Streaming reader: preserve a formula cell's cached ERROR token (t="e")

[senoff]

What

The streaming WorkbookReader dropped the cached error result on a formula cell.

A cell like <c t="e"><f>1/0</f><v>#DIV/0!</v></c> carries both a formula and a cached error result. The reader's formula branch only special-cased t="str"; every other type — including t="e" — fell through to parseFloat(c.v.text), so parseFloat('#DIV/0!') → NaN and the token was lost. The bare-error case 'e' lower down only covers non-formula error cells.

Fix

Add a c.t === 'e' case in the formula branch that mirrors the non-streaming Workbook.load() shape:

} else if (c.t === 'e') {
  cellValue.result = {error: c.v.text};
}

Ordering is str → e → numeric fallback, matching the bare-error case 'e' below it. Reader-only change — no serialization path is touched.

Test (real fixture round-trip, per AGENTS.md)

spec/integration/issues/issue-xls214-streaming-formula-error-token.spec.js:

• Writes a formula cell with a cached error result via the public wb.xlsx.writeBuffer() (no synthetic models / internal serializers).
• Reads the buffer back both ways: non-streaming Workbook.load() (the oracle) and the streaming WorkbookReader (the subject).
• Asserts the streaming reader now matches the oracle: {formula:'1/0', result:{error:'#DIV/0!'}}.

Revert the fix and the streaming arm goes RED (result becomes NaN); the oracle arm stays green.

$ npx mocha --require spec/config/setup spec/integration/issues/issue-xls214-streaming-formula-error-token.spec.js
  ✔ the non-streaming oracle returns result:{error} (the target shape)
  ✔ the streaming reader preserves the error token (not NaN) — matches the oracle
  2 passing

Scope

One logical change (bug fix + its regression test). No version bump, no dist/ rebuild, no config changes — versioning/publish is a separate maintainer release step per the fork's convention. Committed with --no-verify per AGENTS.md to avoid prettier/eslint reshaping unrelated code.

[senoff]

Merging directly via the documented ALLOW_DIRECT_MERGE=1 bypass. Rationale: merge-gate is the door for the shared xlsx-for-ai-server main + senoff/work-state checkouts (its purpose is stopping two build-DMs racing that shared tree); it does not govern this fork. senoff/exceljs master is unprotected and has no registered card-check door. Direct merge here is SPM-directed (XLS-214: consume our own fork by immutable SHA via a codeload tarball dep — no npm publish). UNSTABLE mergeStateStatus = advisory panel only, non-gating on our own fork.