Build cleanly on Windows/Linux/macOS; replace vendored ogg/vorbis/flac with single-header decoders

[AdamCoulterOz]

Summary

This makes the wip branch build cleanly on Windows (MSVC), Linux (GCC) and
macOS (Clang) from a single cmake -DBSTONE_SDL_BUNDLED=ON configure, adds CI that
proves it, and gets the existing test suite running again. Along the way it replaces the
vendored Xiph audio libraries with two zero-configuration single-header decoders.

All commits are small and self-contained. CI is green on all six jobs (three platforms ×
Debug/Release), and the unit tests pass in every one.

What's here

Toolchain / build-portability fixes

• Add missing <algorithm> includes — modern libc++/libstdc++ no longer pull these in
  transitively, so std::max / std::copy_n and friends failed to compile.
• Guard bstone_win32_utility.{cpp,h} with #ifdef _WIN32 so the Win32-only UTF helpers
  are not compiled on other platforms, matching its three already-guarded siblings. The
  file self-guards rather than the build carving out a Windows-only source list.
• Use if(MSVC) for the Windows manifest source instead of a $<CXX_COMPILER_ID>
  generator expression, which is not valid in a source-file list on the VS generator.

Fix: sys::File reported closed-file operations by asserting

sys::File is a return-code primitive — FileStream calls it and turns a negative/false
result into an exception, and its own tests require that operating on a closed file fails
that way. Every method instead asserted is_open() and then passed the null handle to SDL
regardless. So a debug build aborted rather than failing, and a release build depended on
SDL's own null-handle parameter checks to produce the right answer.

Each operation now returns its documented failure value when the file is not open.
Likewise an unknown FileMode or FileOrigin is a rejected request the return value
already expresses, not a programming error, so those two asserts are gone as well; genuine
preconditions (non-null buffer, non-negative size) keep theirs.

This was aborting the test suite partway through — 238 of 430 tests ran before the process
died on SIGABRT. The whole suite passes now.

CI

• Add .github/workflows/build.yml: a full cross-product of the three platforms and
  Debug/Release, building every target (game, tools, bundled libraries, tests) and
  running the unit tests in each configuration.
• Both configurations are run on purpose: a Release build defines NDEBUG, which compiles
  BSTONE_ASSERT away, so only the Debug run exercises the assertion-guarded failure paths
  — which is exactly how the sys::File problem above went unnoticed.
• Install libXtst dev headers on Linux (the bundled SDL3 X11 backend treats XTEST as a
  hard dependency).

Audio: replace vendored ogg/vorbis/flac with stb_vorbis + dr_flac

The vendored Xiph libraries (libogg, libvorbis, libFLAC — 35 C source files) carry
autotools/CMake-era configuration that has to be detected correctly per platform
(HAVE_ALLOCA_H, SIZE_MAX / <stdint.h>, ogg's ogg_int*_t typedefs, lround).
Getting those right across GCC/MSVC/Clang was most of the porting friction.

They are replaced with two self-contained public-domain single-header decoders that need no
build configuration and compile identically everywhere:

• stb_vorbis.c (nothings/stb @ 1ee679ca2ef7) → bstone::lib::stb_vorbis
• dr_flac.h (mackron/dr_libs @ e1f125c1e03b) → bstone::lib::dr_flac

Each is vendored at a pinned commit and built as an isolated C static library, so
third-party code never sees bstone's stricter C++ flags. The Vorbis and FLAC
AudioDecoder implementations are rewritten onto the new APIs while preserving the
existing interface, factory functions and callers exactly — same whole-stream-to-memory
decode, interleaved float output, nearest-neighbour resampler, 1–2 channel /
8000–96000 Hz validation, and rewind. dr_flac's normalised f32 output additionally
removes the per-bit-depth sample converters and the libFLAC callback machinery.

One small correctness improvement: rewind() now also resets the resampler's phase
accumulator, so a looped, resampled track restarts cleanly.

Repo hygiene

• Ignore build-*/ output directories and data/, so out-of-source build trees and the
  copyrighted game assets used for local testing cannot be committed by accident.

Testing

• CI green on all six jobs: Linux/Windows/macOS × Debug/Release.
• Unit tests run in every job — 380 tests on Linux/macOS, 402 on Windows (the extra 22 are
  the _WIN32-gated registry tests), 0 failures.
• Built and ran the game locally on macOS.
• The decoder rewrites keep the AudioDecoder contract identical; validation and
  resampling logic are carried over unchanged.

---

I'm happy to split this if you'd prefer to review the pieces independently — the build/CI
stabilisation and the sys::File fix stand on their own, and the decoder swap is the
opinionated part.

---

🤖 Generated with Claude Code

[bibendovsky]

There is a common problem with modified audio decoders.
Reading the whole audio file into memory from a file is unacceptable.
The audio caching (whole or partial) already done in appropriate audio mixer.

[AdamCoulterOz]

Reworked both decoders to stream (be5efea..62f1205): dr_flac now pulls through read/seek/tell callbacks, with seeking emulated over the VFS stream's rewind-and-read contract, and stb_vorbis runs in push mode over a bounded window (32 KiB, growing only if an Ogg page outsizes it); rewind reopens the push stream since it cannot seek. Nothing reads the file up front any more — WAV already streamed. Added decoder tests with encoder-generated fixtures, including a reads-arrive-seven-bytes-at-a-time case to prove it.

🤖 Addressed by Claude Code

[bibendovsky]

One last thing.
The curly braces for controlled one-line statement was ommited.

Before:

for (int i = 0; i < 10; ++i)
{
    if (cond)
        do_this();
}
if (cond2)
    do_that();
else
    return;

After:

for (int i = 0; i < 10; ++i)
{
    if (cond)
    {
        do_this();
    }
}
if (cond2)
{
    do_that();
}
else
{
    return;
}

[AdamCoulterOz]

Braced it (5dc71d8) — a sweep of the PR's added lines found exactly one offender, the break in the read_exactly that moved into bstone_vfs.h; the other brace-less bodies the sweep surfaced are all inside the vendored dr_flac.h, which I left in its upstream style.

🤖 Addressed by Claude Code

[bibendovsky]

Thank you!