33 lines
1.7 KiB
Markdown
33 lines
1.7 KiB
Markdown
# Root-channel parity diagnostic
|
|
|
|
This is a read-only diagnostic, not a replacement acceptance gate. It requires a
|
|
caller-admitted GPU and the unchanged checkout containing
|
|
`test_channel_root_gpu.py`. It does not run unrelated body/Jet tests.
|
|
|
|
From the checkout, run:
|
|
|
|
```
|
|
python /path/to/channel_parity_diagnostic/discover.py --checkout "$PWD" --output /new/evidence/directory
|
|
```
|
|
|
|
Discovery invokes only the original `test_uniform_matched_root` method, with its
|
|
original loops and exact assertion. It stops at the first assertion failure,
|
|
extracts the configuration and arrays from that original traceback, saves the
|
|
failure verbatim, and serially starts `replay.py` in a fresh Python process.
|
|
The original failure remains a failure: exit 1 means the assertion was captured,
|
|
even if the targeted replay succeeds or does not reproduce it. Exit 2 means the
|
|
original test did not reproduce the failure; exit 3 means a different assertion
|
|
failed and no configuration was guessed. Other errors are retained and raised.
|
|
|
|
Targeted replay compiles only the uniform and one failed blocked configuration,
|
|
serially, caching the two modules. It saves PTX and compiler metadata, all
|
|
numeric and bit-only mismatches with coordinates/hex/ULPs at each step, first
|
|
numeric divergence arrays, same-input localization, and comparison against the
|
|
original multistep call. Signed-zero-only bit changes cannot trigger the first
|
|
numeric divergence. The final replay consistency checks establish whether the
|
|
one-step observation sequence agrees with the original call structure.
|
|
|
|
Outputs retain the original source revision and gate SHA256. No source,
|
|
assertion, tolerance, compiler flags, collision math, or kernel scheduling is
|
|
changed. Findings must be reviewed before any production fix or gate change.
|