Checkpoint GC architecture progress toward #1090

[andrewtdiz]

Summary

Checkpoint toward #1090: make Perry's minor GC structurally cheaper and safer to reason about.

This PR collects the current GC architecture work into a reviewable unit:

• makes copied-minor GC the default-safe path for key churn workloads
• strengthens precise root coverage and copied-minor fallback reporting
• keeps write-barrier/dirty tracking evidence visible in trace output
• adds target collector architecture gates to the memory-stability suite
• splits the large GC runtime implementation into owner modules behind the same public facade
• removes JSON tape scratch/key-cache allocation churn and documents the stable force-evac verifier RSS bound

This is not intended to close #1090. It is a strong checkpoint: default-safe copied-minor behavior, better evidence, and a maintainable module layout for the remaining root/barrier/layout/old-gen work.

Why

#1090's target invariant is:

minor GC cost = live young objects + dirty old cards/pages + mutable roots
not arena size + malloc table + block persistence + sweep walk

This PR moves Perry closer to that invariant by making copied-minor behavior observable and verified, reducing fallback causes, and giving the collector code clearer ownership boundaries before the next architecture work.

Evidence

Verified locally on this branch:

• cargo fmt --check
• cargo test -p perry-runtime --release json_tape -> 6 passed
• cargo build --release -p perry
• git diff --check
• PERRY_GC_EVIDENCE_DIR=/tmp/perry-gc-pr-checkpoint-evidence scripts/run_memory_stability_tests.sh -> 58 PASS / 0 FAIL

Memory-stability highlights:

• JSON churn default/gen-gc: 240 MB under 250 MB gate
• JSON churn mark-sweep: 90 MB under 250 MB gate
• JSON churn force-evac+verify: stable 262 MB under documented 275 MB verifier-only gate
• copied-minor fallback evidence: top_remaining_reason=none
• target collector stress gates passed
• low-pressure benchmark gates passed:
  • 07_object_create: 3 ms / 6 MB
  • 12_binary_trees: 4 ms / 6 MB
  • bench_gc_pressure: 23 ms / 26 MB

Evidence artifacts from the local run were written to:

/tmp/perry-gc-pr-checkpoint-evidence

Deliberately not included

The remaining #1090 production-grade work is still ahead:

• full 11-run benchmark matrix packet from fresh worktrees
• broader old-gen/page layout work
• complete GC type/layout metadata replacement for hard-coded ranges
• broader unboxed-field and typed heap layouts
• further reduction of conservative fallback paths
• making write barriers permanent infrastructure rather than mode-controlled behavior