`group` property is re-enumerated from 0 when slicing a recording with a probe (`split_by`, `remove_channels`) since 0.105

Start by reading ChannelSliceRecording.__init__ and the set_probegroup path that calls _set_group_property_based_on_probegroup; also inspect the corresponding re-attachment path in ChannelsAggregationRecording. Use the minimal reproducer to check how slicing and channel removal affect group, then add regression coverage for the reported cases. Done means existing parent group values survive slicing and aggregation rather than being re-enumerated.

Since 0.105.0, slicing a recording that has a probe attached (split_by, select_channels, remove_channels, channel_slice) no longer keeps the group property of the parent: ChannelSliceRecording.__init__ copies the properties and then calls self.set_probegroup(sliced_probegroup), which runs _set_group_property_based_on_probegroup(..., group_mode="auto") and re-enumerates group from 0 over whatever probe/shank/side combinations are still present. The parent's values are overwritten.

Consequences:

• split_by("group") returns sub-recordings that all have group == 0.
• Removing a whole shank renumbers the remaining shanks.
• Any user-defined grouping on a single-shank probe (set_property("group", ...)) is wiped by any channel removal (everything becomes group 0).

0.103.x and 0.104.1 keep the parent's values. Introduced by #4465 (ChannelSliceRecording re-attaching the sliced probegroup with the new auto mode); the same code is on main.

Minimal reproducer

from spikeinterface.core import generate_recording

rec = generate_recording(num_channels=8, durations=[1.0])  # has a probe attached
rec.set_property("group", [0, 0, 1, 1, 2, 2, 3, 3])
for g, sub in rec.split_by("group").items():
    print(g, sub.get_property("group").tolist())
print(rec.remove_channels(rec.channel_ids[[1, 3]]).get_property("group").tolist())

Realistic probes (384 channels)

import numpy as np
from probeinterface import generate_multi_shank, generate_multi_columns_probe
from spikeinterface.core import generate_recording, aggregate_channels

def summarize(label, rec):
    vals, counts = np.unique(rec.get_property("group"), return_counts=True)
    print(f"  {label:<34} n_ch={rec.get_num_channels():>3}  groups={dict(zip(vals.tolist(), counts.tolist()))}")

def make(probe, n):
    probe.set_device_channel_indices(np.arange(n))
    rec = generate_recording(num_channels=n, durations=[1.0], set_probe=False)
    return rec.rename_channels([f"AP{i}" for i in range(n)]), probe

rng = np.random.default_rng(0)

# NP2-like 4-shank probe, 96 channels per shank, grouped by shank
rec, probe = make(generate_multi_shank(num_shank=4, num_columns=2, num_contact_per_column=48, shank_pitch=[250, 0]), 384)
rec = rec.set_probe(probe, group_mode="by_shank") or rec   # 0.103 returns a new object, 0.105 is in place
summarize("parent", rec)
for g, sub in rec.split_by("group").items():
    summarize(f"split_by[{g}] (first id {sub.channel_ids[0]})", sub)
summarize("remove_channels(10 random)", rec.remove_channels(rec.channel_ids[rng.choice(384, 10, replace=False)]))
summarize("remove_channels(all of shank 2)", rec.remove_channels(rec.channel_ids[192:288]))

# NP1-like single-shank probe, 384 channels, user-defined group = 4 depth bands of 96
rec, probe = make(generate_multi_columns_probe(num_columns=2, num_contact_per_column=192, xpitch=32, ypitch=20), 384)
rec = rec.set_probe(probe) or rec
rec.set_property("group", np.repeat([0, 1, 2, 3], 96))
for g, sub in rec.split_by("group").items():
    summarize(f"split_by[{g}] (first id {sub.channel_ids[0]})", sub)
summarize("remove_channels(10 random)", rec.remove_channels(rec.channel_ids[rng.choice(384, 10, replace=False)]))

4-shank probe, set_probe(group_mode="by_shank"):

Single-shank probe, user-defined groups:

(macOS arm64, Python 3.12, numpy 2.5.3, probeinterface 0.4.0; the 0.105 column is identical on linux x86_64 in the aind-ephys-pipeline-base:1.4.0 image.)

Real-world impact

aind-ephys-pipeline 1.4.0 pins SI 0.105 in its images. aind-ephys-job-dispatch does recording.split_by("group") and dumps each slice to JSON; every slice now reports group 0. aind-units-nwb builds the electrode-group label from that property, so all shanks of a NP2 4-shank probe become <probe>_group0, while the electrodes table written from the full recording still has _group0.._group3. The units export then fails on the first channel of shank 1 (KeyError: 'AP48'). Workaround PR: AllenNeuralDynamics/aind-units-nwb#32 (derive the label from the recording name instead of the property).

Suggested fix

When ChannelSliceRecording (and ChannelsAggregationRecording) re-attach the sliced/aggregated probegroup, keep an existing group property instead of overwriting it, and only derive group from the probegroup when the parent had none. Alternatively copy_metadata could run after set_probegroup, so user values win.