Merge train: #9918, #9950, #9957, #9958, #9976, #9977

[proggeramlug]

Merge train: #9918, #9950, #9957, #9958, #9976, #9977 — the two stacks (#9918 ⊂ #9958 ⊂ #9977 regex, #9950 ⊂ #9957 ⊂ #9976 GC), kept out of #9978 so a failure in ~8k lines of GC and regex work stayed attributable.

The conflict worth recording

In copying.rs, #9976's side of a hunk was empty, which reads as "theirs has nothing to add" and is wrong: it meant #9976 deleted that region, having relocated finalize_dead_copied_minor_from_space_side_allocations into copying_phase.rs with a new return type. Keeping ours produced a duplicate definition and a type error. The correct resolution splits the hunk — keep #9861's LAST_COHORT_SPLIT test helpers, which #9976's base predates, and drop the relocated function. An empty side is a deletion, not an abstention.

The [gc-copy-minor] diagnostic line conflicted twice, because #9861 (landed in #9978) added cohort-split fields while #9976 added pause_us/scan_us and then phases:. Merged to carry all of them; placeholder and argument counts verified equal (22/22) rather than assumed.

Gate fixes

• copying.rs hit 2066 lines. The remembered-set scan and pinned-young preflight move to copying/remembered_scan.rs. Note pub(super) means gc in the parent and copying in the child, so a straight move silently narrows reach and the glob re-export then has nothing public enough — both are widened to pub(in crate::gc).
• family_push_back became dead in a product build. Before gating it I checked that perf(gc): prune the per-object layout tables from a young-entry log (replay of #9895) + a residue histogram — 4 ms per steady minor on cc #9976 dropping its production rekey caller was deliberate: it is, and the replacement comment explains that re-entering the writer funnel mid-walk would enqueue the same family twice. shapes_test_support is now its only consumer, so #[cfg(test)] matches.
• regex_census.rs's rows is extended only under regex-engine: unused-mut without the feature, required with it. The allow is scoped to the feature-off build rather than dropping mut, which breaks the other configuration.
• Five new holders classified. BOX_YOUNG_ROOTS is covered_elsewhere — every address in that minor remembered set is also in the box registry scan_box_roots_mut walks, so the log is an accelerator, not an independent holder. The four test seams and the histogram counter are not_a_gc_pointer.
• PASS1_MARKED re-audited. gc/mod.rs gains two module declarations and one reg_scanner! registration; a scanner registration adds a root SOURCE and runs nowhere between the census boundaries. gc/census.rs widens side_tables() and adds census rows plus a test — reporting that runs from the diagnostic dump, not inside a cycle. LAYOUT_DIAG's entry is deleted because the holder became covered, which the gate calls the receipt.

Validation

run_lint_gates: 3 of 80 FAILED, all accounted for: public-baseline is pre-existing on main (verified on a pristine worktree, red since 2026-07-29), and the two API docs gates are an artifact of this session's CARGO_TARGET_DIR override — regen_api_docs.sh hardcodes $ROOT/target/release/perry. With the binary placed where it expects, regeneration succeeds and the drift check is clean.

suite	passed	failed
perry-runtime	3310	0
perry-codegen	1947	0
perry-hir	630	0
perry-stdlib	132	0