V fork: sync.Mutex.lock() silently ignores pthread_mutex_lock failure — uninitialized mutexes provide no exclusion (Darwin)

Problem

vlib/sync/sync_default.c.v:

@[inline]
pub fn (mut m Mutex) lock() {
	C.pthread_mutex_lock(&m.mutex)
}

The return value is discarded. On Darwin, pthread_mutex_lock on a
zero-initialized pthread_mutex_t returns EINVAL (Darwin's
PTHREAD_MUTEX_INITIALIZER carries _PTHREAD_MUTEX_SIG_init, not all-zeros),
so a sync.Mutex that was never init()'d — e.g. any VALUE-typed __global,
which _vinit zeroes — becomes a silent no-op lock: no error, no panic, no
mutual exclusion. On Linux/glibc zeroed storage happens to equal
PTHREAD_MUTEX_INITIALIZER, so the same code locks correctly there — the bug
ships only on macOS and presents as an "intermittent flake".

Field impact (cx-private vlang#303)

Two SSE subscriber-registry locks were value-typed __global sync.Mutex.
Under parallel-suite CPU load, two executor threads interleaved inside the
"critical section" and a concurrent array append lost an element (server
trace showed both subscribes printing the same corrupted subs=[10, 0] at the
same millisecond). Root-caused and fixed on the cx side by converting to
&sync.Mutex + sync.new_mutex() (cx-private PR for vlang#303); this issue is the
V-layer footgun that let it fail silently.

Suggested hardening (pick one or more)

1. Cheapest: assert/panic on nonzero return in lock()/unlock() (debug
   builds at minimum): an EINVAL here is always a programming error and
   currently costs days of intermittent-flake debugging.
2. -d sync_checked (or default in debug): use PTHREAD_MUTEX_ERRORCHECK
   attrs + checked returns.
3. Checker warning for VALUE-typed sync.Mutex/sync.RwMutex in __global
   blocks (zeroed in _vinit, unusable on Darwin) — suggest &sync.Mutex +
   sync.new_mutex() in a module init().

Upstream

This is not CX-specific — the discard is in upstream vlib/sync/sync_default.c.v
and bites any V program on macOS. Candidate for the next upstream bundle
(precedent: vlang#27458, vlang#27658): options 1 and 2 are V-only and
CX-agnostic; option 3 (checker warning) likewise. Fix here first, then include
in the bundle rather than filing upstream ad hoc.

Repro sketch

__global ( m sync.Mutex  counter int )
// spawn N threads: m.lock(); counter++; m.unlock()
// macOS: lost updates (lock is a no-op). Linux: correct.