Test262: RegExp Symbol.* protocol — 7 remaining spec-edge cases (unicode lastIndex, replacer this, cross-realm)

[nickna]

Motivation

Follow-up to #101, which hardened the RegExp.prototype[Symbol.*] protocol to ECMA-262 §22.2.5 across both modes. That issue is now effectively complete — the Symbol.* surface passes 204/207 interpreted and 201/207 compiled (the remaining 9-per-mode skips are the separate regexp-named-groups feature).

This issue carves out the 7 distinct spec-edge test files that still fail, so the remaining work has an explicit, closeable target.

Remaining failing tests

Interpreted (3):

• test/built-ins/RegExp/prototype/Symbol.replace/coerce-global.js — RuntimeError
• test/built-ins/RegExp/prototype/Symbol.replace/fn-invoke-this-no-strict.js
• test/built-ins/RegExp/prototype/Symbol.split/splitter-proto-from-ctor-realm.js — RuntimeError

Compiled (6, two shared with interpreter):

• test/built-ins/RegExp/prototype/Symbol.match/builtin-infer-unicode.js
• test/built-ins/RegExp/prototype/Symbol.match/builtin-success-u-return-val-groups.js
• test/built-ins/RegExp/prototype/Symbol.search/u-lastindex-advance.js
• test/built-ins/RegExp/prototype/Symbol.split/u-lastindex-adv-thru-match.js
• test/built-ins/RegExp/prototype/Symbol.replace/fn-invoke-this-no-strict.js (also fails interpreted)
• test/built-ins/RegExp/prototype/Symbol.split/splitter-proto-from-ctor-realm.js — RuntimeError (also fails interpreted)

Work clusters

1. Unicode-flag lastIndex advancement (4 compiled tests) — highest leverage

• Symbol.match/builtin-infer-unicode.js
• Symbol.match/builtin-success-u-return-val-groups.js
• Symbol.search/u-lastindex-advance.js
• Symbol.split/u-lastindex-adv-thru-match.js

The compiled exec loop needs surrogate-pair-aware index stepping when the u flag is set (§22.2.7.3 AdvanceStringIndex). The interpreter already passes these, so the C# reference exists — this is an IL port of the unicode advance step.

2. Replacer-fn this in non-strict mode (1 test, both modes)

• Symbol.replace/fn-invoke-this-no-strict.js

The replacer function passed to Symbol.replace must be invoked with undefined this, which in non-strict mode coerces to the global object. Both modes currently mishandle the this binding.

3. Cross-realm prototype provenance (1 test, both modes)

• Symbol.split/splitter-proto-from-ctor-realm.js

The species-constructed splitter must take its prototype from the constructor's realm. This likely needs realm support SharpTS does not model; may be out of scope — acceptable to mark Skipped with a feature tag if realms aren't planned.

Also (interpreted-only)

• Symbol.replace/coerce-global.js — RuntimeError; likely related to cluster 1's global/unicode flag handling on the interpreter side.

Acceptance

• Clusters 1 and 2 flip to Pass in both modes (6 of 7 files).
• Cluster 3 either passes or is explicitly skipped with a realm-feature tag and a note.
• No regressions in RegExp/prototype/* or String/prototype/* baselines.

Related

Follow-up to #101. Part of #69.

[nickna]

Deferred — JS (Test262) conformance track de-prioritized.

SharpTS's core goal is TypeScript language compatibility first. The Test262 / ECMA-262 runtime-conformance work has diminishing returns for that goal, so this issue is being set aside (label: deferred). It is not closed — it remains a valid, tracked item and can be reactivated if priorities change.

Active focus moves to the TypeScript conformance track (#80 and children) and genuine correctness bugs.

[nickna]

Root-cause triage (current main)

Reconfirmed all 7 still fail and traced each. None is a stale/quick win — each needs infrastructure SharpTS doesn't yet have or a deep semantic change, so documenting precisely rather than landing a rushed fix.

Cluster 1 — unicode-mode matching (Symbol.match/builtin-infer-unicode, builtin-success-u-return-val-groups, Symbol.search/u-lastindex-advance, Symbol.split/u-lastindex-adv-thru-match) — now fails in both modes

Two intertwined needs:

• Surrogate-aware matching. /\udf06/u[Symbol.search]('𝌆') must be -1 (a lone low-surrogate pattern must not match the trailing half of a surrogate pair under u); SharpTS returns 1. This is the §22.2.7.3 AdvanceStringIndex / unicode code-point matching behavior and overlaps the broader shorthand/unicode regex-semantics gap in RegExp: \w/\s/\W/\S ECMAScript shorthand classes diverge from JS (.NET engine quirks) #693 — .NET's engine matches by UTF-16 code unit here.
• Reading the overridden unicode flag. builtin-infer-unicode does Object.defineProperty(r,'unicode',{value:…}) then expects [Symbol.match] to infer mode from that override.

Cluster 2 — replacer this in non-strict mode (Symbol.replace/fn-invoke-this-no-strict, both modes)

The replace fn must be called with this = undefined, which in sloppy mode coerces to the global object; the test asserts thisVal === this (top-level this). Probe: the replacer's this is neither globalThis nor undefined, and doesn't equal top-level this. This is the sloppy-mode this = globalThis binding gap (also seen compiled where an unbound this is null, surfaced in test262 language/expressions/call/11.2.3-3_8). Broad change; deferred.

Cluster 3 — cross-realm (Symbol.split/splitter-proto-from-ctor-realm, both modes)

Uses $262.createRealm() + new other.Function(...) and asserts cross-realm prototype provenance. SharpTS has no realm model — out of scope (as the issue anticipated). Recommend marking Skipped with a realm-feature tag.

Symbol.replace/coerce-global (interp RuntimeError)

The RegExp logic is actually correct now (r[Symbol.replace]('aa','b') → "ba" with an overridden falsy global). The RuntimeError comes from the test's first line — Array.print = print — i.e. property assignment on the built-in Array constructor throws in the interpreter (a separate gap, unrelated to the Symbol.* protocol).

Net

• Clusters 1 (regex unicode/surrogate semantics) and 2 (sloppy-this) are the real protocol-adjacent work; both are broad and tracked alongside RegExp: \w/\s/\W/\S ECMAScript shorthand classes diverge from JS (.NET engine quirks) #693.
• Cluster 3 is out of scope without realms.
• coerce-global is gated by built-in-constructor property assignment, not the replace protocol.