Fix #42: extend pivot cache preservation to eliminate Excel repair warning

[senoff]

Closes #42.

Summary

Programmatically-generated pivot tables previously emitted fully-enumerated <sharedItems> for every source field, including non-axial columns (value fields and unused columns like long-text "Description"). On open, Excel flagged the file with:

Repaired Records: PivotTable report from /xl/pivotCache/pivotCacheDefinition1.xml part (PivotTable cache)

This refines the existing pivot-cache hybrid-preservation work shipped in 4.4.0-protobi.9 (#41) by scoping sharedItems enumeration to fields actually used as pivot axes (rows / columns / pages). Non-axial fields now emit a lightweight <sharedItems .../> declaration with type hints from column inspection, and pivotCacheRecords inlines their values (<n>/<s>/<m>) rather than referencing them via <x v="N"/>.

Also corrects a related bug in pivotCacheDefinition: recordCount was cacheFields.length + 1 when it should match the data row count declared on <pivotCacheRecords count="...">. Excel's strict cache validator flags the mismatch.

Root cause

lib/doc/pivot-table.js called makeCacheFields(sourceSheet, allHeaderNames) — every source column got sharedItems aggregated. For a 1500-row workbook with a long-text "Description" column not used in the pivot, this produced ~915 KB of irrelevant <sharedItems> content in the cache definition (and triggered the repair warning). Only fields actually projected onto pivot axes need indexable shared items; everything else can be inlined into the records part.

Change shape

• lib/doc/pivot-table.js: pass only axial field names (rows + columns + pages) to makeCacheFields; inspect non-axial columns for type hints (containsBlank, containsString, containsNumber, containsInteger, longText).
• lib/xlsx/xform/pivot-table/cache-field.js: when sharedItems === null, render <sharedItems .../> with hint-derived attributes only — no enumeration.
• lib/xlsx/xform/pivot-table/pivot-cache-definition-xform.js: derive recordCount from sourceSheet.getSheetValues().length - 2 to match the records count.
• spec/integration/issues/issue-42-pivot-cache-repair.spec.js: 8 new regression assertions.

pivotCacheRecords's existing renderCell already handles sharedItems === null by inlining <n> / <s> / <m> values, so no change was needed there.

Result on the issue's reference fixture

Metric	Before	After
pivotCacheDefinition1.xml size	~915 KB	~4 KB
recordCount	6	1500
Excel repair warning on open	yes	(expected: no)

Test plan

• npm run test:unit — 884 passing
• npm run test:integration — 208 passing (12 pivot tests, including 8 new for [BUG] Pivot cache triggers Excel repair #42)
• npm run test:end-to-end — 1 passing
• Existing pivot integration tests (pivot-tables.spec.js, pivot-tables-with-count.spec.js) still green — no regressions
• Open the resulting .xlsx from the new spec in Excel and confirm no repair prompt (soffice was not available on the dev machine — byte-shape assertions used as a proxy; happy to add a soffice --headless --convert-to xlsx smoke step in a follow-up if useful)

Notes for review

• The fix is limited to the programmatic pivot path. Round-trip preservation (loading an existing pivot via _processPivotCache* and re-emitting raw) is untouched; that path already passes through verbatim XML.
• cache-field.js is conservative about declaring type attributes on lightweight sharedItems: it only emits containsNumber="1" etc. when the column is unambiguously numeric. Mixed-type columns get a bare <sharedItems/> (or with containsBlank/longText hints) and Excel infers from the inline records — this avoids attribute/record disagreement which is itself a repair trigger.
• A small amount of pre-existing-but-flagged lint in the touched files was cleaned up so the pre-commit hook passes (no --no-verify used): a redundant comma after a thrown-error message and an unused destructured defaultValue. Both are cosmetic.

🤖 Generated with Claude Code