The most common support questions on the issue tracker boil down to
"did I install this right?" -- CUDA visible? flash-attn compiled?
HuggingFace auth set up? PAI dataset reachable? -- and today the
answer only surfaces 30 seconds into a long inference run when something
breaks. This adds a fast smoke test that answers all of those without
loading the model.

src/alpamayo_r1/healthcheck.py exposes 8 independent checks and a
`python -m alpamayo_r1.healthcheck` entry point. The exit code is the
number of failures (0 on green, non-zero count for CI). Skipped checks
(no GPU on a laptop, intentional flash-attn skip per README) do NOT
count as failures.

Checks (ordered):
1. python                -- 3.12.x as pyproject.toml requires
2. torch                 -- imports + version, warn if not 2.8.x
3. cuda                  -- device count, name, total memory, CUDA ver
4. transformers          -- 4.57.1 (pinned); mismatches break inference
5. flash_attn            -- importable; SKIP (not FAIL) when missing
                            since SDPA is a documented fallback
6. physical_ai_av        -- the PAI dataset loader dependency
7. hf_auth               -- whoami() succeeds (gated model + dataset)
8. alpamayo_r1           -- the package itself imports

Each check returns a CheckResult(name, status, detail) and is wrapped
in run() so a buggy/unexpected exception in one check is reported as a
fail and never aborts the rest.

Output (this laptop, no GPU/HF/torch fully installed):

    [FAIL] python        got 3.10.0, but pyproject.toml requires ==3.12.*
    [ OK ] torch         torch==2.11.0 (note: pyproject.toml pins torch==2.8.0)
    [SKIP] cuda          torch.cuda.is_available() is False -- no GPU
    [FAIL] transformers  import failed: ModuleNotFoundError(...)
    [SKIP] flash_attn    not importable. Inference still works via SDPA.
    [FAIL] physical_ai_av import failed: ... Run `uv sync --active`.
    [FAIL] hf_auth       Run `hf auth login` and accept the gated terms.
    [ OK ] alpamayo_r1   package importable

Summary: 2 passed, 4 failed, 2 skipped (8 total).

src/alpamayo_r1/test_healthcheck.py adds 7 pytest cases for the
runner: CheckResult immutability, _format_line glyphs/content, ALL_CHECKS
shape, run() exception isolation + ordering, main() exit code semantics
(failures count, skips do not, --quiet doesn't change exit code).

All 8 functional checks verified locally without GPU/HF dependencies --
output above is the actual run on this Mac.

Signed-off-by: lonexreb <[email protected]>