fix[scripts]: emit real diffs from the DaCe determinism check

[havogt]

Motivation

The test_next_dace_determinism session failed on a CI job with:

error: DaCe codegen is non-deterministic: 1/135 comparable program(s) differ

The uploaded artifact could not say what differed. write_diffs only wrote
the program name plus the relative paths whose (relpath, sha256) pair
differed, and run_determinism_check deletes run1/, run2/ and .dacecache
in its finally — so by the time the artifact was uploaded the two versions of
the file were gone:

testee_pyext
src/cpu/testee.cpp

Diagnosing it needed a local reproduction of the whole selection.

Change

_scan now also returns a (relpath, sha256) -> Path map, and write_diffs
uses it to render each differing source as a unified diff (capped at
MAX_DIFF_LINES).

One wrinkle: a name is compiled once per set of parameters, and a variant
carries no identity that survives across runs, so run1's variants cannot be
matched to run2's by digest. The previous placeholder gave up here — which is
exactly the case that occurred, since testee had 34 compiled variants of
which 2 differed. Variants are now paired greedily by smallest diff; pairing
them in digest order diffs unrelated variants against each other and reports
the difference between the variants (here: scan direction) rather than the
non-determinism.

Result

Same failure, replayed against the new code:

testee_pyext
src/cpu/testee.cpp

=== src/cpu/testee.cpp
-- variant: run1 77e1d8db7768... vs run2 345392b365eb...
--- run1/src/cpu/testee.cpp
+++ run2/src/cpu/testee.cpp
@@ -14,10 +14,10 @@
+    int * gpu___gtir_scan_output_carry_0;
+    DACE_GPU_CHECK(cudaMallocAsync((void**)&gpu___gtir_scan_output_carry_0, ...));
     int * gpu___gtir_scan_output_carry_1_0;
     DACE_GPU_CHECK(cudaMallocAsync((void**)&gpu___gtir_scan_output_carry_1_0, ...));
-    int * gpu___gtir_scan_output_carry_0;
-    DACE_GPU_CHECK(cudaMallocAsync((void**)&gpu___gtir_scan_output_carry_0, ...));

The underlying bug this surfaced

Not fixed here (it is a DaCe issue), but for the record — the differing program
is testee from test_scan.py::test_scan_nested_tuple_output, both forward
parametrizations. In dace/transformation/interstate/gpu_transform_sdfg.py:

for inodename, inode in set(input_nodes):
for onodename, onode in set(output_nodes):

input_nodes / output_nodes are lists built in deterministic order, then run
through set() to dedup. Data.__hash__ is hash(serialize.dumps(self)) — a
string hash, so iteration order follows PYTHONHASHSEED. That order is the
insertion order of the gpu_* clones into sdfg.arrays, hence the emitted
alloc/free order. Verified: identical output for a fixed seed, different output
across seeds; replacing both set(...) with an order-preserving dedup makes
the output seed-independent.

Testing

scripts/tests/python/test_dace_determinism.py — 46 pass, incl. four new cases
covering the unified-diff output, smallest-diff variant pairing, unpaired
variants, and truncation. Validated end to end by replaying the real
two-cache reproduction. pre-commit run clean.

https://claude.ai/code/session_01Sf2j16DZhoaJJgy2gWyKUR

[havogt]

cscs-ci run dace-determinism

[havogt]

Not sure if I like the change, but it's certainly more useful to have the diff.