[FabianIsensee]

Hi Francesc,

this follows up #754, as you asked there. Your typed evaluator supersedes #754's source change, so this PR only brings over its tests, adapted to improved-miniexpr at 7eae862, plus one strict xfail for a bug I found on the way.

What it adds (all in tests/ndarray/test_lazyexpr.py):

• test_lazyexpr_result_dtype_matches_numpy: 17 expressions (float functions and true division of integer and bool subexpressions, bool //, % and **, ...) × bool/int8/uint16/int32/int64 × built from functions or from a string. It checks the dtype of expr[:], expr[...], compute() and expr[:7], and the values of expr[:], compute() and expr[:7] against NumPy. Give lazy expressions NumPy's result dtype for division and float functions #754 did not check the values of expr[:], which the bundled miniexpr computed wrongly there.
• test_lazyexpr_non_native_byte_order and test_arctan2_rejects_complex_when_built, as in Give lazy expressions NumPy's result dtype for division and float functions #754.
• test_comparison_with_out_of_range_python_int, a strict xfail (see below).

WebAssembly: #754 marked 30 cases as strict xfails on WASM, because there the bundled miniexpr computed them. With the typed evaluator they pass, also with blosc2.IS_WASM = True set natively, so this PR has no WASM xfails. The WASM CI job checks it on a real Pyodide build.

Legacy evaluator: with BLOSC2_EXPRESSION_EVALUATOR=legacy, 80 cases still fail: 64 give the uniform evaluator's result dtype, and 16 compute integer true division in integer arithmetic on expr[:]. They are strict xfails in legacy mode only (_LEGACY_RESULT_DTYPE_FAILURES), so the list can go when the legacy evaluator does.

Bug in the typed evaluator (strict xfail): blosc2.lazyexpr("a < 300", {"a": a}) with an int8 a raises OverflowError: Python integer 300 out of bounds for int8 on [:], compute() and .sum(), and so does a < 300. NumPy 2 compares (all True); only arithmetic with such an int raises. The error comes from expression_graph._numpy_binary_dtype(), called through native_weak_scalars_eligible(), which treats the comparison like arithmetic. The legacy evaluator casts 300 to int8 (44) instead: 144 True of 200 on [:] and compute(); its .sum() is right.

Not a bug, so not included: the float16 sums of hypot()/arctan2() of bool arrays that I had flagged match NumPy's own float16 sums to one float16 ulp (hypot: 3324 vs NumPy 3322; arctan2: equal). My earlier reproducer compared them with a float64 sum.

Tests (NumPy 2.5.4, Linux x86-64, bundled miniexpr 687e02b and c-blosc2 3.5.0):

• New tests: default (typed) 197 passed, 15 skipped, 3 xfailed; legacy 118 passed, 15 skipped, 82 xfailed; the same with blosc2.IS_WASM = True.
• tests/ndarray/test_lazyexpr.py: typed 1053 passed, 15 skipped, 3 xfailed; legacy 974 passed, 15 skipped, 82 xfailed.
• Full suite: 11184 passed, 389 skipped, 3 xfailed; the 2 failures are the b2view http tests that need aiohttp.
• ruff check and ruff format pass.
• Not run locally: NumPy 1.x (the out-of-range test is skipped there, like the other NEP 50 tests in this file) and a real WASM build.

AI disclosure: the port, the new test and the checks were done with Claude Code; I reviewed them.

Best,
Fabian

🤖 Generated with Claude Code

[FrancescAlted]

Thanks @FabianIsensee for this!

Merged into improved-miniexpr through improved-miniexpr2 (merge commit 2e20a376). The branch is now at 0d707e89 and pins miniexpr 12bf4da; it is the head of PR #741.

The combined stack passed the full test suite (12,841 passed on the default path, 12,765 on legacy) and the Menudet paired qualification, including the exact-pair wheel and tcc JIT tests (run 41998579498). Thanks again!