After noticing that this method uses a lot of memory/takes a lot of time in the Rails app at work, I used Claude to optimize this and then verify that the output was identical from the released gem vs this change on a number of different files.

This should hopefully help resolve some long-standing issues with the performance/memory usage of the gem.

Benchmarked on a generated 20,000-row x 40-col xlsx (mixed shared-string, numeric, and empty cells). Median full-parse time dropped ~1.72x (1.85s -> ~1.08s) and allocations dropped ~3.75x (28.6M -> 7.63M), with byte-identical output verified on both default- and prefixed-namespace files across rows, simple_rows, and rows_with_meta_data (including the self-closing-row + meta-data path).

Changes, in order of impact:

• Resolve the namespace prefix once via a namespace_resolved flag instead of re-checking node.namespaces on every node. Worksheets use a default namespace, so the old prefix.empty? guard never latched and allocated a namespaces hash for every node in the stream (660,002 calls -> 2). This is the bulk of the wall-clock win.
• Hoist node.name/node.node_type into locals (each was read up to 4x per node in the if/elsif chain) and reorder branches so the hottest nodes (<v>/<c>) are tested first.
• Fetch cell attributes with three individual node.attribute('t'/'s'/'r') C calls instead of node.attributes. Reader#attributes is attribute_hash.merge(namespaces), so it built and merged a namespaces hash on every call. With hundreds of thousands of cells the per-cell Hash allocation dominates, so three cheap lookups are both faster and leaner than building and indexing a hash.
• Use node.attribute_hash for the row-opener node (only ~one per row): this avoids the namespaces merge that node.attributes performs while preserving the exact meta-data row hash. Safe because xlsx row elements never declare their own namespaces (the namespace hash is empty for them), so the result is identical -- confirmed byte-for-byte including the self-closing-row + meta-data path.
• In fill_in_empty_cells, drop the redundant .to_a on the column range and use delete_suffix instead of gsub to strip the row number.