Fix CI-breaking registry and diffusion test regressions

[christopherkarani]

Summary

This mission-critical audit run found and fixed deterministic CI blockers introduced by API/catalog evolution.

Issues fixed

1. Compile-gate mismatch in diffusion tests (hard CI failure on non-MLX environments)

   • Tests/ConduitTests/ImageGeneration/DiffusionModelDownloaderTests.swift compiled with #if canImport(Hub) only.
   • Production symbol DiffusionModelDownloader is gated by CONDUIT_TRAIT_MLX && canImport(MLX) && (canImport(Hub) || canImport(HuggingFace)).
   • Result: tests attempted to compile against symbols excluded from the build graph.
   • Fix: mirror production gate exactly in test file.

2. Stale model catalog assertions after provider/model expansion (CI failures)

   • Tests/ConduitTests/Core/ModelIdentifierTests.swift assumed old registry totals/provider distributions.
   • Catalog now includes additional cloud models (kimi) and changed capability distribution.
   • Fix: update provider-specific expectations to current registry shape and strengthen capability/provider invariants.

3. Outdated ProviderType case-count expectation (CI failure)

   • Tests/ConduitTests/Core/ProtocolCompilationTests.swift expected 10 provider cases.
   • Current API surface has 12 (kimi, minimax added).
   • Fix: update expected count and explicit containment checks.

Why this matters

• Restores full testability in environments where MLX traits are disabled.
• Prevents false-negative CI due to stale hard-coded counts.
• Keeps protocol/API conformance tests aligned with the actual public enum surface.

Verification

• swift test --filter DiffusionModelDownloaderTests
• swift test --filter ModelIdentifierTests --filter ProtocolCompilationTests
• swift build
• swift test

All commands pass in this workspace.

Scope

• No production runtime behavior changed.
• Changes are test-only and targeted to correctness of test gating and expectations.