examples/ is outside the mypy lint scope, which let a missing-attribute bug reach main

[shoumikhin]

What

.lintrunner.toml comments examples/**/*.py out of the mypy scope, leaving only examples/arm and
examples/openvino. The mypy lint job is the one gate that runs on every pull request, and it
would have caught a bug that shipped on main and took a separate PR to fix.

Concretely, #22298 fixes a read of a config field that no longer exists. Running the exact pinned
mypy 1.14.1 from requirements-lintrunner.txt with the repo's .mypy.ini against that expression:

error: "BackendConfig" has no attribute "mps"  [attr-defined]

Clean after the fix. So the broken PR would have gone red in seconds. The macOS job that did catch it
lives in trunk.yml, which does not start for a change to that directory, so the bug reached main
and failed there first.

What it would cost

Not free. examples/models/llama/export_llama_lib.py alone reports 20 mypy errors today, 14 of them
not import-not-found, and the lint job runs over all files in scope, so they would all have to be
fixed in the same change. One of those errors looks like a real bug rather than a typing nit.

Filing rather than doing it, since it is a much larger change than the fix that surfaced it, and the
per-directory cost should probably be assessed by whoever owns examples/.

Found while reviewing #22298.

[shoumikhin]

Correcting my own cost estimate on this before anyone picks it up, because I had it badly wrong.

I wrote "about a dozen pre-existing type errors". Measured with the exact pinned mypy 1.14.1 and the repo's .mypy.ini, pointed at the single file the bug was in:

Found 495 errors in 44 files (checked 1 source file)

And enabling examples/**/*.py brings 606 new files into scope: 644 total under examples/, of which only 38 are already covered via examples/arm and examples/openvino.

So the real cost is two orders of magnitude above my estimate. The conclusion still holds that the pinned mypy catches the exact bug ("BackendConfig" has no attribute "mps") in seconds, so the payoff is real, but this is not a small follow-up. It would need either hundreds of type fixes across code with many owners, or a per-directory ignore list that largely defeats the purpose.

Leaving this open rather than attempting it, and flagging that whoever takes it should scope it per-subdirectory rather than flipping the whole tree at once. Sorry for the misleading number.