- MemoryCurationActor: replace Protocol.SessionId qualifier with a
  using alias matching the SessionLlmInvoker.cs:9 pattern. A full
  Netclaw.Actors.Protocol namespace import would have introduced a
  ChatRole ambiguity (Protocol.ChatRole vs Microsoft.Extensions.AI.ChatRole)
  that the bare alias avoids cleanly.
- SidecarDiagnosticsContextTests: drop two defensive
  SessionDiagnosticsContext.SessionId = null pre-test resets. xUnit
  creates a fresh class instance per test method and AsyncLocal is
  rooted in the per-task ExecutionContext; there's no shared state to
  reset. The Assert.Null at the end of each test still verifies the
  using-scope restored properly.
- SubAgentActor: tighten the diagnostics-scope comment from 4 lines
  to 3, dropping the redundant "shares parent's session.log" framing
  that's now covered by the SessionDiagnosticsContext doc.

Three follow-up fixes from the simplify-pass review on the sidecar
diagnostics-context PR. Each addresses something I had previously left
on the table:

1. Stringly-typed sessionId on SubAgentActor.InvokeLlmAsync replaced
   with `SessionId? sessionId`. The call site converts the upstream
   `_toolExecutionContext.SessionId` (which is still typed as string?
   in Netclaw.Tools.Abstractions; out of scope to retype here)
   into the value object once at the boundary.

2. Positional ordering on InvokeLlmAsync normalized: ct now last,
   sessionId after options. The previous "ct between options and self"
   placement was inherited from the original signature and was awkward
   for the new sessionId parameter; the conventional order makes the
   private API harder to mis-call from a future test.

3. Test coverage gap closed. All seven sidecar IChatClient call sites
   now have direct unit tests in SidecarDiagnosticsContextTests:
   - SessionTitleGenerator.GenerateAsync
   - SessionCompactionPipeline.GenerateObservationsAsync
   - LlmSessionActor.InvokeMemoryExtractionCoreAsync (now internal static)
   - SubAgentActor.InvokeLlmAsync (now internal static; null-sessionId
     case also covered)
   - SessionMemoryObserverActor.RunDistillationAsync (now internal
     static; was private instance — extracted to match the existing
     SessionTitleGenerator/SessionCompactionPipeline pattern, _log
     passed as ILoggingAdapter parameter)
   - MemoryCurationActor.TryLlmEvaluationAsync (now internal static;
     same shape — _llmClient, _sessionId, _log all passed as params)

The static extractions for RunDistillationAsync and TryLlmEvaluationAsync
are behavior-preserving refactors that follow the established pattern
of the other sidecar helpers in the same namespace. They make these
methods unit-testable in isolation without spinning up the parent
actor or driving the message protocol.

This closes the coverage gap I previously deferred to the planned
Roslyn analyzer (netclaw-dev#915). Now we don't depend on netclaw-dev#915 landing — the
contract is verified at the unit-test level for every call site.