v0.1.1

Changed

• Linux wheels target a manylinux_2_28 (glibc 2.28) baseline instead of manylinux_2_35. The
  previous tag came from whichever glibc the build runner happened to ship, and it excluded RHEL 9
  and Amazon Linux 2023 by 0.01 of a glibc version; both fell back to the source distribution,
  which requires a Rust toolchain the installing user rarely has. The baseline is now an explicit
  build input, produced with the same pinned Zig the npm platform packages already use, and the
  resulting platform tag is asserted during release verification rather than assumed.
• Dependency requirements of the published cellrune crate are caret ranges instead of = pins.
  An exact pin made the crate unresolvable alongside any dependent needing a newer patch of the
  same dependency, with no remedy available to that dependent. Build reproducibility continues to
  come from the committed Cargo.lock. Two new scheduled gates bound the ranges: one runs the test
  suite against the declared floors on the minimum supported Rust version, and one runs it against
  the newest compatible graph. Both resolve live rather than from the lockfile, so they are
  scheduled alongside the advisory tier rather than gating pull requests: a failure there means
  upstream moved, not that the commit under test did.

Fixed

• Quoted external-workbook references such as '[1]Sheet1'!A1 were reported as supported and
  evaluated to #REF!. Because that is a spreadsheet error value rather than an engine issue,
  IFERROR could hide it, so a workbook could return a plausible number for a formula the engine
  cannot evaluate. Both the capability scan and evaluation now reject the external prefix. The
  unquoted spelling was already rejected, as the formula lexer has no bracket token. A reference
  built at run time by INDIRECT is the deliberate exception: the scan cannot read inside the
  text argument, so it reports the cell as supported, and evaluation answers #REF! there as
  Excel does rather than an engine issue the scan never predicted and no caller could recover
  from. An external prefix written as a saved path keeps its drive letter and is reported once,
  not also as a 3-D sheet range the formula does not contain.
• Calculated zeros could be negative. Iterator::sum for f64 folds from -0.0, f64::min and
  f64::max may return either operand when both compare equal, and Iterator::product inherits
  the sign of an odd number of negative factors, so SUM, MIN, MAX, PRODUCT, and the *A
  variants could each report -0 where Excel reports 0. The boundary every calculated value
  crosses now normalizes negative zero, rather than each kernel being expected to.
• Formula parse-error details labelled every position as token N, but a lexing failure has no
  token stream and was reporting a character offset under that label. Lexing failures now report
  character N and parsing failures continue to report token N.
• The lambda function group dispatched without inspecting the normalized function name, unlike
  every sibling group. A second function in that group would have silently evaluated as MAP.
• A release version bump had to be repeated by hand in more places than the release checklist
  listed, and the ones it missed — a pnpm lockfile, a generated loader shim, the packaged-consumer
  lockfile, and the release verification scripts — fail only during a release run. Every version
  is now derived from the workspace manifest where possible, and a new gate asserts that the
  remaining declarations agree, including the install commands and the supported-version statement
  that ship on the registry project pages. Every check in that gate fails when its pattern matches
  nothing, so a moved file or a regenerated shim cannot report the same green result as a correct
  bump, and the gate has its own tests.
• A Linux wheel could declare a compressed platform-tag set whose first component was a legacy
  alias, which hid a newer glibc requirement behind it from the release verification. Every
  component of the set is now checked against the supported baseline.
• The nightly fuzz tier could not install its fuzzer. taiki-e/install-action has no cargo-fuzz
  manifest, the job deliberately disables the binstall fallback, and the schedule had never fired
  since the repository went public, so the defect surfaced only on the tier's first manual
  dispatch. cargo-fuzz is now built from crates.io with cargo install --locked, which keeps the
  exact-version pin and the registry checksum verification.
• The npm publish job could not authenticate. Trusted publishing requires npm 11.5.1 to exchange
  the job's OIDC identity for a registry credential, and Node 22 bundles npm 10, which publishes
  without credentials and receives the rejection masked as E404. The nine-package dry-run gate
  cannot catch this because a dry run never authenticates, so the failure surfaced on the first
  real upload — before any package was published, leaving the retry clean. The job now upgrades
  to a pinned npm 11 before publishing.

Documentation

• The 0.1.0 entry claimed capability detection for unsupported external, structured, and
  spill-postfix formula forms. Only 3-D and data-table forms have dedicated detection; the other
  three surface as calculation.parse_error, and structured references and spill-postfix
  references still do after this release. The entry has been corrected.
• Added docs/NUMERICS.md, which records where calculated values deliberately differ from Excel,
  the Excel build each statement was measured against, and which function families are unmeasured.
  It is linked absolutely: both published READMEs are also the long description crates.io, PyPI,
  and npm render, none of which resolve a repository-relative path, and a gate now enforces that.
• Installation instructions no longer describe the registry artifacts as pending.