On musl distros (e.g. Alpine) libc and the dynamic loader are combined
into a single module named like ld-musl-x86_64.so.1, where threads start.
Treat it as a valid top frame so musl stacks aren't marked BROKEN, mirroring
the existing glibc libc handling. Add an in-memory nettrace unit test.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

## Why

Traces captured with one-collect on musl-based Linux distros (e.g.
Alpine) show a large fraction of stacks marked BROKEN when opened in
PerfView or TraceEvent. On a real Alpine trace, about 18% of samples
were broken. The root cause is that thread base frames on musl live in
the combined libc + dynamic loader module, named like
`ld-musl-x86_64.so.1`, which `ReasonableTopFrame` did not recognize as a
legitimate thread-start module. We already special-case glibc's `libc`
for the same reason; musl was simply missing.

## What

- `TraceEventStacks.cs`: In `ReasonableTopFrame`, treat any module whose
name starts with `ld-musl-` (case-insensitive, so it covers `x86_64`,
`aarch64`, etc.) as a valid top frame, mirroring the existing glibc
`libc` handling. When matched, a stack rooted there is no longer wrapped
in a BROKEN frame.

- Added `Universal/MuslBrokenStackTests.cs`: a unit test that
synthesizes a tiny in-memory V6 nettrace (no large trace binary added to
the repo) containing a musl loader module, an ordinary library, and two
cpu samples: one rooted in the musl module and one rooted in the
ordinary library. It asserts the musl-rooted stack is not BROKEN while
the ordinary-library-rooted stack still is, parameterized over both
`ld-musl-x86_64.so.1` and `ld-musl-aarch64.so.1`.

## Validation

- A throwaway TraceEvent-based console app run against the real Alpine
trace confirmed broken stacks dropped from 18.4% to 3.7% (eliminating
~6,100 musl-rooted frames). The scratch app was not committed.
- Confirmed the new test fails without the fix (musl stack marked
BROKEN) and passes with it.
- Full TraceEvent suite is green: 2272 passing on net8.0 and 2289 on
net462, no regressions.

## Notes for reviewers

The negative-case assertion (ordinary library still BROKEN) is
intentional to guard against the match being too broad. The test writer
emits a non-zero `syncTimeQPC` so iterating relative timestamps does not
trip a Debug.Assert in `QPCTimeToRelMSec`; this is a test-harness detail
only.

When collecting with /InMemoryCircularBuffer and /CircularMB:N where N is
larger than ~256, PerfView collection succeeds but the subsequent
KernelTraceControl merge fails (e.g. CreateMergedTraceFile returns
0x80280012), and the requested memory grows quadratically.

The buffering-mode branch of GetProperties overwrote properties->BufferSize
with m_CircularBufferMB. BufferSize is a per-buffer size in KB, so a megabyte
value was being written into a kilobyte field. This made each ETW buffer far
larger than the OS maximum buffer size, which the relogger later chokes on
during merge, and it contradicted the MinimumBuffers computation on the line
above (which divides by the 64 KB quantum). It also corrupted the Attach
round-trip that reconstructs buffer settings when stopping the session.

Remove the erroneous assignment so BufferSize stays at the per-buffer
quantum (m_BufferQuantumKB) set earlier, matching the file-based and
real-time paths. MinimumBuffers alone now sizes the in-memory pool, so
(MinimumBuffers * BufferSize) == m_CircularBufferMB megabytes as intended.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

PerfView 3.2.4 can throw `ArgumentOutOfRangeException` when GC Stats or
Heap Analyzer replays `GC/CommittedUsage` events from ETLX. During
conversion, `FixupData()` classifies the raw dynamic event and the ETLX
persists the synthetic event ID, but the original dynamic payload
envelope remains unchanged. On replay, the typed template therefore
still needs to parse that envelope.

This change refreshes the payload layout in `EventPayload` only when
processing a `TraceLog`. Raw ETW and EventPipe dispatch continue using
the payload prepared by `FixupData()`, avoiding redundant parsing.
Fixed-offset `CommittedUsage` fields also return safe defaults for
truncated payloads.

Regression coverage includes valid ETLX replay without `FixupData()`,
reused replay templates with distinct payloads, and malformed payload
access through properties, `PayloadValues`, and `ToXml`.

Fixes #2438

* Enable W4 and warnings as errors for C++ projects

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 5b148be3-5dce-425e-b025-94eb7b8de1d5

* Remove stale sampling debug code

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 5b148be3-5dce-425e-b025-94eb7b8de1d5

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Support embedded portable PDBs for managed symbol resolution

Managed assemblies built with <DebugType>embedded</DebugType> carry their
portable PDB inside the PE image rather than in a standalone .pdb file.
Previously SymbolReader could not read these, so source/line lookup failed
for such modules.

Add the ability to open an embedded portable PDB:

- SymbolReader.OpenEmbeddedPortablePdb reads the EmbeddedPortablePdb
  debug-directory entry from a module and returns a PortableSymbolModule.
  Results are cached under a key suffixed so they cannot collide with
  standalone-PDB cache entries.
- SymbolReader.OpenSymbolFileForModuleFile is a module-oriented entry point
  that prefers a standalone PDB and falls back to an embedded portable PDB.
- PortableSymbolModule gains a constructor that takes a
  MetadataReaderProvider (which owns its own backing memory).
- TraceLog.OpenPdbForModuleFile falls back to the module's embedded portable
  PDB when the on-disk module matches the trace and no standalone PDB exists.

Tests: add an EmbeddedPdbTestApp fixture (built with embedded PDBs) and
cover the happy path, caching, not-embedded/missing-file cases, the
module-oriented entry point (embedded and standalone), and the end-to-end
TraceLog fallback path. All run cross-platform on net462 and net8.0.

* Build embedded-PDB test fixture unoptimized for stable source lines

In Release the C# compiler optimizes EmbeddedTarget.Add, shifting its first sequence point from the 'int sum = a + b;' line to the 'return' line, so the source-line assertions failed in the Release CI leg. Pin Optimize=false on the fixture so its emitted sequence points are stable regardless of the build configuration.

* Remove test-only OpenSymbolFileForModuleFile wrapper

Per PR review: this wrapper was called only from tests, and TraceLog.OpenPdbForModuleFile already provides the standalone-then-embedded fallback for the trace path. Removed it and its three dedicated tests, keeping OpenEmbeddedPortablePdb (the building block TraceLog uses).