ci: a GUI change that only moves a frontend guard's pin runs all 12 python-suite shards #16340

Problem

The frontend's guards are Python tests in repo_tests/, for example i18n_untranslated_ratchet_test.py and frontend_reduced_motion_guard_test.py. A GUI PR that fixes strings or motion has to move one of those pins, which puts a .py file in its diff. .github/filters/python-paths.yml (**/*.py) then runs the whole 12-shard python suite.

Measured on 2026-09-11: of the 9 open PRs that touch the GUI, none is GUI-only. #16313's one non-GUI file is repo_tests/i18n_untranslated_ratchet_test.py, so a terminal-modal fix pays for 12 shards of backend tests (4–10 min of work each, plus 20–47 min queued; see #16332) that it can't affect. A separate "GUI train" wouldn't help, because the cost comes from this trigger, not from how PRs are grouped.

Proposal (the design is open)

When a PR's only Python changes are guard modules under repo_tests/, run those modules, plus anything that imports them, in one job instead of the full sharded suite. Constraints:

• python_filter_covers_its_guards_test.py must keep holding. A change to any tree a guard reads still triggers that guard.
• A change to a guard never skips the guard itself.
• guard(imports): every module under api/ and autobot_shared/ must import inertly (#16198) #16199's import-hermeticity SUBSET mode is the precedent for selecting by what changed.
• A PR that touches both a guard and production Python still runs the full suite.

Acceptance criteria

• A PR whose only .py changes are repo_tests guard modules runs those modules and their importers, not the 12 shards. Measured on a real PR.
• A change to any tree a guard reads still triggers that guard, and the coverage guard stays green, unchanged or tightened.
• A PR that changes repo_tests and production Python together still runs the full suite.
• Sequenced with ci(python-suite): queue the shards first in, first out across PRs, with a per-shard job-level queue: max group (#16320) #16328 (FIFO queue) and ci: python-suite shards run on 4–5-week-old duration files, are 2.7× unbalanced, and appear to pay full-tree collection each #16332 (shard balance), which edit the same job.