linexcel 1.9.0

[1.9.0] — 2026-09-13

Added

• Analysis resource ceilings can be configured through both public APIs.
  Validated limits are recorded in metadata and propagated to workbook loading,
  graph construction, recovery and external workbook reads. A targeted closure
  that exceeds its cell budget is rejected rather than silently truncated.

• AI dossiers retain source name definitions and worksheet scope, including
  constants absent from reference-only graph nodes. Group values identify their
  representative cell, and literal AGGREGATE selectors have deterministic
  meanings. Oversized or locally scoped evidence is explicitly qualified.

• Small-screen graph options have a dedicated keyboard-accessible dialog.
  Views stay visible while filters and layout remain reachable without covering
  nodes with a permanent legend.

• Graph navigation uses a dedicated responsive toolbar with a labeled sheet
  selector, grouped zoom controls and a live zoom percentage, and explicit
  framing actions. Search has a visible submit action and describes its scope.

• Local acceptance runs include both fixtures, screenshots and image analysis
  with Ollama by default. Each run keeps separate artifacts and a manifest
  with source/input/image hashes and per-language coverage. Partial runs and
  failed stages return a nonzero status. Contributor instructions require
  inspecting dashboard tabs, sampled nodes and AI claims against screenshots.

• CI and documentation checks cover stacked PRs targeting codex/** branches,
  so dependent improvements are validated before their base PR reaches main.

• The node panel can frame a node with its direct neighbors. The action
  remains accessible on mobile, where the panel leaves the graph controls
  visible. Search returns to the graph from the other dashboard tabs.

• --target SHEET!A1 limits the analysis to the cells feeding the named
  cell. Repeatable, and several comma-separated cells fit in one flag; the
  library spelling is analyze(..., targets=["Sheet1!A1"]). The engine boots
  without the global evaluate_all, the upstream subgraph of the targets is
  traced (Workbook.trace) and evaluated (evaluate_cells), and its lineage
  is built with relevant defined names, VBA and Power Query context. Dynamic
  references (INDIRECT/OFFSET) or a truncated trace can cause dependencies
  to be evaluated without appearing in the graph; the report states this
  limitation. Without --target the whole workbook is evaluated, as before.

• A warning now flags workbooks of long dependency chains. On a targeted
  run it is exact — the engine's evaluation plan (get_eval_plan) counts the
  layers of the subgraph before it is evaluated. On a full run the plan reads
  zero once everything is computed, so the measure is taken from the lineage
  itself: the longest precedent→dependent path across formula nodes, and the
  largest self-referencing group (a running-total column is one node on the
  graph but a chain as deep as it is long). Either way it is a risk
  indicator, not a duration estimate.

• The viewer uses readable card nodes and zoom-aware labels, offers a
  differences-only filter, and initially uses flow layout for larger graphs.
  The layout choice remains available in the toolbar.

• Browser regression tests run in CI, exercising filters, search, dense
  selections, and desktop/mobile controls in light and dark themes.

Fixed

• Cache ceilings also apply to error overlays and styled empty columns.
  Recovery that runs out of depth does not publish an incomplete scratch result.

• Scratch markers are unique per operation, avoiding collisions with valid
  formula results that happen to equal a former fixed marker.

• Optional documentation parsing rejects structurally complex formulas before
  entering the native AST conversion, and local LET/LAMBDA bindings are not
  mistaken for workbook name references.

• Visual preview is hidden when its content is already available in Sheets.
  Standalone image pages and descriptions without sheet context remain reachable.

• Search feedback is cleared when the query or graph filters change.
  Searches with no matches also clear the previous selection.

• Inline code in AI documentation preserves formula operators. Markdown
  emphasis no longer consumes multiplication signs inside code spans.

• Empty or truncated AI responses no longer become successful documents.
  Failed calls retain their reported token usage; oversized formulas are
  explicitly omitted from bounded dossiers rather than passed as partial proof.

• Legacy NORMSDIST and NORMDIST formulas evaluate through their native
  modern equivalents, preserving original formulas, coercion and error values.

• Local LET and LAMBDA steps are not evaluated outside their scope.
  The complete calculation retains its value; unavailable inner steps state
  that local scope prevents independent evaluation.

• Screenshot descriptions retain their image identity after HTML embedding.
  Pages without sheet context remain accessible, chartsheets preserve page
  ordering, and images can be inspected at native resolution in the viewer.

• Narrow documentation tables preserve whole words and scroll horizontally.
  Group cards show their member count once, keeping the underlying label intact.

• 1904 workbooks use their calendar during engine import. Formatted date
  inputs, date functions and targeted evaluation no longer inherit a 1,462-day
  shift from the engine's default 1900 calendar.

• Circular calculations honor workbook iteration settings. Non-converged
  components and their dependents retain file caches and omit unverifiable
  breakdowns; independent formulas remain calculable. Convergence is qualified
  by the engine's criteria and does not imply a unique fixed point.

• Self-closing XML cells cannot capture a following cell's formula. The
  quarantine scan no longer backtracks quadratically over styled empty cells.

• AI dossiers preserve engine identity, value provenance and uncertainty. Cached or volatile
  snapshots are not labeled recalculated, disagreements and omitted neighbors
  are explicit, and useful calculation steps are retained before extra neighbor
  examples. Prompts in all nine languages reject unsupported explanations of
  discrepancies and extrapolations across sparse groups. Vision descriptions
  and overviews are labeled for checking against the original evidence.

• Viewer Markdown handles nested lists and inline formatting in headings.
  Spreadsheet errors and unevaluated steps no longer use the successful-result
  style. Sheet context displays extraction warnings and unknown dimensions.

• Excel time caches compare correctly with numeric day fractions. Midnight
  (0) and noon (0.5) agree with stored time values without converting plain
  text, booleans, multi-day durations or timezone-bearing values into times.
  The graph retains the numeric reading and carries its ISO time in valueDate.

• Quarantined formulas keep honest value provenance. A stored result is
  attributed to the file, and a missing result is not treated as zero. When a
  deep formula has no cache, dependent calculations and their breakdowns remain
  uncomputed, including IFERROR/ISERROR and dynamic references. Independent
  formulas still evaluate when their independence can be established.

• An engine rebuild preserves the deep-formula quarantine instead of
  restoring formulas that could overflow the evaluator's stack.

• Targeted formula scans read only traced cells. Sparse targets no longer
  consume the sheet scan limit on intervening empty rows or columns, and
  unrelated defined names no longer trigger out-of-scope recalculation.

• Long chains across distinct formula groups are detected. The previous
  traversal skipped every root; the replacement visits each node and edge once
  and handles cycles without recursion.

• Targets support quoted sheet names containing commas and escaped quotes.
  Malformed or empty target entries fail with an actionable error.

• Boolean and mixed-type disagreements are no longer reported as agreement.
  Numeric tolerance, equivalent date representations and Excel error values
  retain their comparison semantics; recalculated errors over stored numbers
  are differences.

• Sheet, type and difference filters compose correctly. Searching only
  matches visible nodes, so hidden results are not selected or centered.

• Group comparison keeps the representative cell's verdict separate from
  the worst sampled verdict. The difference filter includes a group when a
  sampled cell differs without falsely labeling an agreeing representative.

• Disconnected graphs use a compact grid in flow mode, avoiding a tall
  single rank of nearly invisible cards when hundreds of nodes have no links.

Performance

• Large sheet previews use streaming reads. Above 20 MiB of worksheet XML,
  report context avoids materializing the entire openpyxl worksheet. Omitted
  comments and layout metadata are explicitly reported; previews still work
  when worksheet dimensions are absent.
• Batch documentation indexes graph adjacency once, instead of scanning
  every node and edge again for each requested card.
• Formula-free XML sheets skip the quarantine scan. CLI size estimates
  describe worksheet XML rather than incorrectly calling it formula volume.
• Selecting many search results no longer scans the whole graph per result.
  Selection updates are coalesced into one animation frame, and the selection
  collection is assembled in one pass.

Changed

• formualizer 0.9.3 is the minimum supported engine. Parallel evaluation is
  configured explicitly and can be disabled through boot_engine(parallel=False).
• Chartsheets are skipped with a warning so their presence does not prevent
  worksheet analysis. Over-deep formulas are quarantined before evaluation to
  avoid an evaluator stack overflow.
• A failed quarantine retry no longer pays a third from_bytes. The
  rebuild existed because a failed evaluate_all was believed to leave the
  engine reporting no formula at all; on formualizer 0.9.3 the formula map
  survives the failure, so the engine the run already holds is kept (with the
  quarantined cells re-injected by the sweep) and the rebuild is paid only
  when a probe of one formula cell finds the map really is gone.