Rewrite test framework, expand coverage, and fix isolation bugs - #133

- sources/system/unix/common.cpp: __APPLE__/__FreeBSD__ -> COVSCRIPT_PLATFORM_DARWIN/COVSCRIPT_PLATFORM_FREEBSD
- sources/system/common.cpp: _WIN32 -> COVSCRIPT_PLATFORM_WIN32
- unit_tests/test_system.cpp: _WIN32 -> COVSCRIPT_PLATFORM_WIN32

…nd improve memory management

- function::mContext, struct_builder::mContext: weak_ptr -> context_type*
- function::get_context(): return raw pointer instead of shared_ptr
- Remove resolve_ctx() and process_context::teardown_ctx entirely
- function::call_* (4 paths): use mContext directly, drop lock/throw
- struct_builder::do_inherit/operator(): use mContext directly
- fiber_function::context: weak_ptr -> raw, drop lock/throw
- fiber::create: remove context-alive check
- fiber cs_context (win32/unix): weak_ptr -> raw, remove resume expired check
- Delete 4 tests that relied on 'context destroyed' clean-failure:
  fiber_resume_rejected, escaped_function_call_rejected,
  escaped_function_fiber_create_rejected, escaped_type_constructor_rejected
- Update structure escape test comment (method invocation now UB)
- Update SDK docs: all escaped callable invocation = undefined behavior

- struct_builder: remove m_process (shared_ptr<process_context>) entirely;
  type_node lifetime now relies on context holding the process
- structure: change m_process from shared_ptr to raw context_type*;
  finalizer activation still works because context must be alive when
  finalize runs (script function needs ctx->instance)
- Update test: escaped_structure_data_usable → escaped_structure_member_data;
  remove type_node assertions (now UB after context death), keep member data
- Update SDK docs: structure member data remains valid (self-contained domain),
  but type identity is UB after context death

… pin process)

…ing functions

- Section 1 preamble: remove 'APIs backed by weak references' (all
  weak_ptr removed in ee45862)
- Section 1 footer: remove 'type identity nodes' from safe-escape list
  (process pin removed in 3ca83a3)
- Ownership table: token arena owned by 'instance, functions, and
  struct builders' (not just function store)

…egfault

cb10bfe moved function ownership from shared_ptr to unique_ptr+store,
but left struct methods registering via std::move in run_impl. Since
struct_builder::operator() re-runs method statements for every new
instance, the second instance saw a null mFunc and crashed.

Fix: register named functions and struct methods into the function_store
at compile time (codegen), matching how lambdas already work. The
statement holds a raw function* pointer into the store; run_impl only
binds the callable. This eliminates the repeated-move problem and
aligns the registration semantics for all callable types.

Verified: 335/335 unit tests, 69/69 integration tests (was 65 pass + 4
segfault), ASan clean on Linux.

…larity

Remove the 'structure member data remains valid' exception from the
escape contract. All escaped objects are now uniformly documented as
UB after context destruction, with no self-contained subset promised
to survive.

- SDK.md/SDK-zh.md: remove 'member data remains valid' from §1, §4
- Remove test escaped_structure_member_data_usable_after_context_death
  (tested the removed guarantee)
- Fix stale test comment 'callable's owner' → 'function store'
- Update plan.md to match new uniform contract

Verified: 334/334 unit tests, 69/69 integration tests.

Rewrite machine-translated phrasing in SDK-zh.md to match the project
author's preferred style: direct, concrete, no double-negative boilerplate.
Sync SDK.md to match.

Key rewrites:
- '裸非拥有指针，受 context 生命周期约束' → '指向所属 context 的裸指针，一旦销毁即失效'
- '不保证仍可用' → '一律不可再用' / 'no longer usable'
- '保活机制' → '保持存活机制'
- §1 title: remove '(escaped objects)' parenthetical
- Remove redundant '没有垃圾回收器' parenthetical
- Remove '结果是：每项资源...精确回收' paragraph (redundant with ownership model)

Document the breaking changes introduced by this branch:
- function_ptr owner removal
- contains_callable removal
- escape behavior UB (replaces runtime_error)
- get_context() / m_process raw pointer changes
- named function compile-time registration
- cs::invoke new API

…management and update related logic

test: add tests to ensure finalizers are skipped safely when context is dead

- Remove 6 interactive stdin scripts (choice/hash_map/import/optimize/
  recursion/test_coroutine) from the automatic integration list; files stay
  in tests/ for manual runs
- Per-test timeout: run_tests.sh uses 'timeout' when available; Windows uses
  run_tests.ps1 (raw Process API, WaitForExit+Kill; Start-Process' ExitCode
  is unreliable in Windows PowerShell 5.1)
- Enable expected-output comparison: 59 expected files committed, skip_output
  extended to platform/timing-dependent scripts, CRLF normalization on both
  platforms, drop run_tests.bat in favor of the self-contained run_tests.ps1
- New sanitizers CI job (ASan+UBSan) on ubuntu-22.04 for unit and integration
  tests
- CONTRIBUTING: point Windows test instructions at run_tests.ps1; fix 14
  em-dash mojibake occurrences in CONTRIBUTING.md
- covariant_test.hpp: unique XML testcase names with --repeat, --list honors
  --filter, TEST_F runs TearDown when SetUp throws
- test_system.cpp: redirect system.run echo output per platform
- .gitignore: ignore unit test artifacts; stop tracking plan.md (internal
  planning doc)

dlopen with RTLD_DEEPBIND makes the process exit silently under ASan
(google/sanitizers#611), so the extension-loading unit tests would abort
the run before the summary line. The sanitizers CI job now excludes them
with --exclude=extension_load.

matches_filter with an empty pattern returns true, so the exclude check
rejected all tests when --exclude was unset.

GCC 12's libasan misreports a stack-buffer-overflow in the sigaltstack
interceptor when a no-return path (std::exit from system.exit) executes on
a fiber (ucontext) stack that was never registered with ASan. GCC 13 does
not exhibit this. The main build matrix keeps ubuntu-22.04 for GCC 12
coverage of normal builds.

- test_set.csc: sort collected items before printing (hash iteration order
  depends on the std::hash implementation, libstdc++ vs libc++)
- to_string.csc: use type() (cs::ostream, consistent across platforms)
  instead of typeid to_string (C++ demangled name); print hash_map entries
  via sorted keys() and hash_set via sorted items
- limit.csc / numeric.csc: long double extremes depend on the architecture
  (x86-64 80-bit, aarch64 128-bit, macOS arm64 64-bit double) - move to
  skip_output (exit code only) and drop their expected files

std::ios_base::seekdir/openmode are typedef int in the MSVC STL but
distinct enum types in libstdc++, so type() output differs (int vs
cs::iostream::seekdir). Enum type names are a compiler STL detail; the
remaining 18 type() assertions are verified identical on GCC, MSVC and
Clang.

…ents

- Remove no-op --exclude=serial_execution from sanitizer job (env auto-detect already covers it)
- Add if-no-files-found: ignore to unit test report upload
- Delete contains_callable migration-guide bullet (never shipped, cited non-existent APIs)
- Correct structure finalize: instance -> process; add m_ctx==nullptr skip note
- Add non-atomic refcount caveat to English SDK docs (already in Chinese)
- Update stale arena<->function cycle comments in impl.hpp and test_fixes.cpp

- Remove is_script_callable detection from run_finalize; guard on m_ctx
  instead of script/native distinction (same behavior, -15 lines)
- Add storage_transaction RAII class (runtime.hpp) with commit/rollback/discard
- Replace m_snap/m_committed single-slot members with m_tx vector stack,
  fixing re-entrant exec() snapshot overwrite (parallel to m_units)
- compile() reuses storage_transaction; eliminates manual snapshot lifecycle
- Update SDK docs: unified finalize behavior (script/native both scoped
  under context process; skipped with diagnostic when unavailable)
- Verified: 686 unit tests + 63 integration tests pass