffer and invokes the failure owner outside the admission lock. Successful Unary calls still complete only on their response; a request that expires at emission still completes with DeadlineExceeded without requiring a timer callback. Pooled frames retain the observer in their existing completion-state reference.

Ready timed requests can also share one bounded publication batch. The pump acquires one output span, copies the complete retained suffix, then samples the monotonic clock once. It removes expired requests, compacts surviving frames in order, stamps their remaining budgets, and starts FlushAsync before expired-request callbacks or telemetry. A timed batch never waits for new arrivals or the configured batch timer; existing size thresholds, LowLatency flushes, and the 64-frame progress interleave remain enforced.

Validation:

• Full non-incremental Release build: zero warnings/errors.
• Unit: Passed 1825, Failed 0, Skipped 0. Integration: Passed 451, Failed 0, Skipped 0.
• Final focused emission tests: Passed 9, Failed 0, Skipped 0. This includes mixed timed/untimed batches and expiration at the first, middle, last, or every timed request, checking wire order, exact remaining budgets, callback timing, and returned owners/reservations.
• TimedUnaryShouldObserveEmissionWithinTheSendPumpLifetime fails on unchanged b307800 production code and passes with the correction. TimedUnaryDroppedAtEmissionShouldCompleteWithoutDeadlineTimerCallback remains passing.
• Negative control for QueuedTimedRequestsShouldSharePublicationAndDropExpiredFrames: all five original cases fail against pre-batching 6a22b42 production code; all pass with batching restored. The final cases additionally retain an untimed frame inside the compacted suffix.
• The first correction alone (6a22b42), measured with BenchmarkDotNet 0.15.8 on Apple M4, .NET 10.0.11, Server GC, processor count 8, reduces timed loopback Add allocation from 1534 to 936 B/op and completed work items from approximately 8 to 7. Untimed allocation stays 768 B/op. This is allocation evidence for that first correction; measured time differences overlap variation and are not claimed as throughput improvements.

Final BenchmarkDotNet comparison of b307800 versus 0141a85: 64 concurrent Add RPCs per burst, reported per RPC; 8192 bursts per iteration, six warmups, 12 measured iterations, outliers retained, DOTNET_PROCESSOR_COUNT=8, the same Apple M4/.NET 10.0.11/Server GC environment. Timed mean is 5.723 ± 0.1113 us before versus 1.860 ± 0.0364 us after (99.9% confidence half-widths); allocation is 1118 versus 452 B/RPC, completed work items 3.5644 versus 1.3167. Untimed control is 1.729 ± 0.0355 versus 1.726 ± 0.0228 us, with 284 B/RPC in both. Measured lock-contention counters round to zero in all four cases.

The separate four-core server + VM/Mac diagnostic compares 6a22b42 to 0141a85, with a fresh server per case/phase. High-concurrency Add: A1 231974, B 285698, A2 218950 completed/s; B server CPU 2.48 cores versus A approximately 3.66. High Echo(256): 148190 / 164251 / 152175 completed/s, with B p99 2095 us versus A 2347/2353 us. Add p99 was higher (2895 versus 1179/998 us) and A drift was 5.6%; final-main validation must review those explicitly.

These results validate the two focused corrections. Release publication remains blocked on final-main validation and the separate low-concurrency OneWay/upload investigation against published 1.1.1. The release tag has not been created. Final performance validation will use the resulting exact main commit and its gated packages.