Dynamic code: canonicalize flat StringData at parser ingress

Parents: #473, #493, #496 Related: #480

Problem

Dynamic-code entry points currently feed String Values to the lexer through physical asStr() bytes or toStringV() slices. Once prepared flat-Latin-1 storage is enabled, those bytes are one byte per UTF-16 unit rather than canonical WTF-8. Non-ASCII source can then be rejected, mis-tokenized, or trip byte-level assumptions even though its JavaScript String content is valid source text.

Affected shared boundaries include direct/indirect eval, $262.evalScript, ShadowRealm.prototype.evaluate, and Function/GeneratorFunction/AsyncFunction/AsyncGeneratorFunction constructor arguments.

Scope

• Present canonical WTF-8 to every dynamic parser/lexer ingress.
• Borrow existing ASCII and canonical WTF-8 StringCell storage with zero conversion allocation.
• Perform one bounded Latin-1-to-WTF-8 conversion only when physical flat storage requires it.
• Preserve ToString ordering, constructor-realm SyntaxErrors, direct-eval lexical semantics, source locations/debug registration, lone-surrogate handling, and OOM propagation.
• Cover ordinary, realm, ShadowRealm, Function, and generator-family parsing with non-ASCII source.

No-workaround rules

• No representation forcing, replacement character, source escaping, invalid-byte tolerance, parser special case, Test262 exclusion, or hidden permanent canonical copy.
• No stacks, queues, jobs, Promise reactions, Worker, or Map/Set iterator changes.

Acceptance

• Every dynamic parser ingress observes exact JavaScript String content under canonical and flat storage.
• ASCII/canonical input remains borrowed; flat Latin-1 conversion is bounded and OOM-visible.
• Focused tree-walker/VM/realm units, relevant nonzero Test262 accounting, and the full unit gate pass with zero regressions.
• Focused witnesses pass with flat_storage_active disabled and enabled.

Implementation and evidence

• 355329dc establishes canonical WTF-8 at eval, realm evalScript, ShadowRealm.evaluate, and all Function-family parser ingresses.
• Canonical and ASCII StringCells borrow their existing bytes; only physical flat-Latin-1 source takes the existing bounded conversion.
• Focused tree-walker/VM witnesses: 2/2 passed under canonical storage and 2/2 with physical flat storage enabled.
• A fail-first allocator induced OutOfMemory at flat-source conversion before lexing, with 0 successful allocations.
• Test262 dynamic-code subtrees: 620 passed, 0 failed.
• Full unit gate: 1,760 passed, 1 skipped, 0 failed, 0 leaked.

Completed and pushed to main in fix(parser): canonicalize dynamic StringData ingress (#668).

• Direct/indirect eval, realm evalScript, ShadowRealm.evaluate, and all four Function-family constructors now present canonical WTF-8 to the lexer.
• Canonical/ASCII source remains borrowed; only physical flat Latin-1 converts.
• Focused interpreter/VM: 2/2 canonical and 2/2 physical-flat.
• Flat fail-first allocator: exact OutOfMemory, 0 successful allocations, no parser entry.
• Test262: 620/620 across all changed dynamic-code subtrees.
• Full units: 1,760 passed, 1 skipped, 0 failed, 0 leaked.

No stack, queue, job, Promise-reaction, Worker, or Map/Set iterator implementation changed.