fix: release pending stream ownership after local receive-buffer exhaustion

[SunSi12138]

Summary

Fix locally exhausted stream receive buffers leaving live pending-call ownership after the consumer exits. This targets dev only; no merge, main cherry-pick, or 2.0.4 publication is part of this PR.

Root cause

In PooledAsyncStreamDispatcher<T>.DispatchAcquiredAsync, exceeding the 4096-element receive buffer used Complete(ResourceExhausted). That selected the same consumer terminal state as a genuine peer StreamComplete. DisposeAsync therefore skipped ConsumerAbandonedCallback. Closing a stream dispatch state only stops future dispatch; it does not independently remove the pending call and StreamManager route. Standalone RPC cancellation, deadlines, and connection-close cleanup retain their existing independent paths.

Fix

• Distinguish local receive-buffer failure from genuine peer completion with an explicit terminal value, without changing the 4096-item cap or SharpLinkErrorCode.ResourceExhausted.
• Preserve the existing completion-publication protocol (issue test: intermittent full-suite stall on many-core hosts (TUnit engine stops scheduling ~70 tests) #206 double-checked full-fence synchronization), first/second dispatch-state close ordering, active-dispatch drain, discarded-item receive-credit return, and one-shot dispose completion.
• After local receive failure, consumer disposal invokes the existing abandonment callback once, following dispatch drain and credit return; ClientConnection then removes the pending slot, unregisters the inbound route, and emits ConsumerAbandoned Cancel for a published request. True peer completions still skip this callback.
• Keep PooledAsyncStreamDispatcher.cs within the repository's existing source LOC baseline by extracting terminal-publication code into PooledAsyncStreamDispatcher.Terminal.cs (partial class), rather than raising thresholds.

Deterministic regression evidence (test-first)

• RED: test-only commit 59ab11fd4cce53e8b8cf2e7d77bdebf7f3063d69 against dev base f0fee51e77a5a736469bffa837ceb8f31b1f6877: PR Fast run 38157457266 compiled, then failed exactly 3 new regression tests (2060 passed, 3 failed of 2063): actual pending/route leak, missing local-failure callback after ResourceExhausted, and local-completion/Dispose publication race.
• Candidate: 493b9db62db7352ee35b379fde9f56fd2abb6687 (see current PR checks). Four focused cases cover early exit, actual pending/route cleanup and published-request Cancel, full credit return and ResourceExhausted, concurrent/duplicate Dispose with async callback, and real peer completion as a negative control. The published-frame case asserts credit-before-Cancel, without incorrectly imposing Request-before-WindowUpdate ordering: the send pump may prioritize WindowUpdate before Request, provided both precede Cancel.
• Broader validation: PR Fast runs formatting, maintainability, Release build, all unit/generator/load tests, and allocation gate; PR Extended includes full integration (including generated ABI dynamic-module stream lifecycle exit modes), AOT/package/load and codec compatibility. Treat the linked checks as authoritative. A multi-hour generated-ABI soak is separate from these PR gates and is not claimed as performed here.

Review / sequencing

1. Review and validate this PR against dev. Do not merge automatically.
2. After explicit approval, prepare a separate targeted main patch PR for the not-yet-published 2.0.4 release. PR chore(release): promote reviewed dev fixes for 2.0.4 #804 is already merged and is not this fix's feature branch.
3. Prior dump/ABI lifecycle context: Actions run 38148482335 (a dump reader job failed; not evidence of successful ABI validation).

[SunSi12138]

Verification result — head eb9f20cc18c8dec88b8d0e9d5f1a74af5bef7657

Current dev: the fix branch is synchronized with dev aaedfc3f80e8fc00c504a2018577218f224ec3c3 through a two-parent branch-only merge; the PR diff still contains exactly the three fix/test files. All checks below were rerun on that synchronized head and passed. The PR is open, not merged, with base dev.

Test-first RED: PR Fast 38157457266 on test-only 59ab11fd4cce53e8b8cf2e7d77bdebf7f3063d69, before the implementation fix: 2060/2063 unit tests passed, and exactly three added bug-focused tests failed (pending/route ownership, local ResourceExhausted abandonment, and local completion versus Dispose).

Fixed GREEN:

• PR Fast 38159338206: 2063/2063 unit, 271/271 generator, 77/77 load-test cases passed; Release build, formatting, maintainability, reference-boundary and allocation checks passed.
• PR Extended 38159360891: 454/454 integration cases passed (the full integration suite includes the generated API4 collectible dynamic-module streaming exit lifecycle cases for normal completion, cancellation before first item, cancellation mid-stream, consumer break, and service exception); Debug admission race, Release build, AOT transport/topology smoke, NuGet packaging and smoke, one-way demo, load smoke, and cross-platform codec compatibility all passed.
• PR Package Smoke 38159338213: trim, isolated NativeAOT, package contract and NuGet smoke passed.
• CodeQL 38159338191 and extension fault containment 38159338217: passed.

The published-request regression confirmed the outbound pump may prioritize WindowUpdate ahead of Request; the required invariant tested here is that both precede the single ConsumerAbandoned Cancel. No thresholds, deadlines, buffer caps or assertions were relaxed to suppress the original bug.

Scope: no main patch and no merge. A separate multi-hour generated-ABI soak was not executed on this PR; the generated-ABI dynamic-module lifecycle integration tests above were executed and passed. This PR is ready for human review.