[codex] Prioritize old reclaim at parse boundaries

[andrewtdiz]

Summary

• Teach the pending tiny-parse GC boundary to honor old-generation reclaim pressure before scheduling another copied-minor collection.
• Add a regression test that arms parse-boundary collection under old-gen pressure, verifies a full old_gen_bytes reclaim runs, and confirms a rooted old object survives.

Root Cause

bench_json_roundtrip was repeatedly entering parse-boundary copied-minor collections even after lazy JSON/tape churn had moved pressure into old/large-excluded storage. Those minor collections could not reclaim the dead old-gen pressure, so Perry spent time copying/promoting young objects while RSS kept growing.

Evidence

Targeted trace, same worktree before/after this commit, PERRY_GC_TRACE=1:

Metric	Before	After
stdout timing	json_roundtrip:327	json_roundtrip:80
checksum	53735550	53735550
RSS	245588 KB	134528 KB
elapsed	0:00.39	0:00.14
GC events	36 minor	2 full
triggers	36 arena_bytes	2 old_gen_bytes
total GC pause	173554 us	31229 us
max GC pause	28927 us	16222 us
swept bytes	816368	94820920
copied bytes	10514712	0
promoted bytes	3502376	0

Full suite one-run after this change, with absolute JSON output path:

• PERRY_BIN="$PWD/target/release/perry" benchmarks/compare.sh --full --runs 1 --json-out "$PWD/tmp/cycle41-profile/full-after.json" --warn-only
• Target row improved from the pre-change cycle 41 suite run bench_json_roundtrip: 392ms / 236480KB to 91ms / 134660KB.
• The compare script still reports broad regressions against historical baseline commit c89b3ad5; those are existing branch-vs-baseline differences from before this focused PR, not new rows introduced by this patch.

Validation

• cargo fmt -- --check
• cargo test -p perry-runtime gc::tests::triggers
• cargo test -p perry-runtime gc::tests::host_safepoints
• cargo test -p perry-runtime gc::tests::budgeted_step_api
• cargo test -p perry-runtime json::
• cargo build --release -p perry
• Recompiled bench_json_roundtrip.ts with ./target/release/perry compile --no-cache ... and ran the traced benchmark above.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 966cf3c2-d78a-4f1b-93a6-85a2aecd3a4a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/perry-gc-performance-cycle-20260618-41

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}