[okunator]

The segmentation requirement capped Torch below 2.10 solely to match the validated development lock. There is no demonstrated compatibility problem at that boundary. Allow torch>=2.9.1,<3.0, matching the upstream model package's major-version limit, and retain the existing Torch 2.9.1/torchvision 0.24.1/Triton 3.5.1 lock.

Update the 0.3.0 changelog, installation guidance and short maintenance/audit notes to distinguish package compatibility from the reproducible development environment. No library code, CI scopes, new baselines or permanent validation tooling.

Compatibility evidence:

• Published cellseg-models-pytorch 0.1.30 requires Torch >=2.0,<3.0; timm 1.0.20 and torchdata 0.11.0 impose no narrower upper bound.
• Complete segmentation/dev profile on macOS ARM64, Python 3.12.13: changed only Torch 2.9.1→2.14.1 and torchvision 0.24.1→0.29.1; all other packages constrained to locked versions. Full extra resolution and installed requirement check passed.
• All 39 existing offline model/reconstruction tests passed on Torch 2.14.1, including five model families, forward/backward passes, inference, per-model mappings and exact .pt tensor round trips.
• The pinned published HGSC checkpoint matched the existing archived reference: exact labels/shapes/dtypes and unchanged float tolerances (rtol=1e-4, atol=1e-5). Existing checkpoint/image/reference hashes verified; no expected outputs regenerated. Archived checker/reference used only in temporary files.
• Stable Torch/torchvision wheels verified for Python 3.10–3.12, Linux x86_64/aarch64 and macOS ARM64. GPU computation remains unverified.
• Existing PR validation: all 15 jobs passed. The fresh Linux full-profile installation selected Torch 2.10.0/torchvision 0.25.0 and passed model/WSI imports and CPU training.
• All dependency lock records/versions/hashes unchanged; only the root requirement metadata changes. Hooks, seven release guards and v0.3.0 metadata checks passed.

References: model package metadata, Torch 2.14.1 metadata, torchvision 0.29.1 metadata.

Publication remains on hold for the user's release-note review. This PR does not tag, merge or publish the release.