add_hookcall_monitoring undo breaks when called out of LIFO order #815

🤖 Written by Claude Opus 5.5 via Claude Code for the pluggy maintainers; I prompted it, it did the work, I read it.

add_hookcall_monitoring() returns an undo that restores the _inner_hookexec it saw at add time (src/pluggy/_manager.py, undo() sets self._inner_hookexec = oldcall). So undo only works in strict LIFO order:

• undoing an older monitor also removes every monitor added after it
• a later undo can bring back a monitor that was already removed

This applies to enable_tracing() too, since it is built on add_hookcall_monitoring().

Reproducer:

import pluggy

hs = pluggy.HookspecMarker("x")
hi = pluggy.HookimplMarker("x")

class S:
    @hs
    def h(self): ...

class P:
    @hi
    def h(self): return 1

pm = pluggy.PluginManager("x")
pm.add_hookspecs(S)
pm.register(P())
ev = []

def mon(label):
    return pm.add_hookcall_monitoring(lambda *a: ev.append(label), lambda *a: None)

undo_a = mon("A")
undo_b = mon("B")

undo_a(); pm.hook.h(); print(ev)  # [] -- expected ['B']
ev.clear()
undo_b(); pm.hook.h(); print(ev)  # ['A'] -- expected []

The same thing happens on current main. pytest's own uses (--debug, pytester's HookRecorder) undo in LIFO order, so as far as we know nobody has hit this in practice.

Possible fix: keep the active (before, after) pairs in a list on the PluginManager. Each undo removes its own entry by identity, which makes repeated undo harmless, and then rebuilds _inner_hookexec from the list. Calls already in flight keep the chain they started with.

Found while reviewing #814, which fixed this with a much larger change and was closed.