Error: retain realm intrinsic prototypes independently of public bindings

[chrisbbreuer]

Parent #465; related #821 and #822.

Reproduced on pushed 4f7718d before the current candidates: a class call can create its TypeError using a shadowing lexical TypeError binding or a replaced global TypeError constructor. Checking Object.getPrototypeOf(error) against the original realm TypeError.prototype fails in both witnesses. The older instanceof error-name fallback concealed this defect; the prototype-identity check does not.

makeErrorWithProto currently resolves named constructors through self.env.get(name), which is an ordinary mutable binding lookup rather than the realm intrinsic. Other intrinsic prototype helpers also require an audit for the same distinction. Engine-generated errors must use the current or explicitly selected callee realm intrinsic, without invoking user getters or depending on public global replacement.

Acceptance: retain and trace realm-owned intrinsic identities using the existing environment/GC ownership facilities; handle lexical shadowing, replacement/deletion, cross-realm errors and moving GC. Preserve explicit prototype overrides and normal user construction/Reflect.construct semantics. New shared state must follow existing initialization/publication and no-GIL synchronization rules. Do not restore a name-based instanceof heuristic or use mutable public bindings as an intrinsic cache. Unknown engine error names need an explicit Error prototype/name policy, not unprototyped objects silently recognized by tags.

This is a separately reproduced existing error-construction defect, not a passing result or completed implementation.

[chrisbbreuer]

Implementation in progress: private root-Environment-owned Error intrinsic identities, traced/relocated through the existing GC visitor and protected by Environment publication/locking. Constructor ordinary calls now use the active constructor as newTarget and enter its realm; temporary roots cover message, cause and iterator callbacks. Fifteen semantic witnesses pass the Node 24.18.0 reference. Debug builds with 2164 linked tests; GC relocation inventory passes (9 kinds, 27 surfaces, 21 tags). Runtime candidate validation waits for the preceding immutable full suite; no pass or completion claim yet. Unknown engine names receive Error.prototype plus their non-enumerable diagnostic name. Non-Error intrinsic lookup, including AggregateError errors-array realm selection, remains an explicitly separate audit boundary.

[chrisbbreuer]

The retained-foreign-realm candidate passes 9/9 Debug and 9/9 unsuppressed ReleaseSafe+TSan; the broader realm filter passes 50/50 (overlapping the nine new tests), with no failures/leaks/races. Corpus remains 302/302 Error/instanceof and 7155/7160 positive class pairs, no measured regressions; default classes 8426/8426. The pre-landing extension audit is now validating optional ConcurrentAccessError registration and explicit WebAssembly error prototypes, with 2167 linked tests. Separately reproduced baseline defects are tracked in #824 (Array identity/active AggregateError realm) and #825 (BigInt primitive prototype rejection), not claimed fixed by this change.

[chrisbbreuer]

Pre-landing gate update: the extensions candidate scores 9/10 in Debug and unsuppressed ReleaseSafe+TSan. The optional ConcurrentAccessError ownership witness fails; all other new Error/GC/explicit-Wasm tests pass, with no leaks/race reports. The untouched VM own-slot/array fast paths can bypass checkRestricted, tracked as #826. Keeping the failing test and holding this change uncommitted until the access paths are diagnosed and fixed; not weakening the test or claiming integration passed. Source hashes, binaries and failure logs are retained.

[chrisbbreuer]

The #826 candidate removes the failing VM access-control bypass without changing or weakening the #823 witness. All ten Error-intrinsic tests now pass: Debug 10/10 (95.653s), unsuppressed ReleaseSafe+TSan 10/10 (4.543s), no failures/leaks/sanitizer reports. Node passes all sixteen local reference cases. The restriction matrix and direct native-artifact tests also pass in both builds. Final checkpoint-boundary coverage and full integration are still pending; no commit or closure claimed.