[nicksinas]

What

feed_targets filtered non-target keys out of targets.json with a regex heuristic:

NON_BOARD_RE='_|^noarch$|^qcm[0-9]+$'

That rests on "boards never use _". Replaced with IGNORE_TARGETS — an exact-match (grep -vxF) list of the platform/tune keys actually published in 2024/edge.

Why

The heuristic fails silently in the direction that hurts: a future real target whose name contains an underscore gets dropped from the plan, the sweep runs a smaller matrix, and it still goes green. Same failure mode the _fetch rc-1 contract already guards against ("unknown, abort — never treat as empty").

With an explicit list, anything new that is not a real target falls through to install, fails loudly with no avocado-sdk-<name>, and gets added to the list.

Note on icam-540

icam-540 wasn't caught by the old regex, so it was being swept as a top-level target. It's a board of jetson-orin-nx — it's now ignored as a target key and still built as a board via feed_boards. That's the one behavioral change to the matrix.

Trade-off

The list needs a manual edit when the feed adds a tune. Deliberate: the failure mode is a loud red cell instead of a quietly partial sweep.

[jetm]

@nicksinas, I read the diff and ran the old and new filters against the four live targets.json feeds. The grep -vxF -f <(...) quoting is sound, and 2024/next, 2026/next and 2026/edge give the same output as before. I did not run the script or CI, and I did not test macOS BSD grep.

One thing to fix in the description: it says icam-540 is the one behavioral change to the matrix. The old regex _|^noarch$|^qcm[0-9]+$ never matched armv8a or cortexa53/55/57/72/73/76, which have no underscore, so those seven 2024/edge cells are also dropped. Against today's 2024/edge, the old filter keeps 31 keys and the new list keeps 23. The eight that go are armv8a, cortexa53, cortexa55, cortexa57, cortexa72, cortexa73, cortexa76 and icam-540. All eight failed install in 3-6s in the 2026-10-11 Compat Check run, so dropping them is right, but the next weekly report loses eight red rows and the description accounts for one.

Non-blocking: the header of compat-check-targets.sh still says "all dynamic - nothing is hardcoded here" and "deliberately no skip-list". The comment's own test (no avocado-sdk-<name>) can be read from the feed, since sdk/all ships avocado-sdk-* for exactly the 23 kept targets, so the list could be built from it instead of edited by hand.