Ship the turbo extension as a shared core plus a thin extension per PHP version - #6703

The core no longer references anything in the version-specific library:
the pt_abi table, the module-startup registration list and the shadow
plan registry moved into the new Core.cpp, and the functions Abi.cpp
implements per minor became function pointers in pt_abi. The few core
functions the version-specific library calls are PT_CORE_API, everything
else stays hidden.

PHP 8.6 broke the remaining assumptions, now read through pt_abi too:
- zend_class_entry gained ce_flags2 after ce_flags, moving every later
  member (PT_CE(), type-checked to take a zend_class_entry only)
- the ZPP_ERROR_* codes became a uint8_t enum without WRONG_COUNT and two
  Z_EXPECTED_* types were inserted; the parse slow paths now return the
  value (neutral ZPP macros, translated by Abi.cpp)
- zend_is_callable, zend_call_known_function, _php_stream_free and
  _php_stream_read are no longer exported
- zend_arg_info grew a doc_comment (PT_ARG_INFO())
and ZEND_COLD is pinned, since the older headers enable it for GCC only.

bin/shared-core/gate-linux.sh builds in the CI images (GCC, or clang with
GATE_IMAGE/GATE_MAKE_ARGS), link-split-linux.sh links one core from the
8.4 objects plus a 68 KB extension per minor and runs smoke.php with each:
ALL OK on 8.3, 8.4, 8.5 and 8.6. The macOS scripts got a -mac suffix.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017h1GfB7vfJPgwcTHD9Fth6

The last differences between the header sets, each pinned in abi.h or
moved into Abi.cpp:
- the fresh-C-stack switch (a zend_fiber_context, its switch and a
  zend_try, whose SETJMP became sigsetjmp in 8.6) is implemented per minor
- codegen pins for helpers 8.6 re-spelled without changing what they do:
  zend_string_equals_literal_ci / _equals_ci / starts_with_literal(_ci),
  ZSTR_IS_INTERNED, array_init(_size), Z_TRY_ADDREF_P / Z_TRY_DELREF_P
- the comparison treats the zend_result rename (a different mangling of
  the same enum) and the symbol/section name tables as names, not code

clang 14 in the CI images (gate-linux.sh with GATE_IMAGE=turbo-build-clang,
-fgnuc-version=4.3 so the older headers enable ZEND_COLD for it too) now
finds no differing function, data section or relocation outside Abi.o,
Shadow.o, TrustedTypes.o and main.o across 8.3, 8.4, 8.5 and 8.6. GCC
still differs in 15 string-building functions (8.6 const-qualified
zval_get_string() and friends, inline helpers compiled inside the headers).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017h1GfB7vfJPgwcTHD9Fth6

…ake SPLIT=1

`make SPLIT=1` builds phpstan_turbo_core.so (everything but the four
version-specific sources) and a phpstan_turbo.so linked against it, which
finds it next to itself ($ORIGIN, @loader_path); `make SPLIT=1 extension`
links only the extension against a core already in place, which is what a
per-PHP-version CI leg does. `make pgo` and `make strip` handle the pair; a
plain `make` (and the phpize build) still produce the one self-contained
extension.

The boundary is functions only — pt_abi, pt_globals and pt_type_op_infos
are reached through accessors from the version-specific side, since a
Windows DLL can delay-load functions but not data — and main.cpp is
version-specific like Abi/Shadow/TrustedTypes. The extension refuses a
core of another version: it registers nothing and reports "core-mismatch",
which TurboExtensionEnabler rejects. On Windows it loads the core from its
own directory in get_module() (PHP loads extensions with plain
LoadLibrary, which would search the application's directory instead), and
zend_is_true() goes through pt_abi there (no assembler names on MSVC).

Verified in the CI images: a PGO+LTO core built in the 8.5 image (6.73 MB
stripped) with extensions built against it in the 8.3, 8.4, 8.6 and 8.6
ZTS images (68 KB each): smoke.php ALL OK on each; a version-mismatched
pair reports core-mismatch and stays inactive. macOS: the same via make
SPLIT=1, ALL OK on 8.5.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017h1GfB7vfJPgwcTHD9Fth6

…HP version

phar.yml builds one profile-guided core per platform (turbo-compile-core,
-core-musl-arm64, -core-windows for NTS and TS; the jobs turbo-origins now
reuses from earlier runs) and links each PHP version's extension against
it in the existing turbo-compile* legs, whose artifacts carry the core
next to the extension so every test job finds it there.
.github/scripts/turbo-dist-layout.sh maps the artifacts onto the dist
layout (turbo-ext/<platform>/phpstan_turbo_core.so|.dll next to
phpstan_turbo-<minor>.so|.dll) for the aggregate artifact and the commit
job, refusing legs of one platform that carry different cores.
turbo-shared-core-gate (bin/shared-core/check-linux.sh) compiles the shared
sources against every supported PHP's headers with clang and fails on any
difference outside the four version-specific sources; the commit job waits
for it.

config.w32 gains --enable-phpstan-turbo=core / =extension; the extension
DLL delay-loads the core from its own directory. TurboExtensionSelector
offers no binary when the core is missing next to the extension. The
Makefile's default goal stays the extension (it had become `objects`), and
a plain `make strip` leaves a core put in place untouched.

README.md documents the layout, the mechanisms that keep shared code
version-neutral, the gate, and the procedure for supporting a new PHP
version; turbo-ext/CLAUDE.md the rules for code in shared sources.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017h1GfB7vfJPgwcTHD9Fth6