The compatibility baseline is stable
The latest complete coverage measurement is dated 16 September 2026: 2,655 Zend tests pass (65.3% of runnable tests), the core language standard library is 539 / 654 — 82%, and phpr registers 1,017 / 2,143 internal functions.
WordPress remains the strongest application proof. The complete single-site suite contains 30,472 tests and multisite 31,278; each differs from the PHP oracle by one deliberately catalogued test name. WordPress 7.0.1 installs and serves on real MySQL through native MySQLi, with front pages, login, REST, wp-admin, pretty permalinks and media paths covered by byte-parity evidence.
The stable performance scoreboard
Performance is now managed through promoted pins rather than isolated benchmark screenshots. Stable pin s183, promoted on 16 October 2026, records the following R=5 micro ratios against the PHP oracle:
| Category | Stable ratio | August reference |
|---|---|---|
| Arithmetic | 2.1× | 9.3× |
| Property access | 2.3× | 7.9× |
| Function calls | 3.5× | 5.1× |
| Strings | 4.3× | 5.3× |
| Arrays | 2.9× | 3.9× |
| Regex | 2.5× | 3.5× |
The latest clean paired measurement of the full WordPress suite is 1.747× of the oracle. That number was measured on the preceding stable pin s182, so it is intentionally reported separately from the later s183 micro scoreboard. Doctrine ORM remains a harder performance workload at roughly ~7.0×.
What actually moved the numbers
The optimization path is now evidence-driven. Work across the performance arc has included a Zend-style cycle collector over objects and containers, in-node walk marks, re-keying reflection descriptors on their declaring class, growable strings with in-place append when uniquely owned, a keyless hashed-array index, compile-side stub elision, and sealed fast forms that keep hot arithmetic and property operations away from unnecessary temporary Zval lifecycles.
A key result from the measurement program is that raw dispatch is no longer the main story. The remaining gap lives largely inside hot handler bodies: temporary values, references, call/return frames and the semantic work performed around each operation. That finding changed the route from broad VM redesigns to smaller, separately judged transformations.
Performance work is treated as an experiment
The project now has an explicit protocol in REGOLE.md ↗. A lever starts with a stated mechanism and acceptance criterion; A/B runs are interleaved; compatibility is checked against a frozen fail-set by name in multiple execution modes; fixtures are run against both engines; disassembly or mechanism checks are used when relevant; and a promoted binary is rechecked on real workloads.
The stable Rust battery is 1,748 tests, supplemented by sentinel gates. The intent is not merely to report a lower number, but to know why it moved and what semantic boundary was preserved.
Correctness continued after the stable pin
Repository work through 22 October 2026 also closed semantic gaps after s183: recursive object comparison now has a Zend-style recursion guard, and destructor timing at function leave was moved closer to the reference behavior. Three named Zend failures — bug63882.phpt, bug49893.phpt and generators/bug71013.phpt — flipped to pass with no new failures in the frozen gate, reducing that frozen fail-set from 1,412 to 1,409 per mode.
Important measurement distinction: the public global coverage figure remains 2,655 because it comes from the last complete measurement on 16 September 2026. The October fixes are real repository progress, but this site does not add three to the global total until a new promoted full corpus measurement is published.
What comes next
The explicit target is 1× CPU parity with the oracle; ≤3× is an intermediate stage, not the finish line. Calls and strings remain the largest micro categories on the stable pin, while WordPress and Doctrine continue to act as reality checks against local micro wins.
Laravel is queued as the next broad framework validation target after the performance front. Remaining extension surfaces continue on demand, while single-binary distribution and native asynchronous I/O remain longer-term architectural directions.
Sources: README ↗ · COVERAGE.md ↗ · REGOLE.md ↗ · PERF_MAP.md ↗ · current route ↗ · repository snapshot ↗