ponder-lab / Hybridize-Functions-Refactoring / #4160

90%

Build Type

push

github

Committed by web-flow

Commit Message

Never take a one-value points-to set as evidence that a parameter is invariant (#1271)

* Never take a points-to set's size as evidence that a value is fixed

A points-to set may not hold every value a parameter receives: a call outside the analysis, such as Keras's own fit and predict passing training, an argument spread through a starred parameter, or one the engine gives no abstract value, can pass another. So a primitive bound to one constant at every call the analysis sees is still a primitive parameter, a branch condition is fixed only when it's a tensor in every context, and a layer call is fixed only by how its callee is read: the receiver, a module's name, a field written only while its object is set up, or an element of one under a fixed key. Points-to still shows variation, never fixedness.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

* Keep how a layer call's selection is decided as on main

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

* Scan a summary's writes once, not once per field

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

Coverage Stats

20 of 20 new or added lines in 3 files covered. (100.0%)

15 existing lines in 2 files now uncovered.

6811 of 7592 relevant lines covered (89.71%)

0.9 hits per line