Add explicit channel-aware canonical dataset export to OME-TIFF or OME-Zarr #171

Goal

Allow a heterogeneous acquisition to be copied into a canonical MorphoBricks-ready dataset when biological repeats were acquired with different channel counts or channel orders.

Examples:

• repeat 1: [DAPI, ACTIN, PAXILLIN]
• repeat 2: [ACTIN, DAPI, PAXILLIN]
• repeat 3: [DAPI, ACTIN, PAXILLIN, unused]

The user should be able to explicitly map these to one common output schema, e.g.
[DAPI, ACTIN, PAXILLIN], with the fourth channel of repeat 3 deliberately dropped.

Existing code to reuse

The current convert path is deliberately narrower:

• it writes named-axis OME-TIFF copies;
• it corrects/records axis interpretation;
• it preserves pixel values and existing channel order;
• it has strong source fingerprinting, planning, space checks, conversion records, crash-safe writes and reuse.

Do not bolt silent channel manipulation into the normal preprocessing run. Extend/reuse the conversion infrastructure or introduce a clearly named canonicalization operation with the same safety guarantees.

Required plan

Before writing anything, show a table per condition / biological repeat (or the narrowest required scope):

repeat      source channels              output channels             action
rep1        0:DAPI 1:ACTIN 2:PAXILLIN    DAPI ACTIN PAXILLIN         unchanged
rep2        0:ACTIN 1:DAPI 2:PAXILLIN    DAPI ACTIN PAXILLIN         reorder 1,0,2
rep3        0:DAPI 1:ACTIN 2:PAX 3:XYZ   DAPI ACTIN PAXILLIN         drop 3

Every reorder/drop must be explicit and user-confirmed.

Never:

• infer a reorder from image appearance and apply it automatically;
• fill a missing channel;
• duplicate a channel to make counts match;
• change pixel values silently.

If a required canonical channel is genuinely missing, refuse that mapping unless a future design explicitly supports missing-channel datasets.

Output formats

Support a canonical copy in one or both standard forms:

• OME-TIFF;
• OME-Zarr / OME-NGFF.

The format choice is operational; the canonical channel identity/order must be identical.

OME-Zarr export was previously removed from the old conversion path. If reintroduced here, do it as an explicit canonical-export feature with tests and provenance, not as a hidden preprocessing optimization.

Provenance

For every output image record:

• source fingerprint/content identity;
• source axes and channel names;
• canonical output axes;
• source-channel -> output-channel mapping;
• dropped source channels;
• canonical names;
• output format;
• output fingerprint/hash;
• reason/scope for a non-identity mapping.

The copy tree must remain outside the source experiment and must never modify originals.

If OME-Zarr source label images/masks exist, their treatment must be explicit: preserve/copy them correctly or refuse with an actionable message. Never silently lose masks.

Notebook

Offer this only when the folder/channel checks reveal a correctable heterogeneous layout, or under an advanced Create canonical copy card.

The user should be able to:

1. choose the canonical channels;
2. define mappings for affected biological repeats;
3. preview the plan and estimated disk use;
4. choose OME-TIFF or OME-Zarr where supported;
5. write the copy;
6. switch preprocessing to the canonical copy and re-check it.

Acceptance

• Reorder and drop operations are explicit and previewed before writing.
• Output datasets have one canonical channel count/order/name schema.
• Unchanged images may be reused/copied efficiently but provenance remains complete.
• No missing channel is filled or guessed.
• Original images are never changed.
• Interrupted conversion is crash-safe and resumable/reusable using existing conversion principles.
• A conversion record fully describes every channel transformation.
• OME-TIFF and OME-Zarr outputs, when both supported, carry equivalent canonical metadata.
• Final canonical dataset passes preprocessing's normal channel-count/name/order checks.