torch.linalg.eigvals() requires a 2-D matrix input; calling it on the
1-D diagonal vector returned by get_A() raises a runtime error.
get_A() returns the diagonal of A_discrete, so for a diagonal matrix
spectral radius = max(|diagonal|) — use .abs().max() directly on the
1-D tensor, which is both correct and more efficient.

flash-attn is unavailable on Intel Mac, ROCm, and CPU-only environments
(no pre-built wheels). The ImportError fallback is already implemented in
main.py but was not documented. Add a clear note so users on non-CUDA
platforms know they don't need flash-attn and the model will work without it.

Runs on every push and pull request using ubuntu-latest + Python 3.11.
Installs torch CPU wheel (avoids CUDA build overhead), installs the
package in editable mode, then runs pytest on tests/ with verbose output.
Catches import errors and shape regressions on every PR automatically.

Adds mythos_tiny() to variants.py: dim=128, n_heads=4, vocab_size=1000,
n_experts=2, recurrent_iters=2, max_seq_len=64. Targets ~1-3M params so
tests complete in seconds on any machine without a GPU.

Also exports TINY_CONFIG singleton for convenient import in test files.
Updated __init__.py to re-export both mythos_tiny and TINY_CONFIG.

…moke

- tests/test_attention.py: GQA and MLA shape checks, KV-cache population,
  grouped KV-head path, and cache-extended sequence decode
- tests/test_moe.py: router top-k gating, shared expert always-fires,
  output shape and finiteness, single Expert shape
- tests/test_recurrent_block.py: LTIInjection spectral radius < 1 (including
  extreme parameters), LoRA clamp safety, RecurrentBlock depth changes output
- tests/test_smoke.py: full forward + backward pass using TINY_CONFIG,
  spectral radius, generate() shape, GQA variant smoke

~30-50 lines per file; non-trivial regression catchers, no 100% coverage goal.

Covers install, flash-attn fallback note, 8-line smoke test using TINY_CONFIG,
spectral radius verification, backward pass check, switching to production
configs, generate() example, pytest invocation, and a troubleshooting table
for the three known install issues (torch pin, eigvals, flash-attn).

…ams test

With log_A*10 and log_dt*10 the exp() clamp at -20/20 can produce values
that round to exactly 1.0 in float32 (exp(-exp(-20)) ≈ exp(0) = 1).
Relax the strict < 1.0 check to <= 1.0 + 1e-6 which captures the meaningful
stability constraint (no divergence) without being broken by float32 precision.