fix(fspy): record file accesses made inside signal handlers

[wan9chi]

Motivation

The HANDLING_OPEN thread-local guard (#540) suppressed same-thread re-entry into handle_open because, at the time, resolving and reporting an access called interposable libc symbols — getcwd, readlink, fcntl through nix, plus CString::new(format!(...)) — and on Linux an LD_PRELOAD library's own PLT calls resolve to its own exported interposers, recursing until the traced process overflowed its stack.

That vector no longer exists, on either platform for its own reason:

• Linux: the handler's entire call graph is now raw syscalls — path resolution through fspy_nostd's rustix/linux_raw wrappers, fd formatting through itoa, allocation through the mmap-backed pooled bump, reporting through atomics on the shared mapping. Nothing binds through the PLT.
• macOS: the handler still calls libSystem symbols, but interposition is dyld's __interpose section, and dyld never applies interposing tuples to the image that provides them — the same exemption every original() forward has always relied on.

Dropping the guard also fixes a real gap: it silently discarded legitimate accesses made by signal handlers that interrupt an in-flight interception — exactly the case the signal-safe allocator work exists to support — and it removes a std TLS dependency from the hot path, which the std-free preload roadmap has to shed anyway.

A comment on handle_open now records the invariant this rests on: the handler must stay free of bindable libc calls on Linux, and the macOS argument holds only under __interpose-style interposition.

🤖 Generated with Claude Code

[github-actions]

fspy benchmark

linux

dynamic/launch             change  -0.51%  [-11.42% .. +11.42%]  overhead  +281.88%
dynamic/access             change  +0.37%  [-13.83% .. +14.77%]  overhead   +10.58%
dynamic/access-relative    change  -0.60%  [-22.09% .. +23.24%]  overhead   +47.19%
dynamic/access-contended   change  +0.47%  [-14.56% ..  +9.60%]  overhead   +12.80%
static/launch              change  +0.71%  [ -9.17% .. +14.56%]  overhead  +697.50%
static/access              change  +4.46%  [-11.36% .. +19.53%]  overhead  +875.67%
static/access-relative     change  -0.02%  [-17.95% .. +11.36%]  overhead +1518.30%
static/access-contended    change  -0.99%  [-11.35% ..  +8.45%]  overhead +2080.50%

macos

dynamic/launch             change  -0.14%  [ -6.25% ..  +5.37%]  overhead  +244.62%
dynamic/access             change  +0.00%  [ -4.42% ..  +5.16%]  overhead    +1.79%
dynamic/access-relative    change  +1.48%  [-14.30% .. +16.90%]  overhead  +267.58%
dynamic/access-contended   change  +0.74%  [-43.17% .. +13.87%]  overhead    +1.08%

windows

dynamic/launch             change  +2.93%  [ -9.57% .. +15.85%]  overhead   +26.09%
dynamic/access             change  -1.95%  [-26.63% .. +10.24%]  overhead    +1.19%
dynamic/access-relative    change  +2.20%  [-11.53% .. +21.68%]  overhead    +3.81%
dynamic/access-contended   change  +1.38%  [-13.99% .. +25.34%]  overhead    +3.67%