Errors: class early errors and private-name diagnostics diverge from JavaScriptCore

[chrisbbreuer]

Parent: #459
Related: #899

Problem

A 66-case differential over class, private-name, super, TDZ, Promise, async and
iterator errors matches 34/66. Two clusters account for most of it.

Class early errors have no prose at all (8 cases)

Duplicate private name, undeclared private name, a private name outside a class,
delete this.#x, await in a static block, a field named constructor,
#constructor, and super outside a method all collapse to the SAME string:
eval: UnexpectedToken at <line>:<col>.

Two structural problems: there is no per-rule message, so eight distinct spec early
errors are indistinguishable; and the string carries an eval: prefix and a line:col
that JavaScriptCore never puts in .message. This is the class-specific face of #899 —
fixing #899's reason-discriminant mechanism should cover these, and they are a good
first set of rules to carry through it since the parser already detects each one
correctly at the right offset.

Private-name runtime messages need a split AND a merge (6 cases)

JavaScriptCore splits by what the private name IS, where zig-js uses one string:

brand missing on	JavaScriptCore
a field	Cannot access invalid private field
a method or accessor	Cannot access private method or acessor

Note JSC's own typo — one c in "acessor". Exact parity requires reproducing it.

In the other direction zig-js is more specific than JSC, and matching means merging.
For a present brand with a missing accessor half, JSC uses two fixed strings —
Trying to access an undefined private getter / ... private setter — and uses the
setter string even for writing a private method. zig-js emits three more
descriptive strings (Cannot write private method, Cannot write private member: accessor has no setter, Cannot read private member: accessor has no getter).

That merge loses information, so it is a deliberate call rather than an obvious fix:
worth confirming that Home/Bun code actually matches on these before giving up the more
useful wording.

Verification

scratchpad/sweep/probe.sh, against a real JavaScriptCore backend.

[glennmichael123]

Landed the non-lossy runtime half in b6cff0b1 (fix(class): distinguish private brand failure kinds).

The runtime PrivateEnvironment now carries each private binding's declaration kind together with its evaluation-unique storage key. Missing-brand reads/writes therefore preserve JSC's distinction:

• field: Cannot access invalid private field
• method/accessor: Cannot access private method or acessor

This is lexical metadata, so it remains correct through nested classes, closures, direct eval, generators, and forced VM execution; it does not guess from the unbranded receiver. Added eight focused tree-walker + VM cases and retained map collision/OOM coverage.

Verification:

• focused diagnostic unit: 1 passed, 0 failed, 0 leaked
• PrivateNameMap unit: 1 passed, 0 failed, 0 leaked
• test/language/statements/class: 4,367 passed, 0 failed

The information-losing merge for present-brand missing getter/setter/method writes remains intentionally open, as do the parser early-error prose items shared with #899.

[glennmichael123]

Landed f14bb72b (fix(class): retain private access diagnostic source).

Private brand errors now retain the owning member/assignment source through reads, writes, updates, compound assignments, calls, optional calls, tagged templates, nested classes, and direct eval. The bytecode and native property paths carry the same EvaluationSite and route private names through their brand/descriptor checks before ordinary inline caches.

Evidence:

• Existing 66-case JSC diagnostic differential: 34/66 -> 37/66 exact matches.
• test/language/statements/class: 4,367 -> 4,367 passed, 0 failures; flips 0 test262 cases.
• Expanded focused regression: 14 cases in tree-walker mode and required bytecode with JIT disabled/enabled; 1 test passed, 0 failed, 0 leaked.
• A separate GC-sensitive private-eval integration filter was terminated after the local non-collecting zig-gc stub emitted excessive leak diagnostics. It is not counted as passing evidence.

The parser-reason work in #899 and the remaining class diagnostics here stay open. Next slice: explicit class early-error reasons and class-heritage diagnostics, preserving error categories and evaluation order.

[glennmichael123]

Landed b34796fe (fix(parser): preserve class diagnostic reasons and locations). Related: #899, #904.

The parser now records 17 explicit reasons for invalid class declarations. These distinguish field/method/accessor collisions, static versus instance accessor conflicts, forbidden constructor/prototype declarations, and duplicate constructors. Declaration offsets are temporary parser scratch, and the grammar's existing acceptance rules remain intact.

eval and Function render these reasons as JSC-compatible prose. Location stays in non-enumerable line, column, and byteOffset properties. Errors inside nested tagged/untagged template substitutions retain the reason and translate normalized CRLF offsets back into the original source. Shared class-heritage validation also uses JSC's two distinct messages without changing evaluation order.

Measured evidence:

• New declaration/heritage differential: 79/79 exact JSC matches, including every pair of field/method/getter/setter declarations across static and instance placement.
• Existing broad diagnostic differential: 37/66 -> 43/66 exact matches in this commit (34/66 -> 43/66 across the two latest commits).
• Class test262 subtree: 4,367 -> 4,367 pass, zero failures; flips 0 tested cases.
• 55/55 parser tests, plus the two focused diagnostic tests: zero failures and zero leaks. Focused checks cover eval/Function messages, declaration offsets, legal accessor pairs, and nested CRLF template provenance.

Both issues remain open. Undeclared private-name lists, private delete/static-await/super diagnostics, other parser/lexer reasons, and the deliberate present-brand accessor-message merge remain separate work. This is not a claim that all SyntaxErrors now match JSC.

[glennmichael123]

Production-dependency verification update: an isolated clean checkout of bff189a, real zig-gc 0285ef801dd221a4ae8e88b31a09ae31627657a3 and zig-regex c60344dec54f71efe885c62a02e29dfd7ab0f5c0 now passes both GC-sensitive integration filters: private eval contexts survive function escape and suspension (24 scenarios across tree-walker/required VM), and private eval contexts retain suspended receivers across moving collection. Each filter reports 1 passed, 0 failed, 0 leaked. The original local GC shim was not modified. This replaces the prior unverified status for these two filters only, not the full suite. No timing/performance claim: another Home benchmark overlapped because the job used a different lock path; that scheduling mismatch has been identified and future jobs will use the canonical shared lock. Memory guard stayed enabled throughout (3500 MB peak conservative tree footprint, host low-water 33%).

[glennmichael123]

Private-deletion slice landed as #919 in 6256244. It fixes eight optional-private early-error violations and the specific direct-private diagnostic without rejecting valid outer public properties/call results. Focused units and six affected corpus subtrees passed; full verification scope is recorded in the commit and #919. This parent stays open for the remaining parser diagnostics and previously discussed accessor-message decisions.

[glennmichael123]

Landed the semantic early-error slices #920 and #921 in 41104a4. The shared scanner now visits expressions inside destructuring patterns and logical assignments/dynamic imports, and rejects await iteration/asynchronous disposal in static blocks while preserving valid function boundaries. Combined evidence: 25 distinct early-error gains in both execution modes; 72 focused units passed without failures or leaks; 7,462 tested corpus cases retained passes with zero before/after failures. Exact filters and scope limits are in the commit. This parent remains open for remaining diagnostic wording/location work and the previously deferred accessor-message decisions.