perf: move Community batch writes onto dense native rows

[wolfiesch]

Summary

Moves the common Community append() / write_rows() batch path from per-cell nested BTreeMap insertion to compact native dense rows while retaining sparse last-write-wins overlays for later edits and formats.

The stale portion of the audit is corrected here: current Community already crosses PyO3 once per append/bulk buffer. This PR therefore claims removal of the native per-cell tree storage, not removal of per-cell FFI.

Implementation

• stage each Python value/style grid once, then commit monotonic batches as contiguous native rows
• retain the existing sparse path for non-monotonic writes
• merge dense rows with sparse edits deterministically at emission
• preserve dense values when a later format materializes an overlay
• include dense cells in dimensions, calc chains, autofilter evaluation, and streaming-mode eligibility checks
• validate grid bounds before mutating the worksheet
• retain reproducible public-API and direct-Rust benchmark harnesses in the repository

Production-profile evidence

Evidence workflow: https://github.com/SynthGL/wolfxl-community/actions/runs/33599139211
Artifact: https://github.com/SynthGL/wolfxl-community/actions/runs/33599139211/artifacts/9834572130

Method:

• exact base: cc624eabd6f65301ccc5d9adb24a6cb6b0d271d5
• measured candidate production code: 5c6a31b888bdefd50c981703e9ab4fbe3626912b
• both wheels built with fat LTO and one codegen unit
• 25 alternating, isolated-process observations per public lane
• 400,000 cells in each timing lane
• median, p95, and coefficient of variation reported
• three isolated peak-RSS observations on 600,000 unique strings
• same-binary direct Rust sparse-vs-dense storage/emission allocation benchmark
• fixed ZIP epoch and byte-for-byte archive comparison

Public Python API

lane	phase	before median	after median	change	speedup	p95 before/after	CV before/after
append numeric	save	132.112 ms	93.174 ms	-29.5%	1.418×	162.376/97.534 ms	7.4%/2.3%
append numeric	total	162.390 ms	119.485 ms	-26.4%	1.359×	197.452/125.678 ms	7.5%/2.4%
append mixed	save	162.512 ms	115.439 ms	-29.0%	1.408×	175.593/128.472 ms	3.5%/4.1%
append mixed	total	232.207 ms	177.136 ms	-23.7%	1.311×	248.709/197.366 ms	3.3%/3.6%
write_rows numeric	save	117.721 ms	77.914 ms	-33.8%	1.511×	120.056/80.645 ms	2.0%/1.4%
write_rows numeric	total	141.395 ms	97.045 ms	-31.4%	1.457×	145.038/100.290 ms	2.0%/1.4%
write_rows unique strings	save	194.676 ms	149.051 ms	-23.4%	1.306×	214.598/152.604 ms	6.3%/2.0%
write_rows unique strings	total	324.787 ms	270.984 ms	-16.6%	1.199×	344.838/276.234 ms	5.5%/2.1%

The supported claim is therefore 1.31–1.51× faster save/flush and 1.20–1.46× faster end-to-end on the measured Community append/bulk lanes—not the audit's unmeasured 1.5–3× range.

Direct Rust phase and allocations

Sparse vs dense native storage plus sheetData emission, 400,000 numeric cells:

metric	sparse	dense	change
median	56.334 ms	35.209 ms	-37.5% / 1.600×
allocation calls	1,283,355	1,220,037	-4.9%
allocated bytes	140,439,918	92,136,494	-34.4%

Peak RSS

600,000 unique-string cells, three isolated samples per side:

• before: 339,896 / 339,896 / 339,888 KiB
• after: 265,116 / 264,992 / 264,984 KiB
• median: 339,896 → 264,992 KiB (-22.0%)

Correctness and validation

• every corresponding base/candidate XLSX archive in the evidence matrix is byte-identical
• dense and sparse direct XML hashes and byte lengths match
• dense blank/styled-blank rules are covered
• sparse cell/value/style overlays remain last-write-wins
• dimensions and calc-chain scans include dense cells
• autofilter evaluation reads the logical dense+sparse cell
• grid bounds are validated before worksheet mutation
• Python CI passed on the measured code: https://github.com/SynthGL/wolfxl-community/actions/runs/33599145074
• final workflow-free CI passed: https://github.com/SynthGL/wolfxl-community/actions/runs/33599665535

The temporary benchmark workflow was removed after publishing the evidence artifact.