Drop the scaling from attrs once it has been applied

[cboulay]

gain and offset describe stored counts → values. Once the multiply has happened those counts are gone, and every later stage — a filter, a decimation, a re-reference — makes them less true still. Carrying them onward offers a downstream consumer a factor that looks authoritative and reconstructs nothing. Re-digitizing is a real need, but it happens after transformations these numbers never saw, so it takes its range and resolution from the caller instead.

Applying a scaling now consumes it: the output carries unit and no nwb_scaling_* keys. Attrs this module doesn't own are left alone.

reader   : {nwb_scaling_gain: 2.5e-07, nwb_scaling_offset: 0.0,
            nwb_scaling_unit: 'volts', nwb_scaling_applied: False,
            nwb_scaling_voltage: True}
converted: {unit: 'microvolts'}

Idempotency gets stronger, not weaker

It used to rest on the applied flag. Now there is simply nothing pending to apply twice — a second transformer in a chain returns the message untouched. Still never inferred from the dtype.

One consequence worth reviewing

The reader always reports applied=False (describe_stream_scaling never applies), so this transformer was the sole producer of applied=True. Stripping removes that state from the wire entirely — which would have silently killed target_unit retargeting of an already-converted message, a documented and tested feature.

It only ever needed the unit, though: "a prefix change on top of a known unit is well defined." So it now reads attrs["unit"] via a new _scaling_from_unit, and the feature survives with the same semantics. A non-voltage unit (a cursor stream in pixels) returns None there and passes through rather than being reinterpreted.

This is the judgment call in the PR. The alternative was to let chained retargeting go, since nothing produces applied=True any more. I kept it because your reason for stripping — stale gain/offset — doesn't argue against it, and dropping a documented feature silently seemed worse. Easy to remove if you'd rather.

Two knock-ons: scaling_fingerprint had to learn the same fallback, or a stream whose unit changed mid-run would keep a plan built for the old one; and a unit-carrying message from a non-NWB source is now retargetable, which matches the module's stated "graph whose source is not always a file" use case.

Also corrected

Two stale references found while in here, both pre-existing:

• The module docstring said the slicer "applies a file's conversion on read by default". It never applies.
• An error message advised reading with apply_conversion=False — not a parameter that exists.

Tests

TestSpentScalingIsDropped (only unit survives; unrelated attrs untouched; a second pass can't double-scale; a non-voltage stream still gets its file conversion) and TestRetargetOnUnitAlone (unit-only retarget; pixels left alone; a changed unit rebuilds the plan). Existing tests that asserted the restamped scaling now assert it's gone.

294 passed, 1 skipped, ruff clean. Verified on sub-CA001/ses-S006: convert → {unit: microvolts}, chained retarget to volts exact to 1e-6, and concat of two converted streams still merges.