[Guikingone]

Ready for review — ext/intl parity work. The multi-model review findings are fixed, the branch is updated with current main, and the full CI matrix is green on the updated head.

What this PR adds

• A managed ICU 78.3 native package (elephc native add icu) built from the pinned release archive, with a C++ shim owning every ICU object and a fixed C ABI into the elephc-intl bridge crate.
• An AST-only Intl surface (no PHP parsed outside tests): every Intl class is built through synthetic_class, procedural functions through generated declarations, and bridge calls through typed EIR helpers.
• The php-src ext/intl corpus migrated under tests/php_src/intl/, with a generated Rust harness (tests/codegen/intl_php_src_cases.rs) and per-fixture skip statuses.
• Intl object classes for Locale, Normalizer, DateFormatter, TimeZone, Calendar/GregorianCalendar, Collator, NumberFormatter, MessageFormatter, ResourceBundle, Transliterator, Spoofchecker, UConverter, ListFormatter, DatePatternGenerator, IntlChar, IntlIterator/IntlPartsIterator, and the BreakIterator family.

Review fixes (multi-model audit)

A five-model audit of the Intl packet produced confirmed findings; this PR now fixes them all, with regressions in tests/codegen/intl_review_fixes.rs (21 tests):

• ICU handles: Intl wrappers close their bridge handle from __destruct; clone owns a distinct handle, and elephc_intl_live_count() returns to baseline after unset.
• TOCTOU and overflow: the shim registry stores shared_ptrs, so a concurrent close cannot free an object mid-call; oversized bridge payloads are fetched through elephc_icu_intl_fetch_result instead of being truncated to 1 MiB.
• UConverter: the substitution-marker scrub only runs for single-byte destinations, so UTF-16 output such as FFFE 1A01 is preserved.
• IntlDateFormatter::localtime(): reports tm_isdst from inDaylightTime() and no longer dereferences a null SimpleDateFormat; style-only formatters are supported.
• locale_set_default(''): resets to the process system default like PHP, and ini_set('intl.*') alone links the bridge.
• IntlIterator: current()/key() are bounds-guarded and the array-backed iterator keeps its deliberate self-cycle (no destructor), so gc_collect_cycles() still reports it; IntlPartsIterator::getRuleStatus() carries the part's ICU rule status instead of a hard-coded zero.
• NumberFormatter: omitted $type is TYPE_DOUBLE, and ROUNDING_INCREMENT is a double attribute.
• Normalizer: unpublished forms throw ValueError (getRawDecomposition returns "" like PHP) instead of silently normalizing as NFC.
• html_entity_decode(): default/ENT_QUOTES flags and UTF-8 are supported; numeric references still fold under ENT_NOQUOTES when the literal contains no quote entity. ENT_COMPAT, HTML5 doctypes, runtime flags, and non-UTF-8 encodings (or quote entities the runtime cannot decode) are refused at compile time.
• defined(): exact Intl constant membership (including LC_*) instead of prefix matching.
• General compiler fixes: literal str_replace() search arrays apply in source order (the builtin contract stays string-only, so an unlowerable array search is refused with PHP's message), objects at builtin string parameters are coerced only for native SimpleXML views, strtotime() carries $baseTimestamp, large literal range() stops folding past 256 elements, dynamic callable guards evaluate arguments before throwing and use PHP's messages, and unset($object) only defers the safe point for IntlIterator.

Final parity ledger

cargo test --test codegen_tests test_php_src_intl:

434 passed / 0 failed / 48 ignored (previous: 426/8, 425/9, 424/10, 423/11, 412/22, 393/41, 388/46, 378/56, 377/57, 375/59, 373/61, 367/67, 359/75, 351/83, 323/111).

The 48 ignored cases are exactly what php-src itself does on this platform, verified against reference PHP 8.5.10 + ICU 78.3 on macOS ARM64: 47 upstream SKIP (ICU upper-bound gates, 32-bit-only, Windows-only, x86-only, missing optional extensions) and 1 upstream XFAIL (dateformat_format_variant3.phpt). No applicable fixture is skipped by elephc.

Parity work completed in this branch

• MessageFormatter: named-argument formatting through ICU's named entry point, pattern argument types with php's "Inconsistent types" diagnostics, numeric-string/DateTime/IntlCalendar date arguments, default-timezone adoption for date subpatterns, php's parse/format error texts, and the OO/procedural prefixes.
• IntlDateFormatter: TRADITIONAL/GREGORIAN/IntlCalendar calendars, timezone precedence, locale canonicalization and validation, style/calendar/pattern/UTF-8 validation, formatObject error paths, localtime tm arrays, DateTime offset zones, and the procedural aliases.
• NumberFormatter: all ICU styles through unum_open, pattern styles, compact styles, padding/currency text attributes, parse offsets, TYPE_INT64, php's TYPE_* / type-currency ValueErrors, symbol validation, and serialization refusal.
• Collator: SORT_REGULAR/STRING/NUMERIC semantics including whitespace-padded numeric strings, stable ordering, attribute validation, locale-length creation errors, and OO/procedural error prefixes. sort()/asort()/sortWithSortKeys() order by a string key while keeping the original Mixed cells, so object values survive (bug75193) and object __toString side effects keep php-src's collator_sort_modify_during_compare behaviour.
• Locale: compose plural variant/private/extlang arrays, numbered-key handling, lookup tag validation and null-byte diagnostics, locale-length limits, and procedural/global prefixes.
• ResourceBundle: ures_openDirect fallback semantics, sub-table ownership, key iteration, numeric-string keys, localized error contexts, and the fallback warning state.
• Calendar/GregorianCalendar: creation/date/time validation and php's deprecation texts, ICU millisecond clearing, uninitialized comparison arguments, exact Gregorian procedural construction, and IntlCalendar toDateTime error wrapping.
• Grapheme: ICU StringSearch with locale/collation semantics, offset/type ValueErrors, empty-needle rules, negative reverse offsets, grapheme_substr length handling, UTF-8 MAXBYTES/MAXCHARS extraction with next, and collation-aware Levenshtein.
• BreakIterator: debug info (valid/text/typeid), rule-based construction diagnostics and runtime arity/type checks, __clone handle ownership, and code-point method reachability.
• UConverter: user subclasses that override toUCallback()/fromUCallback() run a PHP-side UTF-8 <-> ASCII callback path with php-src's exact callback sequence (reset/illegal/unassigned/close), because ICU's C++ hooks cannot re-enter generated PHP. Converters without overrides and every other encoding keep the managed bridge. The callback marker table is keyed by the bridge handle and cleared on destruction.
• IntlIterator: keeps a self-cycle while alive so a dropped IntlIterator is reclaimed and reported by gc_collect_cycles() exactly like php-src's cycle-management test.
• Normalizer / ListFormatter: raw-decomposition validation, object-conversion errors, UTF-8/range errors, numeric encoding names, and substitution characters.
• ICU bridge: locale canonicalization for formatters, per-call error context/prefix handling, intl_get_error_* status formatting, DateTimeZone unknown-id rejection, and a recipe revision bump (lock revision 40) so the shim is rebuilt.

Compiler/runtime support added for the corpus

• Literal foreach unrolling with per-iteration continue skip flags and distinct source columns, so ReflectionFunction/ReflectionMethod migration loops carry statically-known names; get_extension_funcs("intl") and two-literal range() fold to literal arrays, and literal array facts survive unrelated runtime-state setters (set_error_handler, …).
• Dynamic static-property receivers: $class::$prop / ($expr)::$prop parse through StaticReceiver::Dynamic, the optimizer folds a literal class-string receiver to Named before type checking, and C::$prop =& $local aliases the static slot in both directions.
• ReflectionClass::newInstance()/newInstanceArgs() constructs a statically-known class directly (including $subclass::$ARGS tables); abstract targets and any non-public constructor (declared or inherited) take the reflective path and raise PHP's catchable ReflectionException: Access to non-public constructor of class …. Dynamic callable sites evaluate arguments before raising PHP-shaped ArgumentCountError/TypeError errors, with variadic collector parameters excluded from the per-element type guard.
• gc_collect_cycles() counts unreachable object/callable graph nodes (falling back to total nodes for array-only graphs); unsetting a destructor-less object local only drops a root, so the next explicit collection observes the cycle as PHP does, while classes with __destruct keep the eager safe point that destructor-resurrection paths rely on.
• The WeakMap prelude types WeakReference::get()'s receiver as mixed, so an unrelated class's get signature (Intl's ResourceBundle::get) cannot pin the local and reject the following private WeakMap call. This keeps method_exists() / get_class_methods() on Intl objects working in programs that also inject the WeakMap prelude.
• Dynamic callable sites throw PHP's TypeError/ArgumentCountError classes (not a generic Error), resolve $this and unresolved object receivers without false type mismatches, and leave the language-construct clone() to its descriptor path so argument side effects precede the error. Effectful-key/property and root-drop unsets keep unconditional collection; only destructor-less object locals defer to the next explicit collector.
• Syntactic class constants are mixed until the checker resolves their value, so static defaults like ['x', SomeClass::VALUE] no longer materialize as strings.

Validation

• Review regressions: cargo test --test codegen_tests codegen::intl_review_fixes — 21 passed / 0 failed on the updated head; the full CI run for 945c0d0b46 is green, and the badargs/all-target unit tests now have explicit 300s nextest budgets.
• Contract ledgers after merging origin/main: cargo test -p elephc-builtin-contract — 67 passed / 0 failed.
• Full php-src Intl corpus: 434 passed / 0 failed / 48 upstream skips-XFAIL in CI on 6792ab60be. The harness adaptations restore the fixture INI locale for ini_restore('intl.default_locale') (locale_get_default.phpt) and use a typed-bool helper for normalizer_is_normalized, matching the OO isNormalized results.
• The focused non-ICU regressions pass on the merged head; the ICU-backed module passed 21/21 on the final artifact locally before the merge.
• cargo build is warning-free; git diff --check is clean; generated builtin docs/registry and parity/audit gates are green (Errors: 0; 2881 generated pages validated); the PHP coverage page regenerated to 1265 functions / 191 classes / 1420 constants after the merge.
• libsqlite3-sys is bumped to 0.39.1 for RUSTSEC-2026-0459; cargo test -p elephc-pdo is 140 passed / 0 failed / 2 ignored. The Linux codegen shard cap is raised to 60 minutes (the Intl corpus widened the shards; an aarch64 shard hit the old 40-minute cap).
• origin/main was merged at 67a956b17d (merge commit 3af94a77be), resolving the builtin-count and generated-doc ledgers for the new eval/reflection contracts.
• CI provisions the pinned managed ICU package for the macOS, Linux x86_64 and Linux aarch64 codegen shards. The build-archive steps now live in main’s reusable ci-build-host.yml; elephc-intl is in the shared bridge list there and in the release-artifacts action, and the ICU cache/install steps warm the shards’ key once per platform. examples/intl/ carries ICU recipe revision 40.

Notes

• array_intersect() sparse-key preservation is a separate general gap (elephc-d63l); the Intl case only needs values and is unaffected.
• General by-reference assignment shapes beyond C::$prop =& $local remain tracked by elephc-4ml.12; the static-property form required by bug62915_2 is implemented here.
• Dynamic-callable and dynamic-Intl-surface gaps discovered during the audit are tracked on issue General PHP compiler gaps surfaced by the ext/intl corpus #2276.
• Branch head: 6792ab60be; the PR is mergeable with current main.

[Guikingone]

@nahime0 Ready for review when you have time