linexcel 1.3.0

[1.3.0] — 2026-08-18

Added

• Power Query is part of the lineage (#34).
  A workbook fed by Get & Transform used to show the range the data landed in
  and nothing above it: no query, no M source, no table it read. Each query is
  now a node carrying its M source verbatim, the range it loads into, and the
  sources it names — so Source!A1:B4 → BusyProducts → Loaded!A1:B3 reads
  end to end. A source that lives in the same workbook (Excel.CurrentWorkbook)
  is linked to the table or defined name it reads; a query chaining off another
  is an edge between the two; a source outside the file (File.Contents,
  Web.Contents, Sql.Database, any *.Database/*.Feed connector) is named
  as a node and said, in the panel and in the warnings, to have been named
  rather than read. Connection-only queries are shown as loading nowhere
  instead of being dropped. Read from the DataMashup custom XML part, with
  xl/connections.xml and xl/queryTables/*.xml for the destination; two new
  stats, queries and queriesLoaded.
• An optional multimodal description of each sheet screenshot
  (#46).
  result.describe_screenshots(shots, ...) — --screenshots DIR --vision-docs
  on the command line — sends each rendered sheet to a model that can look at
  it, and the description lands under the image in the Sheets tab, badged read
  from the screenshot so it is never confused with a claim about the lineage.
  This is the one thing linexcel documents from a picture rather than from the
  graph, because colour conventions, conditional formatting, charts and the
  shape of a layout never reach a text dossier. It is opt-in, and separate from
  --ai-docs, because a picture of a sheet shows every row on it: nothing
  sends an image unless you ask. A text-only endpoint raises rather than having
  the image quietly dropped, and --vision-model names the model that looks
  when it differs from the one that writes. Providers may implement
  generate_with_image to serve the same call. Note that this is the only card
  in the report with nothing deterministic behind it: a weak vision model
  invents confidently, and the description is placed under its own image so it
  can be read against it.
• refs_dir= / --refs-dir: the workbooks this one depends on. A cell
  reading '[Budget FY26.xlsx]Annual'!B4 used to be a grey "external
  reference" node with no value, and every formula above it lost its own.
  Three answers are now given, in order: the file names the workbook it
  depends on and the path it declares (always, from xl/externalLinks); the
  values Excel cached across the link are used when it saved them; and,
  given a folder holding the referenced files, the workbook is read and the
  reference evaluates to the value it stands for — the reference is stripped of
  its path and resolved against that folder. The node panel lists each linked
  workbook with its path and says which of the three it got, and the same
  folder is searched for .xlam/.xla/.xlsm add-ins so the VBA a workbook
  calls into becomes part of the lineage, each module tagged with its file.

Changed

• The value panel always separates what the file carries from what linexcel
  recomputed. The card used to collapse to a single figure whenever the two
  could not be compared, so "linexcel agrees with Excel" and "Excel never wrote
  a value here" looked identical on screen. Both columns are now always laid
  out, and a column with nothing to show says so — Not stored, Not
  recalculated — with a sentence naming the reason (a workbook saved without a
  calculation cache carries no value for its formula cells).
• A stretched formula shows one row per sampled cell instead of a single
  headline figure followed by a plain list. Each sampled cell keeps its own
  pair of readings, a row whose two readings disagree is marked on the row, and
  the sample is now spread evenly over the group — first cell, last cell and
  three in between, where a pattern that broke actually shows — rather than
  being the first three, near-identical, neighbours.
• Volatile formulas are reported as not recalculated, on purpose. TODAY,
  NOW, RAND, RANDBETWEEN and RANDARRAY answer differently every time
  they are computed, so a recomputed =TODAY() could never agree with a file
  saved last week — and linexcel called that a divergence, blaming the workbook
  for the calendar. Such a cell now keeps the value the file stores, says Not
  recalculated (volatile) in the recalculation column with a sentence
  explaining why, is not decomposed step by step (every step would carry a
  figure from today's clock), and raises no divergence warning. OFFSET,
  INDIRECT, CELL and INFO — volatile to Excel but stable for a given
  workbook — are still computed.
• Analysis is roughly 2.5× faster on a dense workbook: 246 s → 94 s on a
  47 MB file of 2.6M formulas, with an identical graph. Two costs went away.
  The step decomposition no longer re-evaluates the root step — it is the
  whole formula, so the value the engine already computed for the cell answers
  it; re-running it made a root such as SUM(Calculs!H2:H200001) walk 200,000
  formula cells again, 29 s for one node and 114 s over the run. And table
  detection no longer opens the workbook a second time with openpyxl: declared
  tables are read from the xl/tables/*.xml parts, static ones from the 30×50
  window the engine already holds, which turns 45 s into nothing measurable.
• The per-sheet scan ceiling went from 4M cells to 64M, and the last chunk
  is clipped to what is left rather than dropped whole — a 4,000,000-cell
  budget used to stop at 3,600,000. Sweeping costs about 0.7 µs per cell, so
  the old ceiling bought 1.6 s on the reference workbook and cost that sheet
  its tail; the new one is past any real used range and still bounds the file
  that declares more than it holds (one stray cell at XFD1048576 makes the
  used range 17 billion cells). The warning now names the first row left out
  and says its formulas are missing from the lineage. Reads are chunked by
  cells rather than by rows as well: 20,000 rows of a 16,384-column sheet
  meant 327 million strings in one call.

Fixed

• A step lost the parentheses of the formula it came from, and reported the
  wrong number. Each step renders its own expression from the AST and is then
  evaluated by re-parsing that text — but the parser keeps grouping in the
  shape of the tree, so =D2*(1-Rate) came back as D2 * 1 - Rate. The step
  card showed 2470.06 under a cell plainly holding 1976.208. Parentheses are
  now restored from operator precedence, both in what is displayed and in what
  is evaluated.
• A step could report a value belonging to another cell. The batch
  evaluation reuses one scratch row across nodes, and set_formula does not
  raise on an expression the engine will not take — a structured
  Table[Column] it never loaded, a 3D 'A:B'!A1 — it leaves the cell as it
  was. The step then read back whatever the previous node had computed in that
  column. Each scratch cell is now primed with the sentinel first, so a refused
  expression reads as not evaluated, which is what it is.
• A root step over an array literal (=SUM({1,2,3;4,5,6})) read as #NAME?:
  the AST renders an array as the placeholder {...}, which evaluates to
  nothing. It now carries the value of its cell.