ponder-lab / Hybridize-Functions-Refactoring / #2821

84%

main: 85%

push

github

Impose no eager-effective dtype from a partner that is itself a parameter (#879)

* Impose no eager-effective dtype from a partner that is itself a parameter

Two parameters combined only with each other each pinned to the other's
dtype: the eager-coercion detector read every binary-op partner's tensor
typing, so a symmetric pair decided each other with no fixed point, and
the emitted specification swapped their dtypes (#878, observed on
TensorFlow2.0-Examples' RPN compute_loss, whose orientation flipped
between Ariadne 0.52.82 and 0.52.83).

A partner that is itself a parameter now imposes nothing, the same
exclusion the upstream parameter coercion applies (wala/ML#828), keeping
the two implementations in agreement until #875 collapses this side into
a read. The vendored reduction pins both sides of the change: pre-fix it
emitted the swapped [float32, float64]; the spec now carries the fed
[float64, float32] unchanged, with neither the pin nor the decline
firing.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X8W9PDNUSQFzZRUAYJmRiE

* Decline the conversion for a divergent directly-combined parameter pair

The exclusion alone traded an arbitrary emission for a deterministic but
still-broken one: with each parameter refusing to decide the other, the
specification carried the fed evidence, which for a directly-combined
pair names two dtypes the traced subtraction rejects, and the bare
decorator fares no better, since tracing materializes a weak argument at
its own dtype (the #861 premise). Either-orientation pin breaks one
reading, so no transformation survives, which is exactly the plural-set
meaning of HAS_CONFLICTING_EAGER_DTYPE_COERCIONS.

The detector now records partner operands that are parameters, and the
consumer compares the pair's own evidence: single concrete dtypes that
differ set the conflicting verdict (code 22, no new machinery), while
top or plural evidence on either side stays ... (continued)

26 of 27 new or added lines in 2 files covered. (96.3%)

3245 of 3844 relevant lines covered (84.42%)

0.84 hits per line