Resolve direct string constants before execution and retain requested seeds
for list-producing inputs. Preserve native widget values and scalar reuse.
Cover cold queues, mapped lists and Control combinations with actual core
cache checks and Chrome tests. Limit the Bypass xfail to the known Processor
case so Encode failures remain visible.
Prompt validation runs after the hook and rewrites FLOAT values sent as
whole numbers and __value__ list widgets, so upstream chains containing
them never matched. The fingerprint now applies the same normalization.
Only [node_id, slot] pairs are followed as links; list widget values
crashed the hook.
Each wildcard node keeps only its last static run. ComfyUI's default
and classic caches drop results of earlier prompts, so older records
were never hit. A run that receives syntax clears the record.
Linked populated_text keeps its seed at submission, since its runtime
value is unknown there. When a finished run receives text without
wildcard syntax, the node records that run's seed for the same upstream
inputs. Later submissions with identical upstream inputs reuse it, so
the wildcard node and downstream nodes stay cached.
Upstream nodes that declare IS_CHANGED/fingerprint_inputs or are not
registered are never recorded, matching ComfyUI's own cache assumption.
Clamp finite mask values to [0, 1] before uint8 contour conversion while
preserving input arithmetic dtype, batch handling and combined behavior.
Add CPU and native-workflow regression coverage for range boundaries,
tutorial mask behavior, widget serialization and fresh browser execution.
Document the isolated test environment and regression verification scope.
Preserve saved selections while trimming unused Switch inputs, including empty selected slots. Include native workflow regression tests and their fixtures.
Validation: 8 focused regressions and both catalog restoration cases pass on the final Main integration; 40 expanded E2E cases passed on the revised Switch logic. The original patch fails the same 4 focused cases in both Red runs.
Preserve earlier canvas background callbacks when Mask Rect preview widgets are installed. Reviewed with four hook tests, the native widget catalog in both restoration modes, and non-default Mask Rect export/reload checks.
Include the original LoRA wildcard source together with its regression fixes in one Main update. The merge tree matches the pre-merge validated tree aee243383b.
Keep original automatic-mode candidates and empty selections, retain entry punctuation, and reject cross-drive paths without breaking category discovery.
Keep wildcard sampling on its per-call NumPy generator and remove Python RNG seeding and state restoration. Preserve the negative seed normalization from #1224.
Add 19 seed regressions covering fixed outputs, signed seeds, shared RNG preservation on success and failure, and concurrent consumers. All 117 standalone tests pass; the two concurrency cases fail again when the PR implementation is restored.
Remove the obsolete-socket fixture and its two compatibility cases while retaining fresh hook execution and native reload checks in both restoration modes.
Validation: focused catalog and hook E2E tests pass (3 tests); all 14 remaining catalog tests collect successfully.
Dr.Lt.Data