v0.1.2

Added

• A single data-driven conformance expectation suite. conformance/ holds redistributable
  matrices — every literal, every formula, and the value a recorded oracle saved for it — and
  the normal workspace test run reconstructs every matrix and compares CellRune's calculation
  against the oracle. New functions and syntax extend the same data set.
  The first matrix is the Apache POI FormulaEvalTestData corpus (Apache-2.0): 1,295 cases
  against a saved Microsoft Excel 2013 calculation cache, 1,290 matching within a
  scale-relative 1e-8 and 5 divergences documented case by case. A documented divergence is
  enforced in both directions, so neither the oracle side nor the CellRune side of it can drift
  silently, and a divergence without an explanatory note fails the test. An extraction tool
  (extract_conformance_matrix) accepts ordinary and resolved shared formulas and rejects
  workbook metadata the v1 matrix cannot reproduce.
• A semver invariant inside the existing test suite (tests/public_api.rs). The sixteen exported
  enums that are not non_exhaustive are matched without wildcard arms with every variant
  payload bound at its exact type, and function-pointer constants pin the positional signatures
  of WorkbookSnapshot::new_with_metadata and the monomorphic read → calculate → write entry
  points. This remains one test as the API grows.

Changed

• Rust dependency requirements now declare the oldest API-compatible, advisory-clean versions
  instead of using the release that happened to be newest when each manifest was edited. All
  workspace requirements live in the root manifest, the lockfile still selects the reproducible
  current graph, and every release family a broad range admits is exercised by the floor or the
  newest-compatible suite. Fifteen of the twenty external workspace floors moved down; five
  remain at their current releases because they are the first secure version or belong to the
  same release family (quick-xml, PyO3 and its build helper, and calamine), or because their
  types appear in CellRune's public signatures and a public dependency must stay within one
  semver family (cap-std). The fuzz crate's libfuzzer-sys stays exact-pinned as a tool
  input: its graph is excluded from the workspace and always built --locked, so no suite
  could exercise a range there.
• The conformance/ fixtures now live in binding-contract/, which is what they are: a
  cross-binding API behavior contract exercised by the Rust interop, Python, and Node test
  suites. Two 1 KB fixtures cannot carry an Excel conformance claim, so the conformance/ name
  is reserved for the Excel expectation suite that can.
• The nightly sanitizer campaign remains one job over four stable trust boundaries. It records
  every target's result before failing, so one crash cannot starve later campaigns and new
  functions do not create new fuzz jobs.
• The MSRV check covers --all-targets rather than --lib, so tests, benches, and the
  extraction tool must also build on the declared floor.
• The 50k benchmark is demoted from a scheduled nightly job and a required release gate to an
  explicit observational command, documented in CONTRIBUTING.md, and its output-verifier
  script is removed with the jobs. Release gates are deliberately limited to stable validation
  boundaries that guard irreversible failures; the benchmark keeps its internal correctness
  assertions, and functional correctness remains covered by the generated-workbook tests.
• Release publish jobs are re-run safe. The crates.io, npm, and GitHub-release jobs probe for
  what is already live and skip it — the GitHub release also counts its attached bundles, so an
  interrupted asset upload is completed rather than skipped — and the PyPI upload passes
  skip-existing. A rerun of a partially published tag therefore converges on fully published
  instead of ending red: the final runs of both prior releases stayed red because their
  already-published registries could only fail again or be rejected by hand.

Fixed

• A workbook containing an empty sheet was rejected with xlsx.invalid_worksheet. Excel and
  every mainstream producer serialize an empty sheet as a self-closing <sheetData/>, which
  the reader's required-sheetData rule did not recognize as the element's empty form, so even a
  brand-new workbook saved by Excel failed to open. Empty sheets now read as sheets with no
  cells; a duplicated <sheetData/> is still rejected.

Documentation

• The README carries a Verification section: the public conformance-suite numbers with their
  exact oracle, and the private-audit results (Excel for Mac 16.111 corpus recalculation,
  14-workbook full-span audits, the 250k-formula timing, and the 4-of-4 producer matrix) with
  their provenance and non-distribution stated.