enable faulthandler in every generated script, not just OOM mode

[devdanzin]

Asked for directly while triaging a PyPy fleet: can fusil record a backtrace on SEGV, so we
capture the problem when it happens?

The gap

faulthandler.enable() was emitted only under --oom-fuzz. In every other mode a
SIGSEGV/SIGABRT left the session's stdout simply stopping at the last call line, with no
indication of where the target died:

[c18m11] Queue._init()
Starting thread: [c18m12] Queue.__eq__()
        <- process gone, that's all we get

This is worst for exactly the crashes that need it most. Two SEGVs in the PyPy stdlib fleet
(multiprocessing.dummy, asyncio.queues) are rare, heavily threaded and load-dependent;
10+ standalone replays never reproduced either, so there will never be a second chance to
observe them. The crash-time dump is the only evidence available.

The change

faulthandler.enable() is now emitted first in the prelude, before any other import, so
a crash while the prelude itself is still running is caught too. Wrapped in try/except —
not every target interpreter ships a working faulthandler. The two OOM blocks no longer
enable it themselves; the prelude covers every mode.

--no-faulthandler opts out, for fuzzing signal handling itself or if it ever perturbs
--tsan race signatures (that campaign's dedup format is a cross-repo contract, so it gets
an escape hatch).

Verified live

PyPy 3.11.15, real generated script, SEGV raised inside a worker thread:

Fatal Python error: Segmentation fault

Stack (most recent call first, approximate line numbers):
  File ".../_ctypes/function.py", line 394 in _call_funcptr
  File ".../_ctypes/function.py", line 299 in __call__
  File ".../session-1/source.py", line 9770 in <lambda>      <- the exact fuzzed call
  File ".../threading.py", line 971 in run

It names the crashing source.py line from inside the thread. Bonus: fusil already scores
Fatal Python error at 1.0, so these sessions stop depending on the signal probe alone.

One trap worth flagging

A bare MagicMock() options stub auto-vivifies a truthy attribute, which reads as
--no-faulthandler and silently suppresses the block — the same hazard
test_tsan_generation already documents in a comment. test_oom_fuzz and
test_target_introspect now pin it explicitly, as they already do for no_numpy.

Five tests cover default-on, ordering-before-other-imports, the try/except guard, that the
emitted prelude still parses (it is spliced into a dedent()ed f-string, where a wrong
indent would break every generated script), and the opt-out. Verified each fails without the
change. Full suite green (1260); ruff clean; golden regenerated.

🤖 Generated with Claude Code