[jimisola]

What & Why

To cut CI eval cost on narrow PRs, a PR now runs only the cases its changes reach, not every case of every changed plugin:

• a change under plugins/<p>/skills/<skill>/ runs every case in plugins/<p>/evals/<skill>/;
• a change under plugins/<p>/evals/<skill>/<case>/ runs that case;
• a change to the workflow itself, or a manual run, runs every case of the chosen plugins (as before).

The Pick plugins step becomes Pick cases and outputs case directories; the run loop iterates over them. Stacked on #28 (-j 3), which touches the same loop.

Effect on past PRs

PR	Change	Cases before	Cases now
#23	merge-ready SKILL.md + graders	14	6
#24	one grader	14	1
#20	bambu-3mf script	5	3
#21	new cad-tools cases	5	5

Test Plan

• actionlint clean
• claude plugin validate .
• The step's script run against the diffs of fix(cad-tools): record changed settings so Bambu's GUI keeps them #20, test(cad-tools): add claude plugin eval suites for bambu-3mf and turntable-video #21, ci: run claude plugin eval on PRs with workload identity federation #22, fix(github-tools): merge-ready asks for scope in the reply when AskUserQuestion is unavailable #23, test(github-tools): anonymous-404 grader accepts the skill's own fallback #24 and a manual run for cad-tools: selections as in the table; the workflow change selects all 19
• This PR's own run (workflow changed, so all 19 cases): all pass