perry: soak test recent perf landings against real-world apps before tagging release

[proggeramlug]

Summary

The May-2026 release-readiness session landed several non-trivial perf changes that haven't been tested against shapes outside honest_bench + benchmarks/suite + ECS smoke tests. Each has a plausible regression risk we should de-risk before tagging a release.

Recent landings to soak

Commit	Change	Surface area
56818086	GC bytes-trigger bump after gc-suppressed parse	Any workload calling JSON.parse on > 1 MB blobs
b3e93649	Inline ArrayPush fast path for local-bound arrays	Any local-array-push hot loop
3f5c69a5	integer_locals forward-closure pass	Any program with chained int-stable Lets (let x = a*b; let y = x+c; ...)
4f895dd8	Immutable i32-shadow + i32 init path	Touches every const numeric Let — broad blast radius
cf540b92	Buffer-aware stmt_preserves_array_length	Any for (let i = 0; i < buf.length; i++) shape
f02541a4	FuncRef + Math.* arms in preserves-length walker	Any peephole that runs for (...; i < arr.length; i++) with a body containing function calls or Math methods
aed21292	Static trip-count for-loop unroll	Any tight nested loop with literal bounds and ≤8 iters
817c4b56	>>> 0 correctness fix + channel-vector SIMD reduction	All workloads (correctness); RGB-style image processing (perf)

Soak coverage gaps

• Real-world TS apps: the perry test repo has examples like the postgres bench, the issue-tracker repros (fix(fs): wire fs.appendFileSync end-to-end (closes #226) #231, Array.push from inside an async function silently caps at 16 elements when the array is a function parameter #233, fix(jsruntime): #255 re-entrancy — Perry callbacks can read JS object props #259, axios shim: response.status and response.data are undefined on live HTTPS GET #340, feat(harmonyos): Phase 2 v12 — Tabs / Modal / Menu / Grid (v0.5.432) #357, fix(link): GTK4 gstreamer link + perry-hir stale export (#423) #426); these haven't been re-run end-to-end this session.
• CI parity tests (tests/known_failures.json if any): not exercised this session.
• Customer-facing workloads: any app that imports user code with patterns we haven't anticipated (TS-typed : number accumulators, large numeric constants, string-heavy hash loops).

What's needed

1. Run cargo test --release --workspace (just landed: d826ace4 unblocks the perry crate's 101 tests). Verify all 571 tests pass.
2. Run the parity test suite (or whichever Node-compat sweep exists) and diff against the pre-session baseline.
3. Compile + run a representative set of customer/user TS apps and verify byte-for-byte output match against Bun for at least a 5-minute exercise of each.
4. Document any new known-failures in tests/known_failures.json so CI doesn't surprise us.

Acceptance

• All in-tree tests green.
• Parity sweep delta vs pre-session baseline: zero new regressions.
• 3-5 representative real-world apps verified against Bun.

Related

Related correctness issue: #435 (integer_locals overflow class) — soak should specifically check for the 14_closure shape.
Related perf issue: #436 (image_conv 5%-target).