[jschmied]

Follow-up to #1440, as asked there. One commit, now rebased onto main after #1440's merge.

What

When an image differs between base and head, the tool builds it again from both trees, one build at a time:

• A function counts as changed only if base and head differ again in the rebuild.
• If a tree's two builds disagree on any function, it is reported as unstable, and the exit status is non-zero. A flaky build can never pass quietly.
• Images that match on the first builds are not rebuilt, so a clean run costs the same as before.
• --no-recheck keeps the old behaviour: the first builds' differences count, without a rebuild.

The logic sits in one function, recheck(), and the main loop only builds and prints.

Checks

• python3 -m unittest cuda_sass_identity_test (from tools/zig): 6 tests pass. The 3 new ones use canned dumps:
  • a change both builds reproduce counts once, with nothing unstable;
  • a head whose rebuild matches the base is not a change, but it is unstable;
  • a base whose rebuild disagrees with itself is unstable.
• A real change, NVFP4 fact tiles: fact_swiglu reads each mma A quad in one 16-byte load #1525 (b30d5b9d3 → 1e627ed14, nvcc 13.0):
  • nvfp4_experts has 5 functions changed (tf_nvfp4_fact_swiglu on 5 SMs);
  • after the rebuild: 5 differ again, 0 unstable.
• A simulated unstable build: an nvcc wrapper gives only the base's rebuild -Xptxas -O1. Result: 65 functions unstable, exit 1.
• lean_check: 0 new. zig build test: passes.

[ashhart]

Thanks @jschmied, landed as 80d7c3b with its changelog line. The rebase met #1555's move of the fp4pair control to common/fp_pairs.cuh, so the new usage text names that path; the six unit tests pass.