fix: combined export_batch manifest lost all but the first model entry

[Dinghye]

Summary

Fixes a bug reported by a user: in multi-model combined exports (export_batch with a single output file), the manifest's models list ended up with only one entry — each checkpoint write effectively rewrote it instead of accumulating.

Root cause

CheckpointManager.combined_write_checkpoint returned the jsonable copy of the manifest that write_arrays operates on, not the original dict. run_pending_models rebinds its local manifest to that returned copy after each per-model checkpoint, while the _write_ckpt closure in BatchExporter._run_combined keeps serializing the stale original — so every model entry after the first was appended to a throwaway copy and dropped from both the in-memory result and the on-disk manifest. As a side effect, combined-mode resume would needlessly re-run already-completed models.

Fix

combined_write_checkpoint now merges the writer's path keys (manifest_path, npz_path, npz_keys, nc_path, nc_variables) back into the original manifest and returns that same object, preserving the in-place mutation contract. Per-item exports were never affected (each point's manifest is fully built before its single write).

Tests

• New tests/test_checkpoint_manager.py with two regression tests; the second faithfully mirrors the exporter's closure + run_pending_models loop structure. Verified both fail on the pre-fix code and pass with the fix.
• Full suite: 751 passed, 2 skipped (after merging latest main).