ponder-lab / Hybridize-Functions-Refactoring / #2673

84%

main: 84%

push

github

Model a name-referenced `input_signature` via its sole module-level binding (#835)

* Model a name-referenced input_signature via its sole module-level binding

A supplied input_signature was modeled only when the decorator argument
was a literal list/tuple; a bare name referencing a module-level
constant (the gpt-2 train_step_signature shape) dropped to UNMODELED.
Resolve such a name to the module-level assignment's literal, but only
when the identifier is bound exactly once in the entire containing
module: under that sole-binding rule, Python's scoping cannot bind the
decorator expression to anything else. Attribute stores and call-site
keyword names are not bindings and do not compete; reassignments,
shadowing definitions, imports, loop targets, deletions, and parameters
all decline. The retained node stays the one at the decorator site, and
reconfigure()'s overwrite path gains a Name branch replacing exactly
the reference's span (the bracket scan would have run past the
decorator), leaving the shared constant intact for its other users.

Closes #834.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GcqABsog1rtb4m2CDZdcRD

* Cover the reconfigure Name-overwrite branch and the non-assignment sole binding

Two patch-coverage additions: an end-to-end name-referenced overwrite
fixture (the tighter signature referenced through a module-level
constant; the inferred literal replaces exactly the reference's span
and the constant survives), and a fourth decline control whose sole
binding is a loop target rather than a plain assignment.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GcqABsog1rtb4m2CDZdcRD

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

47 of 56 new or added lines in 3 files covered. (83.93%)

3009 of 3602 relevant lines covered (83.54%)

0.84 hits per line