Multi-class merging with arbitrary class axis placement (resolves #20)

[the-lay]

What & why

Rebuilds the 2022 multiclass Merger prototype (PR #21, unmergeable since the src-layout move) on the current codebase. Resolves #20 — the only externally-reported feature gap, open since 2022: model outputs like (256, 256, 6) for input tiles (256, 256, 4) were rejected by Merger.add because the class axis could only be prepended.

Stacked on #29 (revival baseline) — merge that first; GitHub will retarget this PR automatically.

API (deliberately smaller than the #21 prototype)

• Merger(logits=N, logits_dim=d) — place the class axis at any position of the added tiles and merged output; negative indexing allowed. Default logits_dim=0 reproduces the previous prepend-only behavior exactly (verified by an equivalence test and the untouched pre-existing suite).
• Merger(ignore_channels=True) — expect tiles without the Tiler's channel dimension, for models that consume all channels and emit class logits (the [IDEA] Figuring out how to use this library with TensorFlow multi-class and binary classification #20 scenario) or a single map. Requires channel_dimension on the Tiler.
• merge(argmax=True) — argmaxes the logits axis wherever it lives.

Differences from the #21 sketch, for the record: logits keeps its name and validation (no logits_n rename), argmax stays a bool — #21's argmax: Optional[int] would have silently treated argmax=True as axis 1 — and no buffer renames (weights_sum/data_visits/save_visits untouched), so nothing breaks for existing users.

Implementation notes

• Weights and visits buffers deliberately have no class axis (the window weight is identical for every class); broadcasting happens once at normalization via np.expand_dims.
• With ignore_channels, the applied window drops the channel axis (which is always boxcar-weighted) and tile bboxes come from get_tile_bbox(..., with_channel_dim=False).
• README gains a multi-class segmentation example.

Verification

• 8 new tests (TDD — each observed failing first), including the exact [IDEA] Figuring out how to use this library with TensorFlow multi-class and binary classification #20 scenario, legacy-equivalence, non-trivial window + 50% overlap normalization, batch add, and validation errors
• Full suite: 48/48 pass, coverage 99%, ruff + format clean

🤖 Generated with Claude Code

https://claude.ai/code/session_01PvVX1JaNVaXJ2PJEu4KJzE