D236 - What a sweep proved crashes belongs in the repository, not in a build directory

status: decided

Ten libraries end the process when loaded on real hardware. Establishing that took an iterative sweep against a console: run, find the try record with no matching res, exclude it, run again - ten crashes and ten reboots' worth of patience, for a list of ten names.

That list lived in $BUILD/excludes.txt. A build directory. rm -rf $BUILD loses it, a second machine never had it, and nothing in the history records it. The prose in docs/HARDWARE.md named the ten, but prose is not something a build reads.

It is now data/hardware/crashers.txt, and both build paths consult it:

• scripts/sweep-build.sh seeds a fresh list from it, so a rebuilt build directory does not walk back into ten known process kills one at a time.
• scripts/oops-rebuild-pkg.sh passes it to make pkg as EXCLUDE. It did not pass anything before - make pkg has no exclusions of its own - so every hardware package built through that script had all ten back in it unless somebody remembered to type them.

The file is not a list of libraries this program refuses to probe. It is what one console did on one day, and re-running the sweep with it emptied is a valid experiment - the right one to make when the firmware changes. SEED=0, or EXCLUDE= in the environment, does that without editing anything. Exclusions are how you see what is behind a crash, never how you stop finding them.

The general form: a measurement that cost hardware time to obtain is repository content. Where it happens to be used is a build directory, and the two are not the same file.