docs(xlsx): measure horizontal bar-chart category-count effect on baselines (#1621)

[developer0hye]

Summary

• Discovers and documents that a chart-XML-cache-only category-count patch is a complete no-op against native Excel: it re-derives the income chart's categories from xl/workbook.xml's defined-name array literal and its values from a live formula range, both outside the chart part, and overwrites the chart's own cache on open regardless of what the cache says. scripts/probe_harness.py patches exactly one part per variant, so this is a documented gap in the harness, not a new defect in shipped conversion code.
• Adds build_category_count_probe.py, which patches both parts together (the workbook-level array literal and the chart's c:val range) to actually change category count in native Excel, and uses it to collect a new native measurement: the income chart widened from 5 to 6 categories, closely bracketing the expense chart's own row height from the other side.
• Refines the working model from "round + chart-specific K" (the prior session's still-incomplete reading) to "floor, not round, plus a font-size-only K" — reproduces 18 of the 24 measured rows across three independent geometric configurations (income @ 5 categories x 3 sizes, expense @ 6 categories, income @ 6 categories), replacing the "K depends on which chart" framing with "K depends only on font size; only plot-rectangle edge rows deviate."
• Falsifies one specific edge-row hypothesis: naively snapping the plot rectangle's own continuous top/bottom edges to whole sheet points does not close the remaining gap and introduces a new regression on a previously-exact interior row.
• No runtime code changes — assets/validation/issue-1621/ only.

Related issue

Related: #1621

Testing

• python3 assets/validation/issue-1621/verify_measurements.py exits 0, confirming all 5 documented claims (whole-sheet-point quantization across 24 rows, the floor/round split, the floor+size-only-K model's 18/24 reproduction with its exact 6-row exception set, and the edge-snap falsification).
• python3 scripts/probe_harness.py assets/validation/issue-1621/category-count-cache-only-noop-probe.json --backend office2pdf vs --backend office reproduces the documented no-op finding directly (a real 15.78pt shift against our own converter, layout-identical against native Excel).
• Documentation-freshness review completed across three passes before this commit (numeric drift in a hardcoded "15/17" claim that should have been "18/24" computed dynamically, an incomplete exception-set enumeration, an overclaimed "byte-for-byte" identity, and a docstring that only generalized the illustrative N=6 case — all fixed and re-verified).
• No .rs/runtime config changes — tests skipped per project convention for non-runtime changes.

Visual impact

• No rendered PDF change
• Rendered PDF change or visual evidence added
• Reason: Adds only assets/validation/issue-1621/ measurement data, a probe-building script, and a verification script; no chart rendering or codegen path is touched, so there is no rendered PDF change to audit.

Checklist

• Commits include a Signed-off-by line
• PR scope contains one root cause
• Remaining converter or harness deviations each reference an open issue

🤖 Generated with Claude Code