ponder-lab / Hybridize-Functions-Refactoring / #2417

83%

push

github

Harden the numpy shape precondition: poison ambiguous descriptors and cache the type index (#752)

* Harden the numpy shape precondition: poison ambiguous descriptors, cache the index

Two post-merge review findings on #750:

- colorShapeFrom let the last writer win when a value received shape provenance
  from two conflicting sources (e.g. a phi merge), which could "prove" staticness
  from one source only. Poison such a value to an ambiguous descriptor (null
  source, so it resolves to ⊤/unprovable), the conservative outcome.
- indexTensorTypes was rebuilt on every appliesNumpyToParameters call, i.e. per
  call-graph node of every function. Build it once per function and reuse it
  across the function's nodes; indexTensorTypes is now the shared entry point.

Behavior-preserving on the corpus and the unit fixtures (no conflicting
descriptors arise there); suite 570 green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01He1ULx8nRoHzGicMryry1A

* Make indexTensorTypes package-private

It is only used within the analysis package (from Function); no need to
widen the API surface.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01He1ULx8nRoHzGicMryry1A

* Re-enqueue a value when its shape descriptor changes

colorShapeFrom only re-enqueued a value on first taint, so poisoning an
already-processed value to the ambiguous descriptor never re-evaluated its
uses, leaving a sink decided on the stale descriptor. Re-enqueue on any
descriptor change too; descriptors are monotone (unset -> concrete ->
ambiguous), so this terminates.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01He1ULx8nRoHzGicMryry1A

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

11 of 12 new or added lines in 2 files covered. (91.67%)

2079 of 2559 relevant lines covered (81.24%)

0.81 hits per line