v0.13.0

XLSX reading moves in-house: Rust, Python and JavaScript now use Formualizer's native reader by default, without a second production parser underneath it. Ordinary loading reads the workbook data it needs rather than decompressing every picture and attachment along the way. Source-preserving recalculation keeps its stricter whole-package integrity contract, and Umya remains available for rich document editing and export.

Breaking changes and upgrading

• Retired the Calamine parser and dependency. NativeXlsxAdapter replaces its implementation. The Rust CalamineAdapter name, backends::calamine module path and calamine / calamine_integration feature names remain native compatibility aliases, not an alternate parser or fallback. Python backend="calamine" and JavaScript { backend: 'calamine' } also select native. Most binding users can simply omit backend.
• Changed Calamine-specific Rust error types. The adapter's reader/stream associated error is now IoError. Removed IoError::Calamine(calamine::Error) and From<calamine::Error>; match IoError::XlsxRead(XlsxReadError) instead. Removed the adapter-only external_link_target accessor and test-only relocation-comparator injection hook. See the native XLSX migration guide for replacements and retained interfaces.
• Changed ordinary loading from whole-package payload validation to selective consumption. A corrupt or very large unused image, embedded object or other unread payload no longer prevents value/formula loading. Required parts still undergo CRC, decoded-size and actual-output-limit checks. Ordinary loading is not a certificate of every ZIP member's integrity; strict source-preserving recalc retains its full-package checks and refusal policy.
• Preserved supported rich-document capabilities. Umya reading, mutable workbook operations and rich writing/export remain available where previously supported. Removing Calamine does not remove all spreadsheet dependencies. The former adapter used Calamine's XLSX reader, not its XLS implementation, so this does not withdraw an existing .xls integration.

Highlights

• Added explicit broader template recognition in parser SDK 4.1.0. template_lex::lex_template_for_certified_ast recognises invariant sheet qualifiers, names, whitespace, @ and general functions for callers that independently certify the parsed AST. It is not proof of purity or non-volatility. The existing lex_template retains its published 4.0 conservative subset; native source-family ingestion uses the new opt-in path without narrowing its existing recognition.
• Selected the native XLSX reader by default across Rust and the Python/JavaScript bindings, including the Python API used by Pyodide. It preserves loaded values/formulas, names and calculation settings, date/time behavior, cached external references and inert-sheet diagnostics. Cached links remain read-only: linked files are never refreshed or opened.
• Added configurable ordinary XLSX limits before admission. Rust exposes XlsxLoadLimits and additive native constructors; Python accepts keyword-only load_limits, including through the SheetPort path helper; JavaScript accepts nested loadLimits. Native and legacy backend aliases use the same options and checks.
• Raised the ordinary compressed-input default to 127 MiB (133,169,152 bytes). Expanded data has independent defaults: 128 MiB per consumed part and 256 MiB aggregate actual output. Callers can raise or lower these separately; a compressed-size allowance is not a guarantee that the expanded workbook, formulas or engine state fit in memory.

Loading and safety

• Skipped decompression of unread payloads, including shadowed ZIP entries, while retaining the existing Compatible ZIP structure/index checks and consistent selected-member resolution. Images and attachments remain in the retained source; they are not expanded, decoded, executed or removed by a value/formula load.
• Bounded actual consumed output throughout a load, including lazy external-cache reads and real metadata rereads. Byte counts are based on emitted data, not trusted ZIP size claims. Existing worksheet buffers move to their consumers without a new retained-part cache. Source and part buffers grow geometrically within their allowance rather than requesting a new exact allocation for every chunk.
• Kept required-part failures fatal. CRC errors, declared/actual decoded-size mismatches, resource failures and cancellation cannot disappear into an unavailable external cache or a successful partial workbook. Existing best-effort unsupported-link/name semantics remain separate. Terminal safety failures also prevent subsequent sheet reads and bounds lookups through the same adapter.
• Validated nested options strictly. Unknown keys, booleans, negative/fractional/non-finite values and host-unrepresentable numbers are refused; JavaScript also rejects unsafe integers. Zero is a zero allowance, not unlimited. Explicit XLSX-only overrides are rejected on unsupported loading routes instead of being silently ignored.
• Preserved independent limits and host constraints. XML/cell/grid protections and engine geometry/formula-spool budgets still apply. Cancellation is cooperative; a single XML token is not a hard real-time cancellation boundary. Browser/WASM memory and address-space limits remain relevant even when a caller raises byte allowances.

Performance

• Reduced recalculation ingestion overhead by removing redundant formula payload staging/publication work, streamlining worksheet scanning, and loading native values/formulas without another general-reader pass. These changes build on the recalculation work in 0.12.0. (#537–#539)
• Removed decompression and CRC CPU work for unused ordinary-load parts. The benefit depends on workbook contents; attachment-heavy examples are not representative of every formula/data-heavy workbook. Opaque payloads were already streamed through bounded scratch memory, so skipped expanded bytes must not be described as an equal reduction in peak RSS.
• Kept benchmark comparisons distinct. Reader-transition measurements against an already optimized Calamine control are neither an isolated default-switch speedup nor a total 0.12.0-to-0.13.0 improvement. Stage percentages are not additive. No universal speedup, wheel/WASM-size reduction or peak-memory saving is promised here.

Compatibility and validation

• Added boundary coverage for selected versus unread corruption, Stored/Deflated decoded-size mismatches, exact/overflow output budgets, duplicate and Unicode member names, lazy-cache retries, retained source handles and current-position readers. Required-part checks do not depend on auditing unread content first.
• Expanded binding coverage for configuration validation, native aliases, Python forwarding/stubs and actual WASM/TypeScript loading. Native-versus-legacy-alias tests check API compatibility, not independent parser equivalence. Pre-retirement differential evidence used the real Calamine implementation; Pyodide wheel/SDK execution remains a separate CI gate.
• Retained attribution for the small shared-formula/error routines adapted from Calamine under its MIT license.

Maintenance and versions

• Patched documentation-site and comparative-harness dependencies, including Next.js, PostCSS and transitive dependencies with known security advisories. The documentation site remains a static export. These updates do not change the spreadsheet evaluator's numerical behavior.
• The Rust product crates, native CLI, Python and npm packages are 0.13.0. formualizer-common and formualizer-parse move together to 4.1.0; publish that SDK track before the product track. sheetport-spec remains 0.3.3.