@particle-academy/holy-sheet 2.5.0 → 2.7.0

package/CHANGELOG.md

@@ -0,0 +1,560 @@
+# Changelog
+
+All notable changes to `@particle-academy/holy-sheet` are documented here.
+(The repository is `holy-sheet-js`; the published package has never carried the
+`-js` suffix, and this line used to say it did — a name a consumer would have
+typed into `npm install` and got a 404 for.)
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [2.7.0] — 2026-10-19
+
+### Added
+
+- **`print`** (`orientation`, `fitToWidth`, `footer` with `{page}` / `{pages}`)
+  and **`rowHeights`** (pixels by 1-based row number). Read back, validated and
+  published in the schema. **Nothing to do**: a workbook without them is
+  byte-identical. Same as `particle-academy/holy-sheet` 2.6.0 and
+  `fancy-holy-sheet` 0.6.0; parity holds (`printSetup` fixture).
+
+## [2.6.0] — 2026-10-11
+
+### Fixed
+
+- **A cell value that cannot be represented is REFUSED and names the cell,
+  instead of being silently stringified. In this runtime the old behaviour wrote
+  `"x"` for the cell `["x"]` — the corruption looked correct.** Mirrors the PHP
+  reference fix in `particle-academy/holy-sheet` 2.5.0 ([holy-sheet#9], reported
+  from MOIC production); tracked here as [#2].
+
+  `coerceValue()` ended with "not a primitive, so it must be safe to stringify":
+
+  ```ts
+  if (typeof v !== "string" && typeof v !== "number" && typeof v !== "boolean") {
+    return String(v);   // an ARRAY reaches this line
+  }
+  ```
+
+  An array gets there whenever a plain `rows` form is nested one level too
+  deep — `rows: [[["x"]]]` instead of `rows: [["x"]]`. No object-cell syntax is
+  needed, which is exactly the shape an AI-generated edit produces.
+
+  **Why this port's version was worse than PHP's.** `Array.prototype.toString`
+  joins, so a one-level nesting produced *exactly what the author meant*:
+
+  | cell | before | PHP before its fix |
+  |---|---|---|
+  | `["x"]` | `"x"` | `"Array"` |
+  | `["x","y"]` | `"x,y"` | `"Array"` |
+
+  PHP wrote visible garbage, which someone files a bug about. This wrote a
+  workbook that was silently, undetectably correct-looking — and for a
+  two-element row it merged two cells into one with the data intact enough to
+  look deliberate. A wrong answer that looks right is the expensive kind.
+
+  The guard now accepts an object that defines its own `toString` (PHP's
+  `Stringable`) and refuses everything else, with **two exclusions that are both
+  load-bearing**: a function, because `Function.prototype.toString` returns the
+  function's SOURCE TEXT — the old code wrote a code listing into the cell — and
+  an array, because `Array.prototype.toString` is *also* "its own", so "has a
+  custom toString" is true of the one value that must never pass. A symbol is
+  refused too, where it previously wrote `"Symbol(x)"`.
+
+  The refusal is a `SchemaException` whose structured error names the cell as
+  `Sheet!A1`, with a hint explaining the nesting mistake. That differs from the
+  `sheets[0].rows[1]` schema path used elsewhere, deliberately: this fault is
+  always one cell, and the A1 location is what an operator can act on.
+
+- **A `Date`, or any object with a `toString`, in a plain `rows` row no longer
+  writes an EMPTY CELL.** Found while fixing the above, and it is the more
+  ordinary bug of the two.
+
+  The cell-spec dispatcher tested `isPlainObject(value)` — `typeof === "object"
+  && !null && !Array.isArray` — which is true for **every** non-array object. So
+  a `Date`, a class instance and a `Map` were all read as the `{value, format}`
+  cell-spec form, `value.value` was `undefined`, and the cell came out blank: no
+  error, no validation failure, just a hole where the date should be. PHP's
+  equivalent guard is `is_array && !array_is_list`, and a PHP
+  `DateTimeInterface` is an object rather than an array, so it fell straight
+  through to the date branch and wrote the serial. **The pair's contract is
+  "same bytes either backend", and the byte-parity suite never saw this because
+  its fixtures carry ISO date *strings* rather than `Date` objects.**
+
+  The dispatcher now claims only a plain object literal — which is also what
+  `JSON.parse` produces for every schema arriving as JSON — and anything with
+  its own prototype is treated as a value.
+
+- **`validate()` no longer passes a schema that `normalize()` will throw on.**
+  The same input returned **zero** validation errors, so `Agent.validate()` told
+  a caller the workbook was fine and the write then failed — moving the failure
+  to where it is hardest to attribute, in the exact validate-then-write order an
+  agent uses. `Validator` now reports `sheets[0].rows[0][0]` for an array cell.
+
+  The check is `Array.isArray`, **not** "is an object": the object cell form
+  `{value, format}` has always been legal, and rejecting any object would have
+  traded this bug for a far worse one — every formatted cell in every workbook.
+  There is a test holding that line.
+
+  **What you must do:** nothing, if your cells were already scalars, dates or
+  cell objects. If any schema you build can nest a row one level too deep, you
+  now get a `SchemaException` naming the cell where you previously got a
+  workbook with the joined value in it — read `getErrors()[0].path` for the
+  address. Released as a MINOR rather than a patch because that throw is
+  observable to a caller whose input the validator previously accepted.
+
+- **`CHANGELOG.md` is now in the published tarball.** `files` did not whitelist it, so npm never shipped it: a consumer who followed a link to the changelog — from the README, from npm, or from an upgrade guide — found nothing. Nothing for you to do; the file simply arrives from this release on.
+- Dropped `docs` from `files`: it was declared and never existed, so the package claimed to ship it and did not.
+
+### Security
+
+- `source-map-js` is pinned forward to `^1.2.2` via `overrides`. Versions up to
+  1.2.1 allow an event-loop denial of service through indexed source-map section
+  offsets, and it arrives here transitively through the build toolchain.
+  **Nothing for a consumer to do, and no runtime change**: an npm package does
+  not ship a lockfile, so this governs builds OF this repo, not anything
+  installed FROM it. Recorded rather than left silent because the override it
+  sits beside — `shell-quote` `^1.9.0`, added for an earlier advisory — was
+  carried with no note of why, and had drifted back inside the vulnerable range
+  before anyone looked.
+
+[#2]: https://github.com/Particle-Academy/holy-sheet-js/issues/2
+[holy-sheet#9]: https://github.com/Particle-Academy/holy-sheet/issues/9
+
+## [2.5.0] — 2026-10-02
+
+### Fixed
+
+- **A column's own `width` is honoured. It was documented and silently ignored.**
+
+  `skills/holy-sheet.schema.json` — the tool definition handed to an LLM, shared
+  byte-identically with the PHP and Python ports under a checksum test —
+  describes `columns[].width` as *"Column width in pixels. Same as columnWidths
+  but per-column."* Only the sheet-level `columnWidths` map was ever read, so a
+  width written the way the schema documents it emitted **no `<cols>` element at
+  all**: no exception, no validation error, and nothing from `validateAndRepair`.
+  Silent at every layer.
+
+  Reported against the PHP package as **holy-sheet#8** by the MOIC team, whose
+  owner's complaint was "can't style spreadsheets at all". Measured, that was
+  mostly this: the theme, the header fill, the banded rows and the currency
+  formats all landed, but a real account name truncated and a correctly-formatted
+  currency rendered as `#####` — which is what a reader actually sees. Fixed in
+  PHP 2.4.0; this is the matching twin.
+
+  **Precedence, identical to PHP: the sheet-level `columnWidths` map is applied
+  LAST and wins.** It is the mechanism that already worked, so a consumer who
+  moved to it to route around this bug must not then find a leftover `width`
+  quietly overriding them. The two sources merge rather than replace, and `<col>`
+  elements are emitted in ascending column order — merging two sources means
+  insertion order is no longer column order.
+
+  A `width` now goes through the same `ColumnWidths` rule the map already obeyed,
+  so `"abc"` or `-5` is rejected rather than written, and **`validate()` reports
+  it at `sheets[0].columns[0].width`**. That part matters as much as the fix: the
+  original defect's real cost was that an agent composes against the schema doc,
+  gets no error, and never learns the field was dropped.
+
+  **What you must do: nothing, and do not migrate off `columnWidths`.** If you
+  set `width` and saw no effect, it now works. If you moved to `columnWidths`, it
+  still wins. Output bytes change for anyone who had `width` set — which is why
+  this is a minor rather than a patch.
+
+
+## [2.4.3] — 2026-09-14
+
+### Fixed
+
+- **A `columnWidths` key that is not a column index wrote a `NaN` column.** The
+  normalizer read keys with `parseInt`, so `{"abc": 999}` became a width for
+  column `NaN` in the written file. PHP overwrote column A with it and Python
+  raised; all three follow one rule now, mirroring PHP holy-sheet 2.3.4
+  (`src/schema/column-widths.ts`):
+  - a **key** is a 0-based column index from 0 to 16383, as a string of digits;
+  - a **width** is a non-negative finite number, or a string of digits (`"80.5"`).
+  - `validate()` reports each other entry by path (`sheets[0].columnWidths.abc`),
+    so `write()` and `toBytes()` refuse it.
+  - `validateAndRepair()` turns a one- or two-letter key into its index (`"B"` is
+    1) and drops an entry it cannot repair, and lists both.
+  - The normalizer skips such an entry.
+
+  **What you must do:** nothing, unless a schema carried a width keyed by a letter
+  or by junk. `write()` now refuses it; run it through `validateAndRepair()`, or
+  key widths by index.
+
+## [2.4.2] — 2026-09-14
+
+### Fixed
+
+Six defects in the op code, mirroring `particle-academy/holy-sheet` 2.3.2 (five,
+found by the Python port) and 2.3.3 (one), which are the reference. Each has a
+test ported from PHP's `SheetOpsTest.php`, and each fails against 2.4.1 except
+the second, which this port never had.
+
+- **`Agent.opSchema()` rejected a `set_column_widths` op carrying a list.** 2.4.1
+  allowed only an empty array, but PHP encodes widths keyed 0..n-1 as a list
+  (`[120, 80, 140]`), and a schema read from PHP's JSON carries that list through
+  this port's `diff()` into its op. `columnWidths` is now an object or an array
+  of non-negative numbers, a list indexed by position. The schema matches PHP
+  2.3.2's byte for byte, description included.
+- **A `type` that is not an op type is skipped before anything else.** PHP's
+  loose `switch` let `type: true` remove a sheet. This port's `switch` was
+  always strict, so it never had that defect; `reduce()` now checks the type
+  against `SheetOpSchema.TYPES` first, as PHP does.
+- **A padded address wrote a key of its own.** `set_cell` stored `" a1 "` under
+  `" A1 "`, and `clear_cell` could not reach A1 with it. Both now trim with PHP's
+  `trim()` set (space, tab, LF, CR, NUL, vertical tab; not NBSP) before
+  upper-casing.
+- **A column-width key that is not a column index was read as column A** by
+  `insert_columns` and `delete_columns` (`"abc"` became column 0), and could
+  overwrite column A's width. It is dropped. A digit string (`"007"`) and a
+  negative integer key are still indexes, as they are in PHP.
+- **Two values JSON cannot hold compared as the same.** `SheetDiff.same()`, and
+  so `diff()`, wrote `NaN`, `Infinity`, functions and symbols as `null`, so a
+  cell going from `NaN` to `Infinity` recorded no change. It now throws a
+  `TypeError` for a non-finite number, a lone UTF-16 surrogate in a string or a
+  key (a JS string cannot hold invalid UTF-8, and this is its equivalent), a
+  function or a symbol. Nesting deeper than PHP's `json_encode` depth of 4096
+  throws too, at the same level PHP does (every object and array counts, an
+  empty one included). The comparison is now a loop: the recursion it replaces
+  overflowed the call stack near that depth, at a level that varied from run to
+  run.
+- **An op with a position or count that is not a number moved or unfroze
+  things** (PHP 2.3.3). `add_sheet.index`, `move_sheet.toIndex`,
+  `set_frozen.rows`/`cols` and the row and column ops' `at`/`count` were cast as
+  PHP's `(int)` does, so `toIndex: "last"` moved a sheet to the front,
+  `index: "end"` inserted one there, and `rows: "one"` unfroze the panes. A
+  present value that is not an integer or a string of digits, `null` included,
+  now skips the op; absent keys keep their defaults (`add_sheet` appends,
+  `move_sheet` stays, `set_frozen` uses 0). As in PHP, the check covers every op
+  that reaches a sheet, so a junk `count` on a `set_cell` skips it too. One
+  divergence cannot be removed: `JSON.parse` gives `2` for `2.0`, so a position
+  written `2.0` applies here, where PHP decodes a float and skips the op.
+
+`tests/sheet-ops-parity.test.ts` gains PHP-checked reduce cases for each fix,
+diff cases for a list of widths, a padded cell key and a width key that is not
+an index, and `SheetDiff.same` at 4095 to 4097 levels, which the PHP helper
+builds itself so the nesting never crosses JSON.
+
+**What you must do:** nothing, unless you relied on one of the above. `diff()`
+output changes only for a schema holding a padded cell key or a width key that
+is not an index, where it matches PHP 2.3.3's, and `diff()` now throws on a
+schema holding a value JSON cannot, which JSON input never does.
+
+## [2.4.1] — 2026-09-14
+
+### Fixed
+
+- **`Agent.opSchema()` accepts a `set_column_widths` op with an empty array**, as
+  PHP 2.3.1 now declares. PHP's diff emits `columnWidths: []` when every width is
+  removed (its JSON for an empty map), which the schema rejected; an op stream
+  written by the PHP package failed validation here. `columnWidths` may be an
+  object or an empty array. The parity suite compares the schema with PHP's
+  byte for byte, so the two stay one schema.
+
+  **What you must do:** nothing.
+
+## [2.4.0] — 2026-09-14
+
+### Added
+
+- **`Agent.diff()`, `Agent.reduce()`, `Agent.opSchema()` and `Agent.equivalent()`:
+  a workbook's versions stored as ops** (holy-sheet [#7](https://github.com/Particle-Academy/holy-sheet/issues/7)).
+  Ported from `particle-academy/holy-sheet` 2.3.0, which is the reference: the
+  same algorithm, so the same two schemas give the same ops in the same order
+  in both runtimes, and an op history written by one replays in the other.
+  Hashing xlsx bytes cannot keep a one-cell edit small, because a zip changes
+  nearly every byte, so a version history had to store a whole file per edit.
+  `diff(newer, older)` is the op list that restores `older` from `newer`.
+  - `reduce(a, diff(a, b))` equals `b`, key order aside. The ops are verified
+    by replaying them: a sheet the granular ops cannot reproduce is replaced
+    whole, and so, as a last resort, is the workbook.
+  - One changed cell is one `set_cell`. Rows and columns are aligned by content
+    first, so an inserted row is one `insert_rows` plus its cells rather than
+    every cell below it rewritten.
+  - Schemas that write the same workbook diff to `[]`, so a save without a change
+    records nothing. A columns/rows sheet and the cells it becomes are the same,
+    and so is the creation time the writer stamps on a schema that names none.
+  - `set_cell`, `set_range` and `set_workbook` are fancy-sheets' `SheetOp` shapes
+    and behave as its reducer does (a `set_cell` without a formula clears it and
+    keeps the format). The rest are holy-sheet's: `clear_cell`,
+    `insert_rows`/`delete_rows`, `insert_columns`/`delete_columns`,
+    `add_sheet`/`remove_sheet`/`rename_sheet`/`move_sheet`/`replace_sheet`,
+    `set_merged_regions`, `set_column_widths`, `set_frozen` and `set_meta`.
+  - Row and column ops move cells, merged regions and column widths. They do not
+    rewrite formula text; a formula that changes with an insert is its own
+    `set_cell`.
+- **`SheetOp`**, a discriminated union on `type` (and one interface per op),
+  plus **`SheetDiff`**, **`SheetReducer`** and **`SheetOpSchema`**, exported
+  beside the other building blocks.
+
+`tests/sheet-ops-parity.test.ts` runs the PHP package's `Agent::diff` and
+`Agent::reduce` on every case and requires the same ops, field for field and in
+order: each edit PHP's own suite pins, both ways; seeded random cell edits;
+seeded row and column inserts and deletes; moved rows and raw alignments, where
+the delete-first tie-break decides the result; and the op JSON Schema, byte for
+byte. Two differences cannot be removed, because JS values do not carry the
+distinction:
+
+- PHP counts `1` and `1.0` as different values, so its diff rewrites a `1` as
+  `1.0` with a `set_cell` where this one records nothing.
+- An emptied column-width map is emitted as `{}`, where PHP's JSON writes its
+  empty array as `[]`. Both reducers read either as empty.
+
+**What you must do:** nothing. This only adds methods.
+
+## [2.3.1] — 2026-09-14
+
+### Fixed
+
+- **`Agent.lint()` accepts quoted sheet names** (holy-sheet [#6](https://github.com/Particle-Academy/holy-sheet/issues/6)).
+  `=SUM('My Earnings Projection'!A2:A3)` linted as `#NAME?` because the
+  tokenizer had no case for `'`, so any cross-sheet formula pointing at a sheet
+  whose name contains a space failed. Excel's doubled-quote escape works too:
+  `'Q3 ''Final'''!B2` is the sheet `Q3 'Final'`.
+- **A reference to a sheet that does not exist is `#REF!`**, quoted or not, and
+  the hint names the sheet and lists the ones that exist. It used to lint clean,
+  because the missing sheet's cells read as blanks.
+- **Sheet names match case-insensitively**, as in Excel.
+
+The same fix as `particle-academy/holy-sheet` 2.2.1, with the hint text
+identical in all three runtimes and pinned by a test in each.
+
+**What you must do:** nothing, unless you relied on a formula that names a
+missing sheet linting clean. It now reports `#REF!`.
+
+## [2.3.0] — 2026-09-13
+
+### Added
+
+- **`Agent.read()` and `Agent.describe()` read OpenDocument spreadsheets
+  (`.ods`) into the same schema as `.xlsx`.** Ported from
+  `particle-academy/holy-sheet`'s new reader, which has the full list of what
+  maps (values, formulas translated to A1, repeats, merges, comments, styles,
+  data styles, metadata) and what does not (frozen panes, column widths,
+  fonts, function-name translation). The format is sniffed from the bytes, so
+  the caller's branch on the file type can go.
+
+  Diffed against the PHP reader on the PHP repo's own fixtures, key order
+  included, in `tests/ods-reader-parity.test.ts`: a LibreOffice-converted
+  workbook, a hand-written native file, and a hand-built package of what
+  LibreOffice rewrites on save. `tests/ported-php/OdsReaderTest.test.ts` ports
+  the PHP assertions. Both load the fixtures from the PHP checkout
+  (`HOLY_SHEET_PHP_SRC`, or the sibling directory), as the parity suites
+  already did.
+
+- **`UnsupportedFormatException`**, exported, thrown for bytes that are
+  neither format, with the declared `mimetype` when there is one.
+- **`OdsReader`** and **`FormatSniffer`**, exported beside `XlsxReader`, and
+  `XlsxReader.readFiles()` for an already-unzipped package.
+- **`parseXml(src, { namespaces: true })`** resolves namespaces and keeps mixed
+  content in order. OpenDocument needs both; without the option the parser
+  behaves exactly as before.
+
+### Changed
+
+- **Unreadable bytes now throw `UnsupportedFormatException` instead of a bare
+  `Error`.** It extends `Error`, so **an existing `catch` keeps working: do
+  nothing.** The message for bytes that are not a zip still says "not a zip
+  archive"; only code matching the rest of the old text (`(no EOCD record)`,
+  `missing xl/workbook.xml`) sees different wording.
+
+### Fixed
+
+- **The tool schema announced its own shipped features as unreleased.**
+  `src/holy-sheet.schema.json` is a byte-identical copy of the PHP twin's, and
+  carried the same defect: every formatting field described in the future tense
+  ("Lands in 0.3.", "Lands in 0.5.") under a heading declaring a v0.2.0 that
+  could not style anything, while the package shipped at 2.x with all of it
+  working.
+
+  Updated in lockstep with `particle-academy/holy-sheet`; both checksum pins
+  moved together. See that package's entry for the measurement.
+
+- **An empty `cells: []` validates.** PHP's `describe()` reports a sheet with no cells as `cells: []`, and this validator rejected any array, so a schema read in PHP and written here failed on an empty sheet. The PHP validator had the same defect and is fixed in holy-sheet 2.2.0. A non-empty list is still an error.
+
+## 2.2.1 — 2026-09-10
+
+### Fixed
+
+- **`version()` reports the version this package actually ships as.** It
+  returned `1.0.0` from a 2.2.x release. The constant had drifted because
+  nothing compared it to the packaging metadata — the same shape as every other
+  two-copies-of-one-number failure in this estate.
+
+  `VersionIsSingleSourcedTest` / `version.test.ts` now pins it, so the class is
+  closed rather than the instance fixed. `dark-slide-py` already had that
+  assertion and was the only engine in the family to catch itself.
+
+
+## 2.2.0 — 2026-09-10
+
+### Added
+
+- **A checksum test pinning `holy-sheet.schema.json` to its PHP twin.** The
+  schema is a byte-identical copy of `holy-sheet/skills/holy-sheet.schema.json`,
+  kept in sync by remembering to edit both. Nothing checked that — and this is
+  the file handed to an LLM as the tool definition, so a one-sided edit did not
+  fail a build, it changed what an agent was told the API is on one backend and
+  not the other.
+
+  To change the schema: edit both copies, run either suite, paste the new hash
+  into both tests. The mild annoyance is the mechanism.
+
+  The hash is taken over **newline-normalised** content. Neither repo has a
+  `.gitattributes`, so the file is stored LF and lands CRLF on a Windows
+  checkout — a raw-bytes checksum would have passed on CI and failed on the
+  maintainer's own machine, which is a worse failure than the one being
+  prevented.
+
+### Fixed
+
+- **A sheet with BOTH a table and explicit cells no longer discards the table.**
+  `normalizeSheet` returned the moment `cells` was set, silently throwing away
+  `columns`, `rows`, `totals` and `theme`. A four-row report with one styled
+  title cell wrote a **one-cell workbook**, and `validate()` reported no errors,
+  because the validator explicitly permits both shapes together.
+
+  The PHP twin had the identical bug in the identical place. That is why no
+  parity test caught it: **parity detects disagreement, and the two runtimes
+  agreed.**
+
+  **What you must do: nothing.** A sheet using only one of the two shapes
+  behaves exactly as before. Explicit cells win at their address, and their
+  format is merged over the table's, so a cell asking only for `bold` keeps the
+  column's currency format instead of dropping to bare.
+
+- **Bare scalar cells are written instead of silently emptied.** `{A1: 42}` read
+  `.value` off a number, got `undefined`, and wrote a blank cell — no error,
+  just a hole where the number should be. **This one was Node-only:** PHP has
+  always had the branch, and the port dropped it.
+
+- **A bare `"=SUM(...)"` string is promoted to a real formula**, in both cell
+  maps and rows. Also Node-only: the same schema produced a working formula in
+  PHP and the literal text `=SUM(...)` in Node.
+
+- **A `comment` given as a plain string is no longer dropped** — only the object
+  form was read, and the string form was accepted by the validator and lost.
+
+## 2.1.0 — 2026-08-09
+
+### Fixed
+
+- **`1e300` was written to the sheet as `1e+3`.** The worst bug in this package's
+  history, and it never threw. `toFixed` switches to exponential notation at
+  1e21, and the trailing-zero strip then chewed the EXPONENT rather than a
+  fraction:
+
+  ```
+  (1e300).toFixed(14)         === "1e+300"
+  "1e+300".replace(/0+$/, "") === "1e+3"
+  ```
+
+  So a cell holding 1e300 was written holding 1000. Not a crash, not invalid
+  XML — a different number, in a file someone opens later and believes.
+
+- **`1e21` and above were written as `<v>1e+21</v>`**, which is not valid cell
+  content. This had a second cause the first fix did not reach: the writer only
+  called `formatFloat` for NON-integers, and `Number.isInteger(1e21)` is `true`,
+  so it took a `String(v)` branch. Mirroring PHP's `is_float(...)` test leaks —
+  a PHP int never stringifies exponentially, an integral JS double does.
+
+- **Numeric strings lost their exponent.** `numericStringToNumber` used
+  `parseInt` when the string had no `.`, and `parseInt` stops at the `e`:
+
+  | input | was | now |
+  |---|---|---|
+  | `"1e5"` | `1` | `100000` |
+  | `"1E5"` | `1` | `100000` |
+  | `"2e-3"` | `2` | `0.002` |
+  | `"1e21"` | `1` | `1e21` |
+
+  It is now `Number(v)` — which is what `FormulaLinter` already used for the
+  same job, so the two disagreed with each other as well as with PHP.
+
+- **`NaN` and `Infinity` were written into cells.** Both now write `0`, matching
+  the PHP twin.
+
+### Added
+
+- A `numericHazards` case in the cross-engine parity fixtures. The unit tables on
+  both sides were green while the engines still disagreed on `1e21`; only the
+  byte-for-byte diff caught it. Numbers are now a parity *case*, not a
+  convention.
+
+### Changed
+
+- **BREAKING in effect, though the API is identical:** values listed above now
+  serialize differently, because they were wrong. If you have a golden xlsx
+  containing an exponent-notation string, a value ≥ 1e21, or `NaN`, regenerate
+  it. Anything below 1e21 with no exponent string is byte-identical to 2.0.2.
+
+## 2.0.2 — 2026-08-09
+
+### Fixed
+
+- **`scripts/php-describe.php` was missed by 2.0.1**, so the reader-parity suite
+  still could not find the PHP classes outside the `.agi` envelope and CI failed
+  on the 2.0.1 commit. It now honours `HOLY_SHEET_PHP_SRC` like its sibling.
+
+  Worth saying plainly: CI caught this the very first time the parity suites
+  actually ran. Under the old `skipIf` the incomplete fix would have gone green.
+
+## 2.0.1 — 2026-08-09
+
+### Fixed
+
+- **The cross-engine parity suites now actually run in CI.** They are the
+  strongest guarantee this pair has — every OOXML part diffed byte-for-byte
+  against the PHP `holy-sheet` — and they had never executed on a runner.
+
+  Two things combined to hide it. `describe.skipIf(!HAS_PHP)` skipped silently,
+  and `ci.yml` installed Node only, so the build went green with **zero
+  cross-engine coverage**. Separately, the parity helper autoloaded the PHP
+  sources from a hard-coded `../../holy-sheet/src`, which only resolves inside
+  the `.agi` envelope — so even with PHP present, a different layout found no
+  classes.
+
+  Now: CI installs PHP 8.4 and checks out the PHP repo; the helper takes
+  `HOLY_SHEET_PHP_SRC` (falling back to the sibling path); and a missing PHP
+  **throws in CI** instead of skipping. A green build that asserts nothing is
+  worse than a red one, because nobody investigates green.
+
+  **What you must do:** nothing, unless you run these suites outside the
+  envelope — then set `HOLY_SHEET_PHP_SRC` to the PHP package's `src`.
+
+## 2.0.0 — 2026-08-07
+
+### Changed
+
+- **BREAKING — Node 18 is no longer supported.** `engines.node` moves from `>=18` to `>=22`.
+
+  **What you must do:** on Node 22 or newer, nothing. Note npm only *warns* on an `engines` mismatch while **pnpm fails the install**, so this surfaces differently depending on your package manager. Node 18 is end-of-life and 20 is maintenance-only.
+
+### Why
+
+These are the kit 0.5 platform floors, applied across every package at once so a consumer never has to resolve a mix. **No API changed, nothing was removed, nothing was renamed** — only what the package requires.
+
+This package is past 1.0, so a floor raise takes a **major**. Most of the suite is pre-1.0 and lands the identical change in a minor — that is semver, not a difference in how much changed.
+
+## 1.0.0 — 2026-05-30
+
+Initial Node/TypeScript port of `particle-academy/holy-sheet` (PHP), at
+**feature-parity with PHP 1.2.0**. Zero-dependency, isomorphic (browser + Node).
+
+- Full `Agent` surface: `validate`, `toBytes`, `write` (Node), `validateAndRepair`,
+  `lint`, `fromArray`, `fromCsv`, `read`/`describe`, `toolDefinition`, `version`,
+  plus the `HolySheet` instance class.
+- xlsx **writer** — inline-string/number/bool/date/formula cells, deduped
+  styles (fonts/fills/borders/numFmts/xfs), currency/percent/date number
+  formats, 4 themes, symbolic totals, merged cells, frozen panes, column
+  widths, comments (+ VML), multi-sheet.
+- xlsx **reader** — round-trips the above back to a schema (own + Excel-authored
+  files), via a hand-rolled isomorphic ZIP (STORE write / inflate read) and a
+  tiny XML parser.
+- **Formula linter** — `#VALUE!/#REF!/#DIV/0!/#NAME?/#CIRC!` with off-by-one
+  hints; 15+ functions.
+- **Verified byte-identical** to the PHP engine across sparse, themed/totaled,
+  decorated, and multi-sheet workbooks (cross-engine parity suite).