fix: resolve SparseDistribution crash on zero/NaN probability mass

[Jonathangadeaharder]

Resolves #82

Changes

1. Fallback in SparseDistribution Constructor: Made SparseDistribution.__post_init__ fall back to a valid one-hot distribution on the first available token (greedy choice) if the sum of the probabilities is not finite or <= 0, rather than raising a crash-inducing ValueError.
2. Robust residual_distribution: Added not np.isfinite(total) checks to the fallback paths in residual_distribution (sparse, dense, and non-sparse branches) to prevent NaN values from bypassing target fallbacks.
3. Unit Tests: Added test cases in tests/test_sampling.py verifying fallback recovery on zero/NaN probability mass, as well as a mock test for residual_distribution NaN robustness.

All tests pass successfully.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[xlinbsd]

Reproduced on v0.3.7 — manual patch from PR #83 works**

Environment: M5 Pro 64 GB, macOS 26.5, Youssofal/Qwen3.6-27B-MTPLX-Optimized-Quality, --profile sustained, --reasoning off.

After a few minutes of agentic sessions the server crashes with ValueError: SparseDistribution probabilities must have positive mass, and the client gets Connection reset by peer (os error 54).

Applied the patch from PR #83 manually to site-packages/mtplx/sampling.py — two changes:

1. SparseDistribution.__post_init__ — fallback instead of raise:

# before
if not np.isfinite(total) or total <= 0:
    raise ValueError("SparseDistribution probabilities must have positive mass")

# after (mirrors fast_sampling.py which already has this fix)
if not np.isfinite(total) or total <= 0:
    token_ids = token_ids[:1] if token_ids.size > 0 else np.array([0], dtype=np.int64)
    probs = np.array([1.0], dtype=np.float64)
    total = 1.0

2. residual_distribution — add NaN guard to all 4 if total <= 0: checks:

if total <= 0 or not np.isfinite(total):

Sessions are now stable. Note that fast_sampling.py already has the equivalent fallback at lines 98-100, 151-153 and 208-211 — sampling.py was just missing it.

Please merge PR #83.