ponder-lab / Hybridize-Functions-Refactoring / #2632

83%

main: 84%

push

github

Add an unresolved statically-read-axis safety precondition (#817)

* Add an unresolved statically-read-axis safety precondition

The refactoring can write an inferred input signature that breaks a function which runs correctly today: a wildcard axis is `None` at trace time, and a body that reads that axis statically (`x.shape[i]`, `input_shape[i]`, `K.int_shape`) and consumes it where a Python integer is required then raises, or silently misbehaves through a `[:None]` slice. A dynamic read (`tf.shape(x)[i]`) returns a tensor and is safe, which is the distinction the check draws. Functions whose inferred signature leaves such an axis unresolved now fail with `HAS_UNRESOLVED_STATICALLY_READ_AXES` (code 18), the fourth safety failure in the P1 chain, and the condition also blocks every hybrid reconfigure path via `canReconfigure`.

Per the design settled on #811, `StaticShapeReadAnalysis` mirrors the parameter-flow numpy walk with three inversions: a static `.shape` read is the event of interest rather than laundered, carrying per-parameter provenance and covered axes narrowed by constant subscripts and prefix/suffix slices; `tf.shape`/`size`/`rank` are laundered instead; and descriptors join by union, so lost provenance widens toward declining, the conservative default the precondition requires. The sinks are the enumerated consumption sites observed in the corpus, held as string constants: a weight extent passed to `add_weight`, a `tf.reshape`/`tf.range` target (with element-through-container flow onto the written list), and integer arithmetic over a dimension. Keras lazy-`build` reachability covers both directions: the function's own `build` is a call-graph sibling under the `__call__` trampoline and is seeded by name-based lookup anchored to the first non-receiver parameter, while a sublayer's `build` is forward-reachable and seeded with unknown provenance, mirroring `subtractBuildProtocolContributions`'s reachability model. The verdict resolves each... (continued)

336 of 392 new or added lines in 4 files covered. (85.71%)

2778 of 3330 relevant lines covered (83.42%)

0.83 hits per line