Binding a call is a different operation from looking up a name, and that is where the rule got expensive. Presence checking builds the set of names bound at module scope from the AST and diffs it against what the pack references [3]. Call binding parses the upstream module at whatever git ref you care about, reads the real parameter list, and tries to match every call site the pack makes against it; a call that cannot bind raises TypeError the moment it runs [6].
The first failure is scope. A pack imports `calculate_weight` from `comfy.lora` and then defines its own `calculate_weight` in the same file, so the bare call resolves to the local one. The rule resolved it against upstream and reported a hard break on working code, and any rebinding does the same thing: a local def, a function parameter, a loop target, a walrus, a later import of the same name [8]. The patch is a conservative file-wide shadow set [9].
The design decision underneath is the part worth stealing. The author deliberately did not tighten the import checks to match, because a presence-check false positive is a warning somebody glances at and dismisses, while a call-binding false positive is a red build; same codebase, two tolerances, picked by what the wrong answer costs [10].
The second failure is worse, because it selects for care. One pack sets a flag from `hasattr(comfy.model_prefetch, "GRAPH_MODULES")` and calls `cleanup_prefetched_modules` with the arity that matches your installed version. At any given ref one branch binds and the other is dead code, and the rule read the dead branch and graded the pack WILL BREAK [11]. Dead code cannot raise the TypeError the rule was measuring. The author's line is that a lint which taxes the careful and ignores the careless is worse than no lint [12]. The heuristic that fixes it: if a sibling call to the same function in the same file binds, treat the failing one as a shim, report it, never fail the build, and let `except TypeError` soften a call the way `except ImportError` already softens an import [13].
The validation run was 20 cloned packs, 1,273 Python files, 1,116 call sites into `comfy.*`, 10 monkeypatches, every hit hand-verified against pack source and ComfyUI's git history [7]. Two findings survived, both true, with 18 of 20 packs silent [14]. That is roughly 0.18 percent of call sites, about one per 558 [17], and 10 percent of the corpus flagged [18]. For that rate to transfer you would need your pack mix and your pinned upstream ref to look like this sample; against an older pin the same rule reports different code.
The interesting finding is in ComfyUI-Easy-Use, at 2.7k stars, whose BrushNet path passes `scaled_fp8` to `comfy.ops.pick_operations`. That parameter was removed in commit 43071e3de, PR #11000, in December 2025; v0.3.77 is the last release that accepts it and v0.4.0 the first that does not [15]. The post then stops mid-sentence, at the moment the author goes to file the bug and finds an existing issue [16], so the public record does not say whether the tool found anything the tracker did not already know.
