v0.23.1

xl 0.23.1 — the 0.23.0 dogfood: typed #N/A on legacy lookups, the batch putf parse gate and formula-unparseable lint, Excel's General text rule, General and scaling commas in custom formats, rsvg-convert on stdin

Wave 30 — the 2026-09-15 dogfood of 0.23.0 (103 probes across the CLI, the scripting library and
the tjc-modeling template lane; every recorded 0.19.x quirk re-tested as fixed) filed seven
issues, run as six worktree-isolated clusters, each adversarially reviewed with the LibreOffice
oracle: the typed #N/A on legacy lookups (#662), the batch putf parse gate and the
formula-unparseable lint rule (#663), the rsvg-convert adapter (#664), Excel's number → text rule
(#665), General and scaling commas inside custom formats (#666), three CLI polish items (#667)
and the skill drift the dogfood read (#668).

• xl lint gains formula-unparseable (#663): a repair-tier finding for a <f> whose text
  the parser refuses in the classes Excel certainly repairs — a text that ends early by its own
  shape (an open (, an unterminated string or sheet name, a trailing operator), a ] or }
  closing a (, more than 8192 characters. A complete text the parser cannot finish (NOT, a
  legal defined name read as its prefix operator) is not a finding. Nesting is not judged: the parser's
  128-level depth budget counts every chained operator segment, so a flat 130-term B2+B3+…
  chain Excel opens intact fails it while a 100-deep nest passes (PR #679 review; the parser
  side is #680). One finding per sheet names the
  cells and quotes the first offending formula (capped at 80 characters); the DOM and SAX
  scanners fold identically. Shapes the parser does not yet accept but Excel opens — TRUE()
  as a call, an extra closer after a complete expression, a , or space inside parentheses
  (Excel's union and intersection reference operators) — are left to xl audit. Only the CLI
  wires the rule; the one-argument WorkbookLint.lint(path) library entry cannot see the
  parser and does not carry it (#674). A first cut flagged Excel's union and intersection
  operators as repairs; the adversarial re-review caught it with the LibreOffice oracle and the
  rule was narrowed before release.
• General inside a custom number format (#666): the keyword renders the value in General
  form wherever it sits in a section — General"A" on 2021 is 2021A (the FY-suffix idiom),
  "FY"General is FY2021, General" units" on 12.5 is 12.5 units, General;(General)
  parenthesises negatives — where it used to print the word. Quoted "General" and \G
  escapes stay literal; the whole-code General path (ECMA-376 §18.8.30) is unchanged.
• Thousands-scaling commas (#666): a comma after the last digit placeholder of a section
  divides by 1000, once per comma, before rounding and grouping — $#,##0.0,,"mm" on 1,500,000
  is $1.5mm (the IC-exhibit idiom), #,##0, on 1,234,567 is 1,235, 0.0,% still applies
  the percent first. Commas between digits keep grouping. Display only: stored format codes are
  byte-identical. FormatCodeParser.FormatToken gains Scale and General;
  FormatPattern.hasThousands is now true only for grouping commas.
• -o on a read-only verb is a usage error that names the flag (#667): xl -f f.xlsx view A1:B2 -o out.xlsx says view is read-only and does not take -o/--output with the hint to
  drop it (exit 2, USAGE, the same in the --json envelope; -i/--in-place likewise), where
  decline used to report Unexpected argument: view. Read-only is read from the verb registry
  (Needs.output false on every form of the verb); new, whose output is its positional, is
  excluded.
• Breaking: a VLOOKUP, HLOOKUP or MATCH miss is the typed #N/A (#662): the three legacy
  lookups raised a host failure ("VLOOKUP exact match not found") where XLOOKUP returns the error
  value, so IFNA(VLOOKUP(miss),0) left its cell uncached and ISNA(VLOOKUP(miss)) answered
  FALSE — =IF(ISNA(VLOOKUP(2030,A3:B7,2,FALSE)),0,1) cached 1 where Excel shows 0, with no
  warning and nothing in xl audit. Exact and approximate misses now raise #N/A on the error
  channel with the diagnostic kept as context; IFNA, ISNA, ISERR, IFERROR, ISERROR, ERROR.TYPE
  and an IF condition all see it; an unguarded miss caches #N/A (the cell is no longer
  "uncached" to audit, and recalc counts it as an error value rather than a RECALC_ERRORS
  warning); xl eval prints #N/A (error) at exit 0. An index below 1 is #VALUE! and one past
  the table #REF!, as in Excel, in place of the former host failure. The lookup-miss law
  (legacy miss ≡ XLOOKUP miss ≡ NA()) is pinned in ErrorValueLawsSpec; LibreOffice displays
  the recalculated book identically.
• Breaking: batch putf parses every formula before writing (#663): the single-cell,
  values[] and from-dragging shapes, in memory and under --stream, and --dry-run, refuse
  a formula the parser rejects with BATCH_OP_INVALID (exit 2), the 1-based op index and the
  same caret diagnostic the putf verb prints; an unknown function is refused with its
  "Did you mean" suggestion, as the verb does; no output file is created. The --stream putf
  verb, which had no gate at all, now raises the in-memory verb's FORMULA_ERROR. =SUM(A1:A2 used to
  reach the file at exit 0 with xl lint clean and Excel's repair prompt waiting. The gate is the
  evaluator's parser, so the shapes it does not yet accept — TRUE()/FALSE() as calls, array
  constants, structured references, the union and intersection reference operators — are now
  refused in batch as they always were in the verb (#669 tracks the parser).
• Breaking: a number becomes text by Excel's General rule (#665): &, CONCATENATE,
  text-typed function arguments, a numeric literal in a text position and TEXT(x,"General")
  render at most 15 significant digits with trailing zeros stripped, plain while the unsigned form
  fits 20 characters and in E notation past it (1E19 is 10000000000000000000, 1E20 is
  1E+20, 1/3 is 0.333333333333333), where BigDecimal.toString carried the stored scale —
  ="Total "&B8 shipped Total 1070.0, =A1&"" on a <v>2.0</v> cell 2.0, =1/3&"" 34
  digits, and =(A1&A2)+1 failed on 2.03.0 where Excel returns 25. Dates still concatenate as
  their serial (#561); TEXT with an explicit format is unchanged; =2.50&"" re-prints
  byte-identically (the literal is no longer folded to a string). One rule,
  NumFmtFormatter.generalText (now public beside isGeneralCode), backs every site; LibreOffice
  agrees on the plain range (its own & conversion diverges from Excel in E notation, noted in
  docs/LIMITATIONS.md).
• describe --full lists the lifted print names (#667): the reader lifts a modelable
  sheet-scoped _xlnm.Print_Area / _xlnm.Print_Titles into the sheet's page setup, so the
  loaded name table describe --full printed (text and --json) omitted them while the
  metadata-only describe and names listed them. Both forms now list the effective table a
  write emits; PrintNames is private[xl] for it.
• Skills (#668): the xl-cli skill's rasterizer text names the four-backend automatic chain
  (ImageMagick is opt-in only, the sole webp backend), evala --at is described as positioning
  the displayed grid (the verb writes no file), batch values is a flat row-major array; the
  xl-scripting quick reference carries the real shapes (seedDataTablesReport() returns a
  DataTableSeedReport, editIn/EditScope.of take a SheetName, Edit.Put takes a Loc and
  a format), the post-solve checklist names excelErrors, and certified is dated to 0.19.2.
• view --format png on the default rasterizer chain (#664): the rsvg-convert adapter passed
  a bare - as the input file; librsvg 2.40 through 2.54 reject it (Error opening file …/-),
  so the chain fell through to RASTERIZER_UNAVAILABLE on Debian images although
  xl rasterizers listed the binary. The SVG is now piped on stdin with no positional, which every
  librsvg reads; verified on 2.40.20, 2.54.7 and 2.61.3.
• import CSV date columns carry a date format (#667): a column detected as ISO dates was
  typed as dates but written with no number format, so the serial (46037) showed where
  put "2026-01-15" shows the date; import now attaches the same Date format detection picks,
  for the positioned import and --new-sheet alike. The O(1) --stream --new-sheet path into a
  new workbook still writes no styles (#675).