Fix issues found in code review

[BobsBlazed]

Eight findings from a review of 1.2.0. None changes behaviour for a correctly-paired LoRA and model — the first two matter when the pairing is wrong.

1. The dot-boundary fix never reached the FLUX/SDXL classifiers

The Universal loader matches block stacks on the raw dotted key, because normalising . → _ erases the separator that distinguishes an outermost stack from one nested inside a block. The two dedicated classifiers never got that fix, so an SDXL key:

input_blocks.4.1.transformer_blocks.0.attn1.to_q.weight

matched the inner transformer_blocks and was read as double-stream block 0.

Anchoring on . alone turned out to be insufficient — the nested stack is also dot-preceded. The patterns are now anchored to the start of the module path after an optional known prefix, which is the property that actually distinguishes them. Verified against 782 authentic FLUX and 1696 authentic SDXL keys: unchanged.

keys → node	before	after
SDXL → FLUX	all 1696 in Early Downsampling	1674 Other Tensors, 22 Final Output Layer
FLUX → SDXL	all 782 in Other Tensors	unchanged

2. Wrong loader for a model now warns instead of failing silently

Picking the wrong node was invisible: the keys still match — ComfyUI builds the key map from the model, so the LoRA loads — they just all land in one bucket, and the sliders quietly stop meaning what they say.

A mismatch check now fires when ≥90% of UNet tensors land in a single block, on both the info output and the log. A threshold rather than "all of them", because cross-family names overlap slightly: real SDXL keys through the FLUX classifier leave 98.7% in the catch-all and send the rest to the shared proj_out token, which exact equality would wave through.

Measured against real key sets:

MISMATCHED (must warn)
  SDXL keys -> FLUX node : WARNING: 99% of UNet tensors landed in 'Other Tensors'...
  FLUX keys -> SDXL node : WARNING: 100% of UNet tensors landed in 'Other Tensors'...
CORRECT (must stay silent)
  FLUX->FLUX, SDXL->SDXL, SD15->SDXL, and SDXL/SD15/FLUX->UNIVERSAL : all None

3. Import-order dependency in the shared registries

bobs_universal mutated four of bobs_blocks's dicts at import time. Confirmed consequence: after that import, this line in tests/test_bobs_blocks.py raises KeyError: 'UNIVERSAL':

blocks = {"FLUX": ALL_FLUX_BLOCKS, "SDXL": ALL_SDXL_BLOCKS}[family]

It passed only because the two test files happened to load bobs_blocks under different module names, so they got separate instances — luck, not design. Families now register through an explicit register_family(), and the test drives off ALL_BLOCKS.

4–8. Smaller items

• Per-family tooltips. FLUX/SDXL and Universal each define a block named "Text Encoder"; BLOCK_TOOLTIPS.update() meant the Universal wording overwrote the other. Tooltips are now stored per family.
• Output Head over-matching. The _head_ substring also caught multi_head_attention and head_dim_proj. Now anchored to a real head module (head, head.head, out.<n>).
• Unknown preset names log a warning instead of silently behaving as Custom (only reachable from hand-edited workflow JSON).
• FLUX negative block indices accepted, matching the Universal loader's handling of SD3's joint_blocks.-1.
• Dropped a 1.01 bucket boundary guarding a fraction that cannot occur — depth_fraction maxes at (n-0.5)/n.
• Bonus, found while writing the tests: _mismatch_warning was keying off len(grouped), which is dense in production but sparse in a unit test. It now uses its blocks argument, which was otherwise unused.

Testing

71 tests, up from 55, still with no ComfyUI or torch dependency. New tests/test_node_behaviour.py covers node registration, widget order, output signature, report formatting and the mismatch warning — including the exact 1674/22 distribution from finding 1, so the threshold can't silently regress to exact-equality.

The real-ComfyUI integration harnesses were re-run: 8,000+ authentic keys across 11 architectures, still nothing falling through to Other Tensors, and the node end-to-end run through a real ModelPatcher is byte-identical (block isolation still exactly 384/1933).

pyproject.toml bumped to 1.2.1.

---

Generated by Claude Code