Second of #739's items. Item 1 (the trace tape) landed in #761; this is the one flagged as most severe of the remainder.

The bug

exit is deliberately uncatchable: builtin_exit sets g_exit_requested, and CHECK_ERROR consults it to refuse routing the unwind to a try handler. But the flag was a process global that nothing ever reset — not eigs_close, not eigs_state_destroy, and main leaves it set on purpose. So once any script in any state called exit, every later eigs_eval_string in the process ran with exception handling silently disabled: a raise inside try went to vm_error_halt instead of the catch handler.

Reproduced through the embed API — unreachable from the CLI, where exit ends the process, which is why it had never been seen:

before exit -> CAUGHT
after  exit -> (null)        <- catch handler never fires again

For a long-lived host running untrusted snippets — the sandbox_run case, and the seam EigenOS M11 consumes — one line of script permanently corrupted the semantics.

The fix

g_exit_requested/g_exit_code move onto EigsThread, beside the g_has_error/g_try_depth that CHECK_ERROR reads them with — they were the only members of that flag family still process-global. The request is cleared at host eval entry, using the idiom repl.c:67 already had for g_has_error ("don't carry a prior line's error into this one"). It applies to the eval that raised it and no later one. exit stays uncatchable.

The exit code is additionally latched on EigsState. Per-thread alone silently dropped exit of N inside a spawned worker to 0 while main returned success — a regression this change introduced, which the suite had no coverage for. The thread flag drives the uncatchable unwind; the state latch decides the process's status. Side benefit: a worker's exit no longer erases a genuine main-thread error.

Two reader-side fixes fall out, both the same root cause — a bridge macro is not a process global

• main.c's REPL path read the flag after eigs_thread_detach(), leaving nothing to read it through; the REPL returned 0 instead of the requested code. It captures before teardown now, as the script path already did. Every other read was audited — all are on attached threads.
• eigenscript.h's comment still said "Process-global (exit terminates the whole process)" directly above the new note contradicting it. Merged into one block.

Validation

Both new checks were validated red before being trusted:

The embed check ships a before-exit control that passes in both runs, so the check isolates the right thing rather than the harness.

Behavior verified unchanged: exit of 7 → 7, exit inside try → 3 (still uncatchable), REPL exit of 5 → 5, worker exit of 9 → 9.

• release 3274/3274
• asan+ubsan 3278/3278, asan-http 3374/3374 — leak tally 0 in both
• freestanding symbol gate stage 1+2 OK; embed smoke OK; exit suite 6/6

Refs #739 — items 3 (g_task_suspend_request global while its scheduler is per-thread) and 4 (s_osr_threshold, g_sandbox_*, g_stream_file, g_model, g_db_conn) remain open.

🤖 Generated with Claude Code