[SunSi12138]

Problem and change

Calls with no input streams entered a reservation method whose captured lambdas caused the compiler to allocate callback state before the existing zero-stream early return. Allocation tracing of the release's 64-message OneWay workload identified SharpLinkServer.<>c__DisplayClass348_0 from ReservePreInvocationRequestStreams as an approximately 64-byte per-message allocation.

Move captured callback construction into private streaming-only helpers, after the existing zero-stream guards. Apply the same guard boundary to pre-admission reservation. Input-stream reservation, compression decoding, retained-byte limits and terminal cleanup retain their existing implementation.

Validation

• Full non-incremental Release solution build: passed, 0 warnings / 0 errors.
• Targeted pre-invocation retention, pre-admission stream-budget and OneWay deadline tests: 7 passed / 0 failed / 0 skipped.
• Full unit tests: 1825 passed / 0 failed / 0 skipped. Integration tests: 451 passed / 0 failed / 0 skipped.
• All eight NuGet packages packed successfully and verified against commit f61e1642316d9c14f48f477f97c82cc95bb3f909.
• BenchmarkDotNet 0.15.8: identical loopback workloads against frozen 0141a85 and f61e164 source builds; .NET 10.0.11 / macOS arm64 / M4 / server GC / 8 reported processors. Four dry cases passed; measurements used 8192 invocations, 6 warmup iterations, 12 measured iterations, unroll 1, with outliers retained.

Workload	Before	After	Allocation
64-message OneWay batch with server completion barrier, normalized per message	1.206 ± 0.0270 μs	1.200 ± 0.0223 μs	219 → 155 B/message
Client upload, eight 4 KiB items	94.91 ± 0.991 μs	95.13 ± 1.076 μs	36.40 → 36.41 KB/upload

Errors are BenchmarkDotNet 99.9% confidence-interval half-widths. The OneWay result removes exactly 64 B/message (29% of the measured allocation) while time remains statistically unchanged. Upload time also remains unchanged within uncertainty. Timing measures client and server together over loopback; it does not replace final-main hardware ABA.

This addresses an independently measured allocation issue. It does not claim to resolve the separate baremetal TCP delivery stalls seen in release ABA diagnostics.