CI: PR cargo-test never executes crates/perry integration suites; main-push full run starves under merge trains (documented near-miss) #5960

Documented near-miss from 2026-07-04: PR CI's change-scoped cargo-test (scripts/ci_test_scope.py) runs unit tests only — Running unittests src/lib.rs lines exclusively in the job log. The crates/perry/tests/*.rs integration/e2e suites are compiled (they appear in the changed-file scope listing) but never executed on the PR path. They only run on the full scope (push to main / --full).

Consequences observed today:

• PR fix(hir): capture-snapshot P0 pair — resolved-name re-registration + assignment tracking #5938 landed with its own acceptance suite (capture_rereg_renamed_class.rs) failing — same_class_name_across_factories_keeps_snapshots_separate was red at the very commit that introduced it, through green required checks. (Fix in flight: the static-call/static-field lowering arms bypass class_renames.)
• The intended safety net — the full run on main push — was cancelled repeatedly by the per-branch concurrency group during a rapid merge train (three squash-merges within ~20 minutes each superseding the previous run), so main went ~a workday with no completed full cargo-test.

This is a design trade (e2e suites compile binaries and take tens of minutes; scoping keeps PR latency sane), not a bug per se. Suggested cheap hardenings, any subset:

1. Merge-train discipline (process, zero infra): after an admin merge, let the main-push full run complete before the next merge lands — or re-dispatch it after the last merge of a batch. (I'm adopting this immediately for my own merges.)
2. Concurrency carve-out: give the main-branch full run cancel-in-progress: false (queue instead of cancel) so merge trains can't starve it.
3. Scoped e2e tier: when the changed files include crates/perry/tests/<suite>.rs, run those suites (they're the PR's own acceptance tests — the exact class of miss that happened here).
4. A nightly full run as a backstop (if not already present).

Refs: #5938 (the near-miss PR), the fix PR for the renamed-class static dispatch, run 28698034143 (an old-commit re-run that was the only completed "full" signal today, itself red on the known load-sensitive byob stall).