feat(ai): semantic-search backfill worker + Ollama compose profile (RECOMMENDATIONS #4a)

First activation slice of semantic search (docs/discovery/05+06):
- EmbeddingBackfillWorker: fills Email.Embedding in small batches (32, 2s pause,
  newest first; subject+snippet, truncated) so interactive load is never starved.
  Exits immediately when the embedding provider is unavailable — AI-off deployments
  are untouched, Ollama hiccups back off instead of crashing the host.
- compose: optional 'ollama' service under the 'ai' profile (persistent model
  volume, commented GPU passthrough for the RTX 3080); api gets
  Ai__OllamaBaseUrl pointing at it.
- LiveDb test proves the worker persists 768-dim vectors against real pgvector
  (deterministic fake provider); LiveDb classes serialised into one xUnit
  collection (concurrent MigrateAsync on a fresh DB races — found while testing).

Verified against REAL Ollama locally: pulled nomic-embed-text in a container and
confirmed the /api/embeddings contract the provider uses returns 768-dim vectors.
Full suite 55 green (4 LiveDb vs real pgvector); format clean.

Next slice: hybrid RRF ranking in SearchService once embeddings exist.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/InboxIntel.Infrastructure/InboxIntel.Infrastructure.csproj

@@ -32,4 +32,7 @@
     <ProjectReference Include="..\InboxIntel.Application\InboxIntel.Application.csproj" />
     <ProjectReference Include="..\InboxIntel.Domain\InboxIntel.Domain.csproj" />
   </ItemGroup>
+  <ItemGroup>
+    <InternalsVisibleTo Include="InboxIntel.IntegrationTests" />
+  </ItemGroup>
 </Project>