feat(ai): IEmbeddingProvider foundation (Null + Ollama)

AI-foundation slice 1 (per docs/discovery/06). Adds the embeddings abstraction that
semantic search / near-dup / find-similar will build on: IEmbeddingProvider with a
NullEmbeddingProvider (AI-off: IsAvailable=false, empty vectors) and an
OllamaEmbeddingProvider (/api/embeddings, nomic-embed-text). DI registers embeddings
by AI mode (Ollama local, else Null). Kept separate from IAiProvider since embeddings
are a distinct model/capability. AI remains fully optional — callers fall back to
lexical search when unavailable.

Buildable/testable without a live Ollama; unit test locks the Null fallback contract.
Build + all 41 tests pass.

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

tests/InboxIntel.UnitTests/EmbeddingProviderTests.cs

@@ -0,0 +1,23 @@
+using FluentAssertions;
+using InboxIntel.Infrastructure.Ai;
+using Xunit;
+
+namespace InboxIntel.UnitTests;
+
+/// <summary>
+/// The Null embedding provider is the AI-off / no-model path. Semantic features must be able
+/// to detect unavailability (IsAvailable=false) and fall back to lexical search — this locks
+/// that contract so "AI is never required" can't silently regress.
+/// </summary>
+public class EmbeddingProviderTests
+{
+    [Fact]
+    public async Task Null_provider_reports_unavailable_and_returns_empty_vectors()
+    {
+        var sut = new NullEmbeddingProvider();
+
+        sut.IsAvailable.Should().BeFalse();
+        (await sut.EmbedAsync("anything")).Should().BeEmpty();
+        (await sut.EmbedBatchAsync(new[] { "a", "b" })).Should().BeEmpty();
+    }
+}