Fix get_parent same-time-tag bug in LineageClustering

[DonNabla]

Fix get_parent same-time-tag bug in LineageClustering

Summary

Tie-break by spatial proximity to the daughter's creation position when multiple parent interactions in lineage_cluster.get_parent share the same time tag.

For fast γ events that Compton-scatter at one site and photoabsorb at another within G4's sub-ns time-tag precision (~0.1 ns of travel time, bit-identical float64 t), the original time-cut + array-last logic returned the last same-time parent step in array order — which can be the wrong vertex. The Compton e⁻ daughter at site A then inherited the lineage of the γ's photoabsorbtion step at site B, collapsing the two physical scatter sites into a single cluster.

What's affected

Empirically the bug only flips event-level SS/MS classification for single-γ-decay isotopes (K40, Co60). Cascade chains (Th, U, Ra) and β-only sources (Pb212 with nucleusLimits excluding daughter chain) are unaffected, because either (a) cascade γ's have distinct decay-time tags so get_parent's time-cut already disambiguates, or (b) there's no γ Compton+phot pair to trigger the same-time-tag tie in the first place.

For K40/Co60, ~65 % of physically multi-scatter events end up reclassified as single-scatter, biasing the MS templates by ~50 % vs WFsim.

Validation

5 isotopes × 20 jobs × 100k primaries each (InnerCryostatFlange for K40/Co60/Th232/Ra226; WholeLXe for Pb212; SR0 Fermi-Dirac SS/MS classifier):

isotope	topology	vanilla MS-ratio vs WFsim 2026	patched MS-ratio vs WFsim 2026
K40	1.46 MeV single γ	0.385	1.006 ✓
Co60	1.17+1.33 MeV 2-γ cascade	0.351	0.998 ✓
Th232	chain cascade γ's	0.997	0.996
Ra226	α + 186 keV γ + chain	1.002	1.000
Pb212	intrinsic β (no γ)	1.000	1.000 — bit-identical h5

Single-event spot check (K40 g4id=39232):

	site A (-35.6, -21.8, -0.2)	site B (-33.2, -18.4, -0.6)	alt_cs2_wo_timecorr	SS/MS
WFsim epix	423.27 keV	1036.87 keV	154,972 PE	MS
Vanilla fuse	0.06 keV	1460.76 keV	128 PE	SS ✗
Patched fuse	423.95 keV	1036.87 keV	189,862 PE	MS ✓

The diff

# Among parents with t <= particle.t, pick the latest time and then
# tie-break by spatial proximity if multiple steps share that latest time.
candidates = parent_interactions[parent_interactions_time_cut]
candidate_lineages = parent_lineages[parent_interactions_time_cut]
t_max = candidates["t"].max()
same_time_mask = candidates["t"] == t_max
if same_time_mask.sum() == 1:
    idx = int(np.where(same_time_mask)[0][0])
    return candidates[idx], candidate_lineages[idx]

same_time = candidates[same_time_mask]
same_time_lineages = candidate_lineages[same_time_mask]
distances = np.sqrt(
    (same_time["x"] - particle["x"]) ** 2
    + (same_time["y"] - particle["y"]) ** 2
    + (same_time["z"] - particle["z"]) ** 2
)
nearest = int(np.argmin(distances))
return same_time[nearest], same_time_lineages[nearest]

The existing nearest-in-time fallback (when no parent step satisfies t <= particle.t) is preserved unchanged.

Alternatives considered

I also prototyped and validated two position-free alternatives:

• Process-match: daughter.creaproc == parent.edproc (G4's causal mapping). Closes ~85 % of the gap.
• Precompute hybrid: {trackid: parent_step_index} map computed once per event using process-match → spatial fallback, making get_parent O(1). Also ~85 %.

Both are causally explicit but converge on the same coverage. The spatial tie-break in this PR catches the additional ~15 % of edge cases (daughters whose creaproc isn't directly in the parent's edproc set, e.g. secondary brems and cascaded photo-absorptions). A true hybrid (process-match prefilter on the same-time-max candidate set with spatial tie-break inside) is the principled long-term refinement — worth a follow-up PR but not blocking.

Test plan

• Existing fuse unit tests pass
• Spot-check K40 g4id=39232 produces the 423/1037 keV split documented above
• Cross-isotope MS-rate ratios vs WFsim land within statistics of the table above on a re-run

[DonNabla]

Happy to get the opinion of the fuse expert for this change. Mostly relevant for material ER simulation.

[HenningSE]

Looks good to me, approved