Fix crash in fault injection and flaky thread affinity test

[saxena-anurag]

Problem 1: Crash in fault injection due to unserialized dbghelp API calls

Fault injection tests crash with FAIL_FAST_GUARD_ICALL_CHECK_FAILURE (0xc0000409) inside dbghelp!diaGetLineFromAddr when multiple threads concurrently trigger fault injection. The crash manifests as a use-after-free of internal DIA SDK COM objects (RAX = 0xfeeefeeefeeefeee, the MSVC freed-heap fill pattern).

The root cause is that _cxplat_decode_symbol calls SymFromAddr and SymGetLineFromAddr64 without synchronization. These dbghelp APIs are explicitly documented as not thread-safe, and concurrent calls corrupt their internal state.

Additionally, IMAGEHLP_LINE64.SizeOfStruct was never initialized, violating the API contract for SymGetLineFromAddr64.

Fix

• Add a static std::mutex inside _cxplat_decode_symbol to serialize all dbghelp API calls.
• Zero-initialize IMAGEHLP_LINE64 and set SizeOfStruct as required by the API.

Repro

The crash occurs in concurrent fault injection tests such as _test_custom_maps_concurrent_insert_delete_and_query where multiple threads simultaneously allocate memory, triggering the fault injection stack-trace capture path.

Call stack

dbghelp!diaGetLineFromAddr+0x160
dbghelp!SymGetLineFromAddr64+0x69
unit_tests!_cxplat_decode_symbol+0x17f
unit_tests!_cxplat_fault_injection::log_stack_trace+0x281
unit_tests!_cxplat_fault_injection::is_new_stack+0x2c0
unit_tests!_cxplat_fault_injection::inject_fault+0x20
unit_tests!cxplat_fault_injection_inject_fault+0x3a
unit_tests!cxplat_allocate+0x5c

Problem 2: Flaky threads test failure

The threads test intermittently fails with:

REQUIRE( old_affinity == 0 )
with expansion:
  1 == 0

This is a pre-existing flaky failure (also seen on unrelated PRs like #310) that depends on Catch2's random test ordering — it reproduces when dpcs runs before threads.

Root Cause

When KeSetSystemAffinityThreadEx is called at DISPATCH_LEVEL (e.g., via KeRevertToUserAffinityThreadEx), it only updates _usersim_group_affinity_cache but does not update _usersim_group_before_raise_irql. When KeLowerIrql subsequently restores thread affinity, it uses the stale _usersim_group_before_raise_irql value (from the original KfRaiseIrql call), overwriting the intended affinity change.

In the dpcs test:

1. KeSetSystemAffinityThreadEx(1) — pins to CPU 0, cache becomes {Mask:1}
2. KeRaiseIrqlToDpcLevel() — saves _usersim_group_before_raise_irql = {Mask:1}
3. KeRevertToUserAffinityThreadEx(0) at DISPATCH — updates cache to {Mask:0} but _usersim_group_before_raise_irql remains {Mask:1}
4. KeLowerIrql() — restores from _usersim_group_before_raise_irql, setting cache back to {Mask:1}

Fix

Update _usersim_group_before_raise_irql in KeSetSystemAffinityThreadEx when called at DISPATCH_LEVEL, so KeLowerIrql restores the correct affinity.

Validation

• All 75 usersim tests pass without fault injection
• All [ke] tests pass consistently with fault injection (CXPLAT_FAULT_INJECTION_SIMULATION=4)
• The specific dpcs,threads ordering that previously failed now passes reliably