Sample-adaptive regression collection + multi-output adaptive linear regressor

[cboulay]

Summary

Adds a new SampleAdaptRegressor collection that wires resampling, trigger-driven sequence-to-sequence sampling, optional time-lag windowing, and an adaptive linear regressor into a single reusable graph. Along the way the underlying adaptive regressor gains multi-output river support, a new SeqSeqSampler unit and a learn-side Flatten wrapper are introduced, and SelfSupervisedRegression/LRR get an automatic block-based channel clustering option plus an identity pass-through before fitting.

Changes

New

• ezmsg.learn.collection.SampleAdaptRegressor (src/ezmsg/learn/collection/sample_adapt_regressor.py): collection composing ResampleUnit, SeqSeqSamplerUnit, optional Window + Flatten, and AdaptiveLinearRegressorUnit. Routes labels through resampling against the signal time base, drives training from triggers, and switches between direct inference and windowed/flattened inference based on decode_window_dur.
• ezmsg.learn.process.seqseqsampler (src/ezmsg/learn/process/seqseqsampler.py): buffers continuous signal and value streams keyed on a shared time base, then emits per-trigger AxisArray samples whose attrs["trigger"].value carries the matching label window. Handles off-by-one mismatches by picking the best-aligned subwindow (bounded by MAX_ONE_TO_ONE_SAMPLE_MISMATCH) and drops/defers triggers that fall outside the buffered span. Includes max_buffer_dur cap on per-stream memory.
• ezmsg.learn.process.flatten (src/ezmsg/learn/process/flatten.py): thin wrapper around ezmsg.sigproc.flatten.FlattenTransformer that detects the (win, time, ch) windowed-feature case and injects a structured lag CoordinateAxis (with integer lag and "t-i" label fields) into the inner sample dim before delegating, so the merged-axis struct in the output carries real lag metadata ("t-2/c0" style labels). Outside the lag case it delegates unchanged.

AdaptiveLinearRegressor

• Multi-output river path: when the trigger target has multiple channels the transformer builds a per-label dict of river GLMs and routes learn_many/predict_many per output, preserving label order from the target axis.
• Loads pickled multi-output models (dicts of river GLMs), applies l2/optimizer overrides across all sub-models, and deep-copies optimizers so each sub-model has independent state.
• New _normalize_axis_label helper handles structured-dtype channel axes (extracting "label" sub-fields and unwrapping numpy scalars) so dataframe construction and prediction-template labels work with composed axes coming out of the new Flatten.
• Inference now runs as a pass-through (returns empty AxisArray) until a model exists rather than requiring template to be set first, and builds a prediction template from the signal when no fit-time template is available.

SelfSupervisedRegression / LRR (src/ezmsg/learn/process/ssr.py)

• New block_size setting on SelfSupervisedRegressionSettings: when channel_clusters is None, auto-generates contiguous block clusters of this size. Threaded through _validate_clusters, the SSR compute path, and LRRTransformer.
• LRRTransformer._process no longer raises before fit — it lazily initialises an identity AffineTransformTransformer (respecting block_size/min_cluster_size) so signal flows unchanged until weights are learned.

Dependency bump and API updates

• pyproject.toml: ezmsg>=3.9.0, ezmsg-baseproc>=1.7.0, ezmsg-sigproc>=2.23.0.
• Drop zero_copy=True from RNNUnit, TorchSimpleUnit, TorchModelUnit, and TransformerUnit signal subscribers to match the new base interface.
• tests/integration/conftest.py: NoiseSrc.OUTPUT_SIGNAL switched from ez.OutputStream to ez.OutputTopic.

Tests

• New tests/unit/test_sample_adapt_regressor.py, tests/unit/test_seqseqsampler.py, tests/unit/test_flatten.py.
• Expanded tests/unit/test_adaptive_linear_regressor.py covering multi-output river training/prediction, structured-axis label handling, and pre-fit pass-through.
• Expanded tests/unit/test_ssr.py covering block_size clustering and LRR identity pass-through before fit.