Fix negative secondary mass crashing log_transform during rollout

decode_secondaries() applied inv_log_transform() to the model's raw
predicted log_mass directly. Since that value isn't itself the output
of log_transform, exp(log_mass) can undershoot _EPS, making
inv_log_transform(log_mass) = exp(log_mass) - _EPS go slightly
negative. Once that secondary spawns a track and its mass is fed back
in as conditioning for a further rollout step, log_transform(mass)
computes log(mass + eps) with mass <= -eps, producing a non-finite
value and raising.

Clip log_mass to log(_EPS) before inverting so the resulting mass is
guaranteed >= 0 (matching the invariant the surrounding comment
already assumed, but didn't enforce).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

giant/data/transforms.py

@@ -629,9 +629,16 @@ def decode_secondaries(
                 pre_dir[valid], dir_local[valid, i]
             )
 
-    # mass is non-negative by construction (inv_log_transform of a real
-    # number is always > 0); clip to 0 for padded/invalid slots rather than
-    # leaving a spurious small positive floor from the log inverse.
+    # log_mass is a raw model prediction, not itself the output of
+    # log_transform, so exp(log_mass) can undershoot _EPS and make
+    # inv_log_transform(log_mass) = exp(log_mass) - _EPS go slightly
+    # negative. That negative mass then blows up the next log_transform
+    # call on this track's mass once it's fed back in as conditioning for a
+    # further rollout step (giant/rollout.py -> build_cond_features ->
+    # _physical_cond_columns). Clip log_mass so its inverse is guaranteed
+    # >= 0 before that can happen; clip to 0 separately for padded/invalid
+    # slots rather than leaving a spurious small positive floor.
+    log_mass = np.clip(log_mass, np.log(_EPS), None)
     sec_mass = np.where(sec_valid, inv_log_transform(log_mass), 0.0).astype(np.float32)
     sec_charge = np.where(sec_valid, charge, 0.0).astype(np.float32)
 

tests/test_phase2.py

@@ -540,3 +540,29 @@ def test_decode_secondaries_mass_charge_round_trip_with_normalizer():
     )
     assert sec_mass[0, 0] == pytest.approx(938.27208943, abs=1e-2)
     assert sec_charge[0, 0] == pytest.approx(1.0, abs=1e-4)
+
+
+def test_decode_secondaries_extreme_negative_log_mass_stays_nonnegative():
+    from giant.data.transforms import decode_secondaries, log_transform
+
+    N = 1
+    e_sec = np.array([5.0], dtype=np.float32)
+    n_sec = np.array([1])
+    pre_dir = np.array([[0.0, 0.0, 1.0]], dtype=np.float32)
+
+    sec_cont = np.zeros((N, K_MAX, 6), dtype=np.float32)
+    sec_cont[0, 0, 0] = 10.0  # stick logit -> ~all of e_sec
+    sec_cont[0, 0, 1:4] = [0, 0, 1]
+    sec_cont[0, 0, 4] = -50.0  # raw model prediction: extremely negative log_mass
+    sec_cont[0, 0, 5] = 1.0
+
+    _, _, sec_mass, _, _ = decode_secondaries(sec_cont, n_sec, e_sec, pre_dir)
+
+    # A raw model prediction isn't itself the output of log_transform, so
+    # naively applying inv_log_transform can undershoot zero (see
+    # decode_secondaries) — which then crashes the next log_transform call
+    # once this mass is fed back in as conditioning during rollout. The
+    # float32 residual from clipping can land a hair below zero, but must
+    # stay well above -eps so log_transform(mass) stays finite.
+    assert sec_mass[0, 0] > -1e-8
+    log_transform(sec_mass[0, 0])