Mirrors the .NET reference's `if (!subscription.IsBuffered)` guard
at `MxNativeSession.cs:361-381`. The Rust port previously emitted an
`UnAdvise` frame for both plain and buffered subscriptions; the
buffered server-side registration is unwound by the engine when the
`RegisterReference` handle goes away, so emitting an `UnAdvise` for
buffered entries is at best a no-op extra frame and at worst could
race with the engine's own teardown.
Fix: branch `Session::unsubscribe` on `SubscriptionEntry::mode` (the
discriminator F45 added). For `SubscriptionMode::Buffered { ... }`,
skip the `un_advise` call and proceed directly to registry cleanup.
For `SubscriptionMode::Plain`, retain the previous behaviour.
The registry-entry probe runs first (separate lock acquisition) so
the `is_buffered` decision doesn't hold the NMX-client mutex
unnecessarily — common case where the entry is plain still acquires
the NMX lock immediately after.
The metrics counter `record_unadvise()` still fires on every public
`unsubscribe` call regardless of mode — it tracks consumer-side
unsubscribe rate, not wire-frame rate. That matches what dashboards
expect from the public API.
New unit test `unsubscribe_skips_un_advise_for_buffered_subscription`
issues a plain subscribe (recorded as 1 RPC), mutates the registry
entry to `SubscriptionMode::Buffered`, calls unsubscribe, and
asserts the recorded RPC count stays at 1 (no UnAdvise emitted).
The existing `subscribe_populates_registry_unsubscribe_clears_it`
test serves as the negative control for the plain branch.
Workspace 794 → 795 tests; clippy clean; rustdoc clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Joseph Doherty
1a1830f3bf