[mark14wu]

Summary

An address built from a reduction of loaded values fails under the sanitizer:

seq = idx - tl.max(tl.where(cumsum_lens <= idx, cumsum_lens, 0))

NotImplementedError: Concretize for op max is not implemented

When an address depends on loaded values, the sanitizer concretizes the address's subtree (replace_subtree), and ReduceSymbolicExpr had no concretize. This PR adds one for all of its ops: sum, max, min, argmax, argmin, xor_sum, and reduce_or, computed with NumPy in the reduction's result dtype. max and min use fmax/fmin on floats, so they ignore NaNs as Triton's maxnumf/minnumf do; argmax and argmin return the first index, as with tie_break_left=True.

Tests

• New in tests/end_to_end/test_sanitizer.py: test_sanitizer_concretizes_reductions_in_addresses, a kernel that loads from offsets given by each of the seven reductions of loaded indices. With a large enough source tensor it expects no report; with smaller ones it expects reports at exactly the offsets whose reduction value is out of bounds. All 3 cases fail on main with the error above.
• pytest tests/: 366 passed, 6 skipped on main; 369 passed, 6 skipped with this PR (Python 3.12, Triton 3.8.0).
• ruff check and ruff format --check (v0.1.9) pass on the changed files.

🤖 Generated with Claude Code

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.101s	0.100s	-0.4%	20 / 20
gemm_oob	0.113s	0.112s	-0.8%	20 / 20
indirect_load	0.021s	0.021s	-0.9%	20 / 20
nested_loop	0.225s	0.223s	-1.1%	20 / 20
block_pointer_loop_advance	0.120s	0.122s	+1.3%	20 / 20
liger_jsd	0.134s	0.133s	-0.2%	20 / 20
flaggems_layernorm	0.379s	0.377s	-0.4%	20 / 20
swiglu	0.164s	0.164s	-0.1%	20 / 20
cross_entropy	0.944s	0.936s	-0.8%	20 / 20
fused_linear_jsd	0.203s	0.202s	-0.2%	20 / 20
Total	2.404s	2.392s	-0.5%	N/A

Iterations: 1 warmup + 20 measured
Samples are shown as main / PR; long pytest benchmarks may use fewer samples.