Fix evented exec and long-lived scope waiter retention

[lePereT]

What type of PR is this? (check all applicable)

• Bug Fix
• Documentation Update
• Code Refactor
• Performance Improvements
• Test

Description

This PR fixes two backend/runtime issues in lua-fibers and adds regression coverage for both.

First, it fixes unbounded memory growth in long-lived scopes caused by cancelled one-shot waiters being left as tombstones. Scope-aware blocking operations race their main operation against the scope cancellation/fault one-shots. When the main op wins, the losing one-shot waiter is cancelled. Previously, cancellation cleared the callback but left the waiter record in the one-shot waiter array. Long-lived scopes performing repeated stream operations could therefore accumulate large numbers of cancelled waiter records. This was visible during repeated upload-shaped stream transfers.

Second, it adds an event-driven luaposix process backend for LuaJIT + luaposix. The existing SIGCHLD/self-pipe backend remains appropriate for PUC Lua, but LuaJIT is unsafe with a Lua-level SIGCHLD handler in this path. The new posix_reaper backend keeps the parent scheduler fully evented by using a lightweight reaper process and a sentinel pipe, similar in shape to the existing nixio compromise.

Changes included:

• Add exec_backend.posix_reaper for LuaJIT + luaposix.

  • Uses parent -> reaper -> child.
  • Reports child completion to the parent over a sentinel pipe.
  • Keeps the parent scheduler event-driven through the existing poller.
  • Avoids LuaJIT signal-handler instability.

• Update exec backend selection.

  • pidfd remains preferred where available.
  • sigchld remains the PUC Lua luaposix backend.
  • posix_reaper is selected for LuaJIT when using the posix family.
  • nixio remains the nixio-family backend.

• Fix one-shot cancellation retention.

  • Cancelled one-shot waiters are now physically unlinked rather than left as tombstones.
  • Signalling takes ownership of the waiter array before invoking callbacks, so cancellation during signalling remains safe.

• Add backend-family cycling tests.

  • Runs relevant I/O tests in separate subprocesses for each backend family.

  • Prevents silent mixed-family fallback.

  • Covers:

    • ffi fd/poller/exec family
    • posix fd/poller/exec family
    • nixio fd/poller/exec family

• Add one-shot waiter regression coverage.

  • Verifies cancelled one-shot waiters do not accumulate.

• Add a small upload-shaped regression test to the standard suite.

  • Repeated UNIX-socket to stream to file transfers inside a long-lived scope.
  • Asserts scope cancellation/fault one-shot waiter arrays do not grow.
  • Kept deliberately small by default.

Related Issues, Tickets & Documents

No linked issue.

Relevant context:

• Long-lived scope memory growth was observed during upload-shaped stream workloads.
• The issue was traced to cancelled one-shot waiters accumulating on long-lived scope cancellation/fault one-shots.
• LuaJIT with luaposix SIGCHLD handling was found to be unstable enough to avoid a Lua-level signal callback for process completion.

Screenshots/Recordings

Not applicable. This PR has no visual changes.

Manual test

• yes

Manual test description

Manually tested with both LuaJIT and PUC Lua.

Commands run:

cd tests
luajit test.lua
lua test.lua

Both complete successfully, including:

• existing unit tests
• I/O tests
• exec tests
• backend-family subprocess tests
• one-shot regression tests
• upload-shaped regression tests
• scope tests

Added tests?

• yes

Added tests:

• backend-family cycling tests
• one-shot waiter cancellation regression
• small upload-shaped long-lived-scope regression
• optional larger upload stress harness

The upload-shaped test can be controlled with:

LUA_FIBERS_UPLOAD_STRESS_MB=4
LUA_FIBERS_UPLOAD_STRESS_REPEAT=3
LUA_FIBERS_SKIP_UPLOAD_STRESS=1

Added to documentation?

• no documentation needed

README changes are not required for the public API surface.