Release 2.0.4-beta.3: Dependency compatibility + Documentation

Bugfixes and compatibility improvements. No new features.

Core fixes:
- Framework detection for web API models (Issue #48)
- Video-only model filtering from vision capability
- Page size detection for memory metrics (macOS)
- Model switch log timing (after load completion)

Compatibility:
- hub 1.x + transformers 5.0 support
- Python 3.9-3.14 verified (494 tests passing)

Testing infrastructure:
- Benchmark schema v0.2.0 (hardware profiling, system health)
- Benchmark template v1.0 (automated JSONL→Markdown reports)
- Memory timeline visualization (memplot.py)
- Unified model filter (build_model_object single source)

Documentation:
- Multi-Modal Support section in README (Vision subsection)
- JSON API 0.1.5-0.1.6 marked Stable
- Vision promoted from alpha to beta status
- Removed conceptual drift and outdated references

See CHANGELOG.md for complete details.

TESTING.md

@@ -19,6 +19,17 @@ For current test counts, version-specific details, and complete file listings, s
 - Delete operations fail if not in test cache (`MLXK2_STRICT_TEST_DELETE=1`)
 - Live tests never modify user cache without explicit environment variables
 
+**Unit Test Limitations:**
+
+MLX Knife has two test categories:
+1. **Unit tests** (~500 tests, fast, mocked) - verify code structure
+2. **Live E2E tests** (real models, slow) - verify actual functionality
+
+**Why both are needed:**
+When dependencies like `transformers` or `mlx-lm` update their APIs, unit tests (which mock these libraries) continue to pass, but real model loading breaks. Only live E2E tests catch these issues.
+
+**Example:** transformers 5.0 changed tokenizer initialization - unit tests passed (mocked API), but vision models failed to load in production. Live E2E tests caught the issue immediately.
+
 ## Quick Start
 
 ```bash