Validate probabilities in noise channel constructors (#132)

## Summary

The Pauli channel probability helpers (`error_probs`,
`pauli_channel_1_probs`, `pauli_channel_2_probs`,
`heralded_pauli_channel_1_probs`) accepted invalid inputs and produced
malformed distributions: `probs[0]` could come out negative when
individual entries were outside `[0, 1]` or their sum exceeded 1.
`ChannelSampler._precompute_sparse` then silently dropped those channels
because `p_fire = 1 - probs[0] > 1` doesn't trigger its `p_fire <=
1e-15` early-return — sampling proceeded as if the channel did nothing.

In the normal stim-driven pipeline this is unreachable (stim rejects bad
probabilities at parse time), but it bites anyone constructing a
`Channel` directly.

- Added `_validate_probabilities` to check each input is in `[0, 1]` and
the sum is ≤ 1.
- Wired it into all four Pauli helpers.
- `ChannelSampler._precompute_sparse` now raises `ValueError` when
`p_fire < 0` instead of silently skipping.

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: rafaelha

src/tsim/noise/channels.py

@@ -27,6 +27,17 @@ class Channel:
     probs: np.ndarray
     unique_col_ids: tuple[int, ...]
 
+    def __post_init__(self) -> None:
+        """Validate channel probabilities."""
+        tol = 1e-12
+        if np.any(self.probs < -tol) or np.any(self.probs > 1.0 + tol):
+            raise ValueError(f"Probabilities must lie in [0, 1], but got: {self.probs}")
+        if not np.isclose(np.sum(self.probs), 1.0):
+            raise ValueError(
+                f"Probabilities must sum to 1, but got: {self.probs} "
+                f"(sum {np.sum(self.probs)})"
+            )
+
     @property
     def num_bits(self) -> int:
         """Number of bits in the channel (k where probs has shape 2^k)."""

test/unit/noise/test_channels.py

@@ -162,6 +162,41 @@ def test_chain_with_certain_first_error(self):
         assert_allclose(probs[4], 0.0)  # index 4 (0b100): third error
 
 
+class TestChannelValidation:
+    """Channel.__post_init__ rejects malformed probability distributions so
+    they cannot reach the sampler and get silently dropped."""
+
+    def test_rejects_negative_entry(self):
+        with pytest.raises(ValueError):
+            Channel(probs=np.array([1.2, -0.2]), unique_col_ids=(0,))
+
+    def test_rejects_entry_above_one(self):
+        with pytest.raises(ValueError):
+            Channel(probs=np.array([-0.5, 1.5]), unique_col_ids=(0,))
+
+    def test_rejects_sum_below_one(self):
+        with pytest.raises(ValueError):
+            Channel(probs=np.array([0.5, 0.4]), unique_col_ids=(0,))
+
+    def test_rejects_sum_above_one(self):
+        with pytest.raises(ValueError):
+            Channel(probs=np.array([0.7, 0.7]), unique_col_ids=(0,))
+
+    def test_rejects_helper_with_arguments_summing_above_one(self):
+        # pauli_channel_1_probs(0.6, 0.6, 0.6) returns probs[0] = -0.8.
+        # Wrapping in a Channel surfaces the negative entry as a clear error.
+        with pytest.raises(ValueError):
+            Channel(probs=pauli_channel_1_probs(0.6, 0.6, 0.6), unique_col_ids=(0, 1))
+
+    def test_channel_sampler_rejects_invalid_channel(self):
+        # ChannelSampler wraps probs in a Channel, so invalid distributions
+        # get caught at construction time rather than silently dropped.
+        bad_probs = np.array([1.2, -0.1, -0.05, -0.05])
+        transform = np.array([[1, 0], [0, 1]], dtype=np.uint8)
+        with pytest.raises(ValueError):
+            ChannelSampler([bad_probs], transform, seed=42)
+
+
 def _sample_channels(channels, matrix, n_samples, seed=42):
     """Sample from channels using the ChannelSampler infrastructure."""
     sampler = object.__new__(ChannelSampler)