feat(dynacell): pix2pix3d nucleus predict leaves wired to modernized GAN checkpoints

[alxndrkalinin]

Stacked on top of #449 (base = dynacell-cellpose-watershed-eval).

Prepares the full nucleus predict matrix for the three modernized
pix2pix3d_unetvit GANs trained to completion (iPSC / A549 / joint,
Run-D recipe, 40 ep). 13-file delta on top of the eval branch — which
already carries the modernized engine + overlays and the
hardware_predict_any_gpu migration, so this is config-only.

Changes

1. Wire trained checkpoints into the 8 existing predict leaves (ba72c77)
Replace the REPLACE_ME / stale paths with the best loss/validate_ema
checkpoints from the modernized runs:

trained_on	checkpoint	job
iPSC	epoch=15-step=51200	33501549
joint	epoch=39-step=214400	33502019

Predict uses the EMA generator (use_ema_at_predict=True, engine default).

2. Fix a latent TargetNotFoundError (same commit)
The six a549 predict leaves lacked dataset_ref: {target: h2b}. The A549
manifest keys nucleus by gene (h2b), so the iPSC-side nucleus target
inherited from targets/nucleus.yml would fail to resolve. Bug is present
on the base branch (leaves were never wired, so it was never hit).

3. Add the 4 missing a549-trained predict leaves (a55517e)
a549_mantis had no predict leaves. Added ipsc_confocal + a549 h2b
mock/denv/zikv, best checkpoint epoch=39-step=86400 (job 33501550).
Output stores use the _a549trained infix (fnet3d_paper convention) to
avoid colliding with iPSC-trained predictions in the shared dirs.

Validation

All 12 leaves compose and resolve via submit_benchmark_job.py --print-resolved-config (exit 0), each to the correct test store
(cell.zarr for iPSC; H2B_{mock,DENV,ZIKV}.ozx for a549); all three
checkpoints verified on disk.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR updates the DynaCell benchmark nucleus / pix2pix3d_unetvit predict leaves to (1) point at the finalized modernized GAN checkpoints, (2) fix A549 dataset_ref.target resolution by overriding the nucleus target to h2b, and (3) add the previously-missing A549-trained predict leaves with non-colliding output stores.

Changes:

• Replace placeholder/stale ckpt_path entries with the selected modernized Run-D checkpoints (iPSC-trained and joint-trained leaves).
• Fix A549 predict leaves by overriding benchmark.dataset_ref.target: h2b so manifest resolution succeeds.
• Add the missing a549_mantis/ pix2pix3d_unetvit nucleus predict leaves (including _a549trained output store naming).

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/joint_ipsc_confocal_a549_mantis/predict__ipsc_confocal.yml	Update joint-trained predict leaf to the selected modernized checkpoint.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/joint_ipsc_confocal_a549_mantis/predict__a549_mantis_zikv.yml	Add dataset_ref.target: h2b override and update checkpoint for A549 ZIKV cross-predict.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/joint_ipsc_confocal_a549_mantis/predict__a549_mantis_mock.yml	Add dataset_ref.target: h2b override and update checkpoint for A549 mock cross-predict.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/joint_ipsc_confocal_a549_mantis/predict__a549_mantis_denv.yml	Add dataset_ref.target: h2b override and update checkpoint for A549 DENV cross-predict.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/ipsc_confocal/predict__ipsc_confocal.yml	Wire iPSC-trained predict leaf to the selected modernized checkpoint.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/ipsc_confocal/predict__a549_mantis_zikv.yml	Add dataset_ref.target: h2b override and use iPSC modernized checkpoint for A549 ZIKV cross-predict.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/ipsc_confocal/predict__a549_mantis_mock.yml	Add dataset_ref.target: h2b override and use iPSC modernized checkpoint for A549 mock cross-predict.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/ipsc_confocal/predict__a549_mantis_denv.yml	Add dataset_ref.target: h2b override and use iPSC modernized checkpoint for A549 DENV cross-predict.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/a549_mantis/predict__ipsc_confocal.yml	New A549-trained leaf predicting on iPSC test set, using _a549trained output naming.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/a549_mantis/predict__a549_mantis_zikv.yml	New A549-trained leaf for A549 ZIKV test set with dataset_ref.target: h2b.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/a549_mantis/predict__a549_mantis_mock.yml	New A549-trained leaf for A549 mock test set with dataset_ref.target: h2b.
applications/dynacell/configs/benchmarks/virtual_staining/nucleus/pix2pix3d_unetvit/a549_mantis/predict__a549_mantis_denv.yml	New A549-trained leaf for A549 DENV test set with dataset_ref.target: h2b.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.