[SunSi12138]

Release validation exposed three correctness gaps: dynamic-module tests could request a proxy before the remote manifest arrived, a delayed send-pump wake could flush a timed batch early, and Windows shared-memory shutdown could exhaust its cleanup budget while disposing a mapped view.

The changes wait for manifest readiness, recheck the queue and original flush deadline after a wake, and close transient mapping views without the accessor’s unnecessary explicit file flush. The Windows failure-dump helper now captures the failing process with the installed diagnostic tool and waits for capture before final cleanup.

Validation:

• Full local unit suite: Passed: 1819, Failed: 0, Skipped: 0. Full local integration suite: Passed: 451, Failed: 0, Skipped: 0.
• Two deterministic timed-batching cases fail with the old send-pump logic and pass with the correction.
• Two mapping lifecycle cases cover closing either peer first, preserved shared bytes, invalidated closed-side memory, idempotent cleanup, and final file removal.
• Windows diagnostic comparison: baseline recorded 1,270 mapped-view disposals over 100 ms, up to 3,469 ms. Corrected disposal passed three 300-second restart workloads with zero measured cleanup stages over 100 ms, zero unexpected failures, and zero unobserved task exceptions. Diagnostic instrumentation remains on its separate branch.
• The accessor publication/stop stress test retains all 100,000 rounds and assertions, with a two-minute worker budget so hosted macOS scheduling does not cancel the test at its previous 30-second bound. Focused verification: Passed: 1, Failed: 0, Skipped: 0.
• Focused whitespace validation passed. The updated candidate is running Release Gate before merge.

The change preserves the existing cleanup budget and still joins and releases owned resources. It adds no per-RPC state or allocation. Idle send-pump BenchmarkDotNet results remain 2.589 → 2.579 µs and 536 B/op; the manual-clock timed-batch measurement was inconclusive and is not claimed as a pass. The final-main package ABA comparison remains a separate prerequisite for publication.