ponder-lab / Hybridize-Functions-Refactoring / #3710

89%

main: 90%

push

github

Excuse a summary's writes to the object it allocates (#1141)

* Excuse a summary's writes to the object it allocates (#1140)

Code 39 took two kinds of engine summary for rebinding writers: the
synthesized constructor of a NamedTuple without an __init__, which
stores its arguments into the new object's fields, and a TensorFlow
operation's summary, such as tf.concat's, populating the tensor it
returns. Each allocates the object afresh and runs no other code, so
its writes set the object up rather than rebind it. These are the
objects tf.function flattens into its inputs on every call. An opaque
argument whose field another function rebinds is still refused, and
so is a class attribute that module code rebinds, since program
source can write an object again after allocating it.

The refusal message now names the field, the object's class, the
rebinding function, and the write's file and line.

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

* Read a field holding a summary's object as state (#1140)

Excusing a summary's writes to the object it allocates left a gap. A
NamedTuple held in a field that another method rebinds was converted,
and a trace kept the old object. The rebinding write is to the field
holding the NamedTuple, which isn't tensor-typed, so it was never read
as state. A field read whose value may be an object a summary allocated
now counts as a state read, whatever the value's type. An object passed
in as an argument is read through no field, so its fields stay excused.

The excuse also now requires each of the summary's writes to the field
to be on the value its own allocation of that site defines, rather than
any write by the allocating node.

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

* Don't read a super() proxy's fields as state (#1140)

The front end lowers each super(... (continued)

72 of 78 new or added lines in 2 files covered. (92.31%)

5376 of 6056 relevant lines covered (88.77%)

0.89 hits per line