ponder-lab / Hybridize-Functions-Refactoring / #2845

85%

main: 85%

push

github

Revert the container-precedence change, keeping its reproduction (#891)

The change let a container verdict outrank a parameter's own tensor
typing, so a parameter with a container anywhere in its points-to set
withheld its signature rather than emitting one reduced from the sites
that pass it unwrapped (#890, for #888). Its whole justification was that
the offending call site could not be told apart from the conforming ones,
and that premise is false.

Probed on this fixture's shape, at the default context depth, on both a
plain function and a Keras layer `call`: the bare-tensor typing lands on
the call-graph node whose context names the rejecting caller, the tuple
site contributes no tensor type at all, and `Gate.call` and `Outer.call`
elsewhere in the suite carry two nodes with divergent types. The
attribution exists; `Parameter.inferTensorTypes` discards which node
supplied what when it unions across them. Retaining it lets the offending
site be excluded and the correct nested spec emitted, which is strictly
better than withholding and does not rest on points-to membership.

Withholding on doubt also cost more than it was measured to: a spurious
container silently loses a signature, an affected already-hybrid function
loses its P4 reconfiguration for the benign already-optimal verdict, and
the broad container set (dict, set, `enumerate`) carried no fixture at
all. None of that was measured at corpus scale before it landed.

The fixture stays, with its assertions flipped to state the wrong
emission (one `TensorSpec` of the unwrapped site's shape) so the
reproduction survives as a two-sided anchor; `tuples_only` stays as the
contrast that localizes the defect to the mixed evidence. The suite runs
687/687.

Reopens the structural half of #888.


Claude-Session: https://claude.ai/code/session_01R1emLt6bvQa993huMzTaKu

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

2 of 2 new or added lines in 1 file covered. (100.0%)

3313 of 3919 relevant lines covered (84.54%)

0.85 hits per line