fix(runtime,codegen): bundled-zod capture-class family — statics dispatch, value-super, ctor args, constructor identity (#6530)

[proggeramlug]

Fixes #6530.

Next.js's bundled zod (next/dist/compiled/zod/index.cjs) broke in three ways, all rooted in capture-carrying classes — classes nested in a (webpack-wrapper) function whose methods reference outer locals, which Perry materializes as per-evaluation heap class objects instead of INT32 ClassRefs. zod's whole ZodType family is this shape (the helper namespaces o/i/k are captured), which is why z.string().optional().parse(undefined) threw while trivially-reduced mimics passed.

Defect 1 — post-class statics on capture classes are invisible to ClassRef static dispatch (repro 1)

ZodOptional.create = (e, t) => new ZodOptional({...}) lands on the fresh class object's own fields, but a sibling method body (ZodType.optional()) compiles the same call as js_class_static_method_call(ClassRef(cid), "create", …), which consults only the cid-keyed CLASS_DYNAMIC_PROPS side table. The miss silently returned the receiver — so .optional() evaluated to the ZodOptional class ref, and .parse(undefined) then threw Cannot read properties of undefined (reading 'errorMap').

Fix (field_set_by_name.rs): mirror named-field writes on a marked class object into CLASS_DYNAMIC_PROPS under its template class id (internal __perry_* markers excluded; own-field write unchanged; last-wins across evaluations, matching the existing template-cid compromise).

Defect 2 — value-super with a class-object parent constructs-and-drops (repro 2, part of repro 1's family)

new ZodEffects({...}) from a sibling method takes the dynamic-parent super leg (js_fetch_or_value_super). With a class-object parent it dispatched js_native_call_value, whose W2 class-object arm constructs a fresh discarded parent instance instead of running the parent constructor on this — ZodType's ctor never ran, _def and the bound methods stayed undefined (transform/default/catch/refine, and the record/object repro's _getType crash).

Fix (fetch_globals.rs): add a class-object arm mirroring the existing INT32-ClassRef arm — resolve the template cid stamped on the class object and run run_class_constructor_on_this_flat on this.

Defect 3 — user args swallowed as capture fallbacks at direct-ctor call sites (repro 2)

The bare-identifier new C(...) codegen path assumed the HIR always appends the class's captures as trailing LocalGet args. The HIR only does that where the captured locals are in scope (the declaring function); inside a sibling class's method nothing is appended — so the tail-split treated the one user arg as a capture fallback: the synthesized ctor got an empty rest array, super(...[]) ran the parent ctor with no def.

Fix (lower_call/new.rs): treat the tail as appended caps only when each trailing arg is LocalGet(id) matching the synthesized __perry_cap_<id> param it binds to (exact id + order, all-or-nothing — the HIR emitter's invariant).

Defect 4 — .constructor / .name identity collapse (repro 3)

Instances of capture classes resolved .constructor to a bound constructor-method closure (underlying func = nearest ancestor ctor), so x.constructor === Sub was false even for bases and every subclass reported the base's name (z.string().constructor.name === "ZodType"); zod's describe() re-constructs via this.constructor and produced base instances. Separately, three static-resolution walks (classref CLASS_DYNAMIC_PROPS chain, classref resolve_proto_chain_field, class-object pinned-parent recursion) inherited name from the parent edge — but name/prototype are own properties of every constructor, never inherited.

Fix: new CLASS_OBJECT_VALUES registry (cid → marked class object, populated by js_object_mark_class, GC-rooted in both the mut scan and the snapshot machine exactly like CLASS_DYNAMIC_PARENT_VALUE), consulted first in the instance .constructor arm; name skipped in the three inheritance walks (prototype also skipped in the pinned-parent recursion).

Validation

All three issue repros now match node byte-for-byte against the bundled zod:

z.string().optional().parse(undefined)  → undefined
z.record(z.string(), z.object({p: z.string(), q: z.boolean().optional()}))
  .parse({a:{p:"x"}})                   → { a: { p: "x" } }
z.string().constructor.name             → "ZodString"
z.object({}).constructor.name           → "ZodObject"

Every ZodType combinator was probed (optional/nullable/array/promise/or/and/transform/default/brand/catch/describe/pipe/readonly/refine/nullish) — all now produce real instances with correct _def.typeName.

Boot-time manifestation: Next 16's real next/dist/server/config-schema.js (the gscmaster standalone boot blocker) now initializes under Perry, and configSchema.safeParse(<valid config>) matches node. The validation-failure path hits a separate pre-existing gap (new.target undefined in a CJS-nested class X extends Error ctor — bundled zod's ZodError), filed as #6535 with a reduced repro; it reproduces identically on a pre-fix main build.

New gap test test-files/test_gap_class_capture_forward_static.ts covers the shape without the zod bundle (fails pre-fix, byte-identical post-fix). A full scripts/run_gap_tests.sh sweep is running locally; result will be posted as a comment. perry-codegen/perry-runtime lib tests pass (the two gc::tests::teardown parallel-run failures reproduce identically at the base commit — pre-existing parallel-suite flakiness; they pass with --test-threads=1).

Summary by CodeRabbit

• Bug Fixes

  • Fixed subclass constructors with captured values so forwarded arguments and super(...) calls work correctly.
  • Preserved per-evaluation class identity when accessing constructor and invoking inherited static methods.
  • Corrected constructor name and prototype property resolution for subclasses.
  • Ensured class property updates and garbage collection retain the correct class objects.
  • Fixed dynamic parent construction so base-class fields are initialized correctly.

• Tests

  • Added regression coverage for captured class values, forwarded static methods, subclass construction, and parent initialization.

[coderabbitai]

📝 Walkthrough

Walkthrough

The change validates appended constructor capture arguments during lowering and adds per-evaluation class-object storage with GC integration. Runtime lookup, writes, static-name handling, and super() dispatch now preserve class identity, with regression coverage for nested capture-carrying classes.

Changes

Capture-Carrying Class Identity

Layer / File(s)	Summary
Constructor capture validation crates/perry-codegen/src/lower_call/new.rs	Validates trailing __perry_cap_* LocalGet arguments before treating captures as present during new lowering.
Class-object registry and GC roots crates/perry-runtime/src/object/class_registry/{state.rs,gc_roots.rs}, crates/perry-runtime/src/object/class_registry.rs, crates/perry-runtime/src/object/class_registry/parent_static.rs	Stores per-evaluation class-object pointers by class ID, re-exports the registry items, registers class objects, and scans them as GC roots.
Class identity and dispatch paths crates/perry-runtime/src/object/field_get_set/*, crates/perry-runtime/src/object/field_set_by_name.rs, crates/perry-runtime/src/object/global_this/fetch_globals.rs	Resolves per-evaluation constructors, restricts inherited name and prototype lookup, mirrors class-object writes, and invokes class parents directly during super().
Capture-forwarding regression coverage test-files/test_gap_class_capture_forward_static.ts	Tests nested-class static forwarding, dynamic-parent construction, constructor identity, initialization, and parsing results.

Estimated code review effort: 4 (Complex) | ~45 minutes

Possibly related issues

• flaky: capture-bearing class factory reports .name === undefined on some builds (build-nondeterministic) #6482 — Related to per-evaluation class identity and constructor capture handling, although its specific CLASS_NAMES registration objective is not changed here.

Possibly related PRs

• PerryTS/perry#5178 — Shares the constructor capture-preservation path in lower_call/new.rs.
• PerryTS/perry#5213 — Shares appended __perry_cap_* argument recognition during new lowering.
• PerryTS/perry#6449 — Shares pinned-parent resolution for per-evaluation class-object statics.

Suggested reviewers: thehypnoo

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly summarizes the bundled-Zod capture-class runtime/codegen fixes and matches the main changes.
Description check	✅ Passed	The description covers the issue, related issue, detailed fixes, and validation, though it doesn't use the template's exact section layout.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/6530-zod-subclass-identity

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[proggeramlug]

Closing in favor of a re-push under a new branch name — this branch never triggered the required Actions checks (known CI-trigger quirk). Replacement PR carries the identical commits.