Preserve XLSX workbook fidelity

[rastagan-git]

Summary

• preserve supported XLSX workbook content while overlaying only intentional enrichment cells on one selected worksheet
• add exact --sheet selection plus pre-provider checks for unsafe OOXML parts, merged table cells, output capacity, path aliases, and checkpoint collisions
• harden atomic no-overwrite publication and delete only checkpoints owned by the current run
• extend the bilingual docs, project Skill, workbook inspector, and regression suite

Root cause

XLSX input was reduced to a pandas DataFrame and then written as a brand-new workbook. That flat-table round trip silently discarded worksheets, names, formulas, formatting, dimensions, and workbook features.

The new path keeps an immutable source package, records only cells intentionally touched by a workflow, serializes the selected worksheet, restores untouched formula caches, and overlays the controlled parts back onto the source package.

Fidelity and safety boundaries

• non-target worksheet content and supported workbook parts are retained
• legitimate formulas and cached results remain unless an output field explicitly targets that cell
• newly fetched formula-like text remains escaped
• merges intersecting the selected table are rejected before provider access; unrelated merges are preserved
• known unsafe features and OOXML parts dropped by an in-memory trial round trip fail before provider calls
• the optional Excel calculation chain may be removed for spreadsheet software to rebuild

Validation

• 76 passed
• Ruff lint and format checks passed
• project Skill validation passed
• offline all-skipped run against the repository's real 示例data.xlsx passed and the result reopened cleanly
• isolated wheel build passed

Addresses the P1 workbook-fidelity portion of #8. Retrieval timestamps, --no-odor schema behavior, python -m aromanexus, and existing-CAS confirmation remain follow-up work.