[F49 steps 2 + 3] live verification: buffered recovery replay + unsubscribe skip

Step 3 (F47 buffered unsubscribe skip):
- crates/mxaccess-compat/tests/buffered_unsubscribe_skip_live.rs.
- Subscribe buffered, sleep so the engine has DataUpdates in flight,
  then call unsubscribe. Asserts Ok return without surfacing transport
  or HRESULT errors.
- Session::unsubscribe (session.rs:2261) probes the registry: if
  Buffered { .. }, it skips nmx.un_advise entirely, mirroring the .NET
  reference's `if (!subscription.IsBuffered)` guard at
  MxNativeSession.cs:361-381. If unsubscribe accidentally emitted
  UnAdvise for a buffered correlation id, the engine would return
  non-zero HRESULT (no matching plain advise to retract) — surfacing
  as a panic.

Step 2 (F45 buffered recovery replay):
- crates/mxaccess-compat/tests/buffered_recovery_replay_live.rs.
- Subscribe buffered, drain >=1 NMX subscription message
  (cmd=0x32 SubscriptionStatus + cmd=0x33 DataUpdate) to confirm the
  wire path is hot pre-recovery, install a RebuildFactory that calls
  NmxClient::create (the same auto-resolving COM-activation path
  Session::connect_nmx_auto uses), invoke recover_connection, drain
  >=1 NMX subscription message post-recovery.
- Verifies the replay branch in recover_connection_core re-issues
  RegisterReference (NOT AdviseSupervisory) for the buffered entry,
  mirroring MxNativeSession.ReAdviseSubscription (cs:538-569).
  Structural property is unit-tested; this confirms the engine
  actually picks back up after the rebuild + replay.

Both tests pass live on this Galaxy:
  cargo test -p mxaccess-compat --features live-windows-com \
      --test buffered_unsubscribe_skip_live -- --ignored --nocapture
  cargo test -p mxaccess-compat --features live-windows-com \
      --test buffered_recovery_replay_live -- --ignored --nocapture

Pulls mxaccess-nmx + mxaccess-codec into mxaccess-compat dev-deps so
the recovery test can build a RebuildFactory closure that returns
NmxClient and bind a typed broadcast Receiver.

design/followups.md F49 -> Resolved (all five steps pass live).
docs/M6-live-verification.md updated with per-step evidence + repro
commands.

F49 is fully closed out. F55 (DCOM-managed INmxSvcCallback, Path A)
and F56 (missing EnsurePublisherConnected + post-RegisterReference
AdviseSupervisory for buffered) were the two real Rust-port bugs
uncovered along the way; both resolved. Remaining post-V1 followups
(F50 Suspend/Activate Frida, F51 ASB type matrix, F52 perf, F53 doc
lint, etc.) are scoped independently and not part of F49.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

design/followups.md

@@ -26,7 +26,14 @@ Between each publish: wait for the crate to be indexed before the next one's `ca
 **Resolves when:** crates.io shows all 9 crates published + the V1 tag is pushed.
 
 ### F49 — Live verification sweep for the M6 features
-**Status:** Steps 1, 4, 5 resolved 2026-05-06 (`docs/M6-live-verification.md`). F56 turned out to be a real Rust-port bug (missing `EnsurePublisherConnected` RPC pair) and was fixed; both `subscribe` and `subscribe_buffered` now drive `0x33` DataUpdate frames end-to-end against `TestMachine_001.TestChangingInt`. Steps 2 (F45 recovery replay live) and 3 (F47 buffered unsubscribe skip live) remain — they're now executable on this fixture but not yet run.
+**Status:** **Resolved 2026-05-06.** All five steps pass live against the local AVEVA install (`docs/M6-live-verification.md`):
+- Step 1 (F36 buffered subscribe) — `tests/buffered_subscribe_live.rs`
+- Step 2 (F45 buffered recovery replay) — `tests/buffered_recovery_replay_live.rs`
+- Step 3 (F47 buffered unsubscribe skip) — `tests/buffered_unsubscribe_skip_live.rs`
+- Step 4 (F40 metrics smoke) — `tests/metrics_smoke_live.rs`
+- Step 5 (F54 OnWriteComplete) — `tests/lmx_write_complete_live.rs`
+
+F55 (DCOM-managed `INmxSvcCallback`) and F56 (missing `EnsurePublisherConnected` + AdviseSupervisory after RegisterReference for buffered) were the two real Rust-port bugs uncovered along the way; both are resolved.
 **Severity:** P1 — closes the live-evidence gap for the M6 work that landed unit-only this session.
 **Source:** F36, F40, F45, F47, F54 closeouts — each ships with unit tests but most were not exercised against the live AVEVA install in this session.
 **Blocked-by:** F12 hardening (`Session::connect_nmx_auto` returns `RPC_S_SERVER_UNAVAILABLE` (1722) under `cargo test`'s tokio multi-thread runtime — see "Live attempt 2026-05-06" below). The COM-activation path itself works in isolation (`cargo run -p mxaccess-rpc --example com-marshal-probe --features windows-com` succeeds), so the failure is downstream — likely a COM apartment threading issue when CoInitializeEx runs on a tokio worker thread.