Build and test on macOS as well as Linux

[Aduneer]

Everything FlashTerm does outside the standard library is POSIX rather than Linux — termios, ioctl, dirent, wcwidth — and date.cpp and event.cpp have carried their gmtime_r/localtime_r branches since the log landed. So this was expected to pass.

It did not, three times over, and that is the point of the PR.

What changed

os: [ubuntu-latest, macos-latest] across the existing compiler and mode axes, with one pair excluded: g++ on a macOS runner is a shim for Apple clang rather than GCC, so that combination would run the clang entry a second time under another name. Six jobs rather than eight.

	optimised	sanitised
ubuntu · g++	✓	✓
ubuntu · clang++	✓	✓
macos · clang++	✓	✓

macos-latest is Apple Silicon, so this also builds for arm64, where char is unsigned by default and every x86 build so far has had it signed.

What it caught

The code compiled clean. Nothing below is a macOS-only defect — each is something that was already true and that only a second platform could show.

1. FLASHTERM_SEED did not fix the order it promised to fix

Six of forty-seven golden cases failed, and every one was a multi-card case.

std::shuffle is specified to produce a uniformly random permutation and nothing about which one, so libstdc++ and libc++ deal the same seeded deck in different orders. A scripted session answers cards in the order it expects to meet them, so transcripts recorded against libstdc++ were simply wrong on libc++ — surfacing as every answer being graded incorrect.

FLASHTERM_SEED is documented as fixing the review order. A promise that holds only on the standard library you happened to build against is not that promise. It is now an explicit Fisher–Yates:

• std::mt19937 stays. The standard pins its algorithm exactly, down to a required value for the 10000th draw — the generator was never the problem.
• bounded() draws an index by rejection sampling, since std::uniform_int_distribution is unspecified in the same way.

Nothing was wrong with the randomness and no real session is affected. One transcript changes on Linux, review-multi-card, whose input is remapped rather than re-recorded so the case still exercises what it was written for — the same correct, hint-then-partial, incorrect, correct, correct. Its summary is still Correct: 3 · Partial: 1 · Incorrect: 1. Two lines of the input file change.

2 and 3. Two golden cases could observe the machine

Same class as the bug that made run.sh pin the environment for every case rather than per case.

wc pads its count with leading spaces on a BSD userland, so a binary fixture was described as <binary, 597 bytes> on macOS and <binary, 597 bytes> on Linux. Stripped rather than reformatted, since the value is compared as an integer as well as printed.

review-audio-unplayable pointed FLASHTERM_TTS and FLASHTERM_PLAYER at /bin/false to reach the "No audio available for this card" path. There is no /bin/false on macOS — false lives in /usr/bin — so FlashTerm rejected an override it could not execute, fell back to the built-in candidate list, found macOS's own say, and it succeeded. The case had stopped testing a failed playback at all.

fake-audio.sh now fails on demand via FAKE_AUDIO_FAIL, the same shape as fakebin/piper and FAKE_PIPER_FAIL, and the case asks for that instead of naming a command it hopes exists. Its transcript is byte for byte what it was, so what it checks is unchanged — it merely checks it everywhere now.

Also

• README's Quick Start now says which platforms are built and tested, which it never did.
• README's Contributing paragraph said "all four with -Werror", which is now six.
• CHANGELOG.md gains an Unreleased section. Its historical 0.2.0 and 0.3.0 entries still say four jobs and are deliberately left alone: they record what was true at those releases.

Checks

Green on all six jobs. Locally before each push: make check on gcc and clang, optimised and sanitised (-fsanitize=address,undefined -fno-sanitize-recover=undefined), plus the env -i clean room. 1184 unit checks, 47 golden cases.