Update the README and add architecture documentation

[BobsBlazed]

Documentation only — no code changes.

Why

The README had drifted behind three releases. It described the project as two nodes for SDXL and FLUX, told readers to "choose either SDXL or FLUX", showed an info example predating the architecture line added in 1.2.0, and stated the wrong licence (GPL-3.0; the repo is Apache-2.0).

README

• Leads with the three nodes and a table of when to use each, instead of bolting the Universal loader onto a two-node description. Previously the Universal bullet read "A third loader covers everything else" while the line above still said "Dual Model Support".
• Refreshed info example — a real one, including the architecture: QwenImage transformer_blocks[60] line, with every column explained.
• New Troubleshooting section covering what the node can now actually report: nodes missing from the menu, the mismatch warning, a LoRA matching nothing, a high Other Tensors count, and divergence from the built-in LoraLoader.
• Notes the SDXL node also covers SD1.5 and SD2 — true since the classifier is UNet-shaped, but never stated.
• Separates verified from expected-to-work. The 1.2.0 notes implied one list of supported architectures. SD1.5, SDXL, SD3, FLUX (full and pruned), AuraFlow, PixArt, LTX-Video, Lumina, Qwen-Image and Wan were exercised against real ComfyUI key sets; HiDream, Chroma, Mochi, HunyuanVideo and Cosmos were not, and are now labelled as such rather than listed alongside.
• Licence corrected to Apache-2.0.
• Changelog moved to CHANGELOG.md — it was 90 of the README's 250 lines and growing one section per release.

docs/ARCHITECTURE.md (new)

The design knowledge behind the current implementation, so it doesn't have to be re-derived:

• Why classification runs on the canonical target key rather than the LoRA's own names, and the two consequences — (key, offset) tuples for fused weights (534 of them in a real FLUX key map), and model/CLIP routing without name guessing.
• Why stack matching uses the raw dotted key and not the normalised one, with the SDXL nesting case that motivated it, plus why the FLUX/SDXL classifiers additionally anchor to the start of the path.
• How runtime stack discovery and the depth axis work, including negative-index resolution for SD3's joint_blocks.-1.
• How to add an architecture or a whole family — and the widget-order compatibility contract, which is the easiest thing here to break silently.

Verification

Every factual claim in both documents was checked against the code rather than written from memory: both FLUX block-range tables, the Universal and SDXL block lists, the preset names, all documented stack names against STACK_ORDER, the test count, and all internal links and anchors.

OK  README FLUX double_blocks ranges match code
OK  README FLUX single_blocks ranges match code
OK  Universal block table matches code
OK  all documented stack names present in STACK_ORDER
OK  all internal links and anchors resolve

That check is what caught the licence error and the over-broad architecture list.

Note on versioning

No version bump — publish.yml only fires on pyproject.toml changes, so this won't republish to the registry. The registry keeps 1.2.1's README snapshot until the next code release. Happy to bump to 1.2.2 if you'd rather the registry listing update now.

---

Generated by Claude Code