sfora-cli 0.9.0 → 0.11.0

package/dist/format/formatAxes.d.ts

@@ -0,0 +1,228 @@
+/**
+ * The format degrees of freedom, as a finite list somebody can read in one
+ * sitting.
+ *
+ * ---------------------------------------------------------------------------
+ * WHAT AN AXIS IS
+ * ---------------------------------------------------------------------------
+ *
+ * A format axis is a dimension along which the BYTES of a document may vary
+ * while its MEANING does not. `*em*` and `_em_` are two points on one axis.
+ * So are ` ``` ` and `~~~`, `# T` and `# T #`, a table's `-` and its `---`.
+ * Nothing on this list changes what a reader sees; everything on it changes
+ * what the file says on disk.
+ *
+ * The list exists because "lossy" and "byte-identical" are both useless as
+ * verdicts on their own. A serializer that reflows a document is not corrupting
+ * it, and a round trip that changes six bytes has not necessarily lost
+ * anything — but until the six bytes have a NAME, nobody reviewing the diff can
+ * tell which of those two sentences is true. Card #288: byte fidelity as a
+ * finite reviewable vocabulary, each entry with a witness that runs.
+ *
+ * The names follow open-knowledge's, whose `parser-drop-closure.ts:22-289`
+ * already spends axis ids the same way — `code-fence:char`, `link:title-quote`,
+ * `list:item-marker-spacing`, `setext:underline`. Where a name diverges it is
+ * because sfora's capture layer draws the line somewhere else, and the axis
+ * says so.
+ *
+ * ---------------------------------------------------------------------------
+ * WHERE THE LIST COMES FROM, AND WHY IT CANNOT BE INVENTED
+ * ---------------------------------------------------------------------------
+ *
+ * The axes are not a taxonomy somebody sat down and wrote. They are DERIVED
+ * from what the capture layer already records: `GEOMETRY_KEYS` in
+ * `editorWalker/sourceGeometry.ts` is the ground truth, and every
+ * `<nodeType>.<dataKey>` in it is one axis's `captures` entry. That is the
+ * whole discipline — an axis with no capture is a word, and a capture with no
+ * axis is a field nobody can name in a review. `__tests__/formatAxes.test.ts`
+ * closes the two against each other in both directions, so neither can drift.
+ *
+ * This file deliberately does NOT import `GEOMETRY_KEYS`. The capture layer
+ * lives under `editorWalker/`, which pulls unified and remark-stringify in and
+ * is excluded from the CLI sync (`packages/sfora/scripts/sync-format.mjs`);
+ * the drop ledger that spends these ids is NOT excluded and must keep
+ * compiling in the tarball. So the coupling is a string in a test rather than
+ * an import, and the test is what makes the string true.
+ *
+ * ---------------------------------------------------------------------------
+ * THREE CAPTURE LAYERS, ONE VOCABULARY
+ * ---------------------------------------------------------------------------
+ *
+ * Sfora has three places where an author's spelling is read and set aside, and
+ * they are not the same machinery:
+ *
+ *   - the mdast geometry capture, which reads spellings off a POSITIONED parse
+ *     tree and stores them on `node.data` (`layer: "mdast-geometry"`);
+ *   - the structured-block parsers, which read a fence body line by line and
+ *     report what they did not keep (`layer: "structured-block"`);
+ *   - the blank-run plan, which reads the gaps BETWEEN blocks and the runs at
+ *     the document's own two ends (`layer: "blank-runs"`).
+ *
+ * The second is why this file is not simply a rename of `GEOMETRY_KEYS`. The
+ * wave-1 drop ledger already spends the id `structured-block:entry-bullet` on
+ * a chat entry authored `* @mara: …` rather than `- @mara: …`, and that
+ * spelling has no `node.data` anywhere — the bullet is dropped, not captured.
+ * An axis vocabulary that could not name it would leave the one verdict it was
+ * built for pointing at a string nobody defined.
+ *
+ * The third is here because the list has to cover what the walker REPLAYS, not
+ * what one module happens to declare. `blankRuns.ts` writes
+ * `sourceBlankLinesBefore` and `sourceEdgeBlankLines`, `emit.ts`'s join reads
+ * both back, and an author's double gap between two sections survives a save
+ * because of it — a degree of freedom as real as a fence character, and one the
+ * first version of this list was silently missing because those two keys are
+ * written by the WALKER onto the tree it builds rather than by
+ * `captureSourceGeometry` onto a parse, and so are structurally outside
+ * `GEOMETRY_KEYS`. Deriving the list from one module's export rather than from
+ * the behaviour is exactly how a vocabulary ends up with a hole in it.
+ *
+ * So the layer is a discriminant and the closure test grades each layer against
+ * its own ground truth: the geometry axes against `GEOMETRY_KEYS`, the block
+ * axes against the `format-dof-axis` entries of `BLOCK_DROP_ADJUDICATIONS`, and
+ * the blank-run axes against the keys a live `captureRootGaps` /
+ * `applyRootGaps` pair actually writes over a corpus. That third ground truth
+ * is BEHAVIOURAL rather than declarative, and the suite says so: the two keys
+ * are declared in a `declare module "mdast"` block, which erases at runtime, so
+ * there is no list to import. A key the corpus never provokes would be invisible
+ * to it, which is a real limit and not a hidden one.
+ *
+ * ---------------------------------------------------------------------------
+ * WHAT A WITNESS PROVES, AND WHAT IT DOES NOT
+ * ---------------------------------------------------------------------------
+ *
+ * Every axis carries at least one DISCRIMINATION WITNESS: two spellings, `a`
+ * and `b`. The name matters, because these are not round-trip witnesses and
+ * calling them that would claim the one thing they do not show. The suite
+ * asserts three things about each pair, and the third is what stops the list
+ * becoming decoration.
+ *
+ *   1. the bytes differ — otherwise there is no degree of freedom here;
+ *   2. an INDEPENDENT reparse reads both to the same tree, positions and
+ *      `data` stripped (`geometryReparsesTheSame`) — otherwise this is a
+ *      content difference wearing a formatting name;
+ *   3. the capture layer TELLS THEM APART — the set of captured keys that move
+ *      between `a` and `b` is non-empty, and across an axis's witnesses it
+ *      covers every key the axis claims.
+ *
+ * Three is the anti-vacuity clause. An axis whose witnesses no longer move its
+ * own capture key is an axis whose capture stopped happening, and the suite
+ * fails on the axis rather than on some distant serializer test six months
+ * later.
+ *
+ * What a discrimination witness does NOT prove is that the spelling survives a
+ * save. Most of these axes are captured and not yet replayed —
+ * `CAPTURED_NOT_REPLAYED` in `sourceGeometry.ts` says which, and names the card
+ * that owes each one. The suite pins the replayed SET rather than a total,
+ * because a total can hold still while the membership churns — and, for one
+ * axis in each replay FAMILY, it serializes an editor document through the real
+ * walker and asserts the author's bytes come back. A scoreboard read off a
+ * hand-maintained array is a scoreboard that can be wrong in the direction
+ * nobody checks.
+ */
+import type { BlockDropReason } from "./blocks/dropClosure.js";
+import type { StructuredBlockLanguage } from "./blocks/structured-block-schema.js";
+/**
+ * Every axis id, as a union so a ledger entry citing one that does not exist
+ * fails to compile rather than fails to mean anything.
+ *
+ * `namespace:name`, lower case, and the namespace is the CONSTRUCT rather than
+ * the mdast node type where the two disagree — `task-list:checkbox-char` sits
+ * on `listItem.sourceCheckboxChar`, and an author who has never heard of mdast
+ * still knows what a task list is.
+ */
+export type FormatAxisId = "blank-lines:doc-edge" | "blank-lines:interior" | "block-gap:after-heading" | "blockquote:marker-spacing" | "code-block:style" | "code-fence:char" | "code-fence:info-padding" | "code-fence:length" | "definition:layout" | "definition:title-quote" | "emphasis:delimiter" | "hard-break:style" | "heading:style" | "heading:trailing-hashes" | "inline-code:fence-length" | "inline-code:padding" | "lexical:backslash-escape" | "link:angle-url" | "link:autolink-form" | "link:title-quote" | "list:bullet-marker" | "list:item-marker-spacing" | "list:item-ordinal" | "list:ordered-delimiter" | "setext:underline-length" | "strike:delimiter" | "strong:delimiter" | "structured-block:entry-bullet" | "table:cell-padding" | "table:delimiter-dashes" | "table:delimiter-padding" | "table:outer-pipes" | "task-list:checkbox-char" | "thematic-break:marker";
+/**
+ * Two spellings of one meaning, for an axis the mdast geometry layer captures.
+ *
+ * Both are whole markdown documents rather than fragments, because the capture
+ * layer reads source OFFSETS and a fragment has none. They are kept as short as
+ * the axis allows so a reader can see the difference without diffing.
+ */
+export interface GeometryAxisWitness {
+    a: string;
+    b: string;
+    /** What moved, in the words a reviewer would use. */
+    note: string;
+}
+/** Two spellings of one meaning, for an axis a structured-block parser drops. */
+export interface BlockAxisWitness {
+    lang: StructuredBlockLanguage;
+    /** The spelling whose bytes survive. */
+    a: string;
+    /** The spelling that provokes the drop. */
+    b: string;
+    note: string;
+}
+export type FormatAxis = {
+    id: FormatAxisId;
+    layer: "mdast-geometry";
+    /**
+     * `<nodeType>.<dataKey>`, exactly as `GEOMETRY_KEYS` spells it. Every
+     * entry must be declared there and every entry there must appear on
+     * exactly one axis.
+     */
+    captures: readonly string[];
+    rationale: string;
+    witnesses: readonly GeometryAxisWitness[];
+} | {
+    id: FormatAxisId;
+    layer: "blank-runs";
+    /**
+     * The `data` keys `blankRuns.ts` writes, by NAME rather than as
+     * `<nodeType>.<dataKey>`.
+     *
+     * The node type is not part of the identity here and pretending it were
+     * would be a lie in the vocabulary: an interior gap belongs to whatever
+     * block happens to follow it, so `sourceBlankLinesBefore` lands on a
+     * paragraph, a code block or a table depending on the document. (The edge
+     * run is always on the root, and is spelled the same way for symmetry.)
+     */
+    dataKeys: readonly string[];
+    rationale: string;
+    witnesses: readonly GeometryAxisWitness[];
+} | {
+    id: FormatAxisId;
+    layer: "structured-block";
+    /**
+     * The drop reasons in `BLOCK_DROP_ADJUDICATIONS` whose `format-dof-axis`
+     * verdict spends this id. Closed both directions by the suite.
+     */
+    ledgerReasons: readonly BlockDropReason[];
+    rationale: string;
+    witnesses: readonly BlockAxisWitness[];
+};
+/**
+ * The list. Sorted by id, because a list nobody can scan is not reviewable and
+ * reviewable is the entire point.
+ */
+export declare const FORMAT_AXES: readonly FormatAxis[];
+/** Every id on the list, for a cheap membership test. */
+export declare const FORMAT_AXIS_IDS: ReadonlySet<FormatAxisId>;
+/**
+ * The axis with this id.
+ *
+ * Throws rather than returning undefined: every call site here is spending an
+ * id that the type system already proved exists, so a miss means the list and
+ * the union disagree, which is a bug in this file and not a case a caller
+ * should be writing a branch for.
+ */
+export declare function formatAxis(id: FormatAxisId): FormatAxis;
+/**
+ * The axis that owns a captured geometry key, spelled `<nodeType>.<dataKey>`.
+ *
+ * Undefined here is the interesting answer, not an error: it means the capture
+ * layer records a spelling the vocabulary cannot name, which is exactly what
+ * the closure test is looking for.
+ */
+export declare function axisForCapture(key: string): FormatAxis | undefined;
+/** Every `<nodeType>.<dataKey>` the vocabulary claims, deduplicated. */
+export declare function claimedCaptureKeys(): string[];
+/**
+ * Every blank-run `data` key the vocabulary claims, deduplicated.
+ *
+ * Kept apart from {@link claimedCaptureKeys} rather than merged into it because
+ * the two are spelled differently on purpose — one carries a node type and the
+ * other cannot — and a single list would have to pick one spelling and be wrong
+ * about half its entries.
+ */
+export declare function claimedBlankRunKeys(): string[];

package/dist/format/formatAxes.js

@@ -0,0 +1,454 @@
+/**
+ * The list. Sorted by id, because a list nobody can scan is not reviewable and
+ * reviewable is the entire point.
+ */
+export const FORMAT_AXES = [
+    {
+        id: "blank-lines:doc-edge",
+        layer: "blank-runs",
+        dataKeys: ["sourceEdgeBlankLines"],
+        rationale: "blank runs at the document's own two ends; `emitMarkdown` strips the terminator and writes no leading run, so without the plan an author's deliberate breathing room at the top or bottom of a body is gone on the first save",
+        witnesses: [
+            {
+                a: "A",
+                b: "A\n\n\n",
+                note: "a bare body against one with a trailing blank run",
+            },
+        ],
+    },
+    {
+        id: "blank-lines:interior",
+        layer: "blank-runs",
+        dataKeys: ["sourceBlankLinesBefore"],
+        rationale: "how many blank lines the author left between two root blocks; mdast-util-to-markdown writes exactly one and mdast has no field for the rest, so a double gap used as a section break closes on every save",
+        witnesses: [
+            {
+                a: "A\n\nB",
+                b: "A\n\n\nB",
+                note: "one blank line between two paragraphs against two",
+            },
+        ],
+    },
+    {
+        id: "block-gap:after-heading",
+        layer: "mdast-geometry",
+        captures: ["heading.sourceContiguousNext"],
+        rationale: "whether a heading is followed by a blank line before its first block; remark writes one unconditionally, so without this axis every contiguous heading in a document reflows on the first save",
+        witnesses: [
+            {
+                a: "# T\nbody",
+                b: "# T\n\nbody",
+                note: "the gap between the heading and the paragraph under it",
+            },
+        ],
+    },
+    {
+        id: "blockquote:marker-spacing",
+        layer: "mdast-geometry",
+        captures: ["blockquote.sourceMarkerSpacings"],
+        rationale: "how many spaces followed each `>`; CommonMark strips one and the rest are the author's, per line",
+        witnesses: [
+            {
+                a: "> q",
+                b: ">q",
+                note: "one space after the marker against none",
+            },
+        ],
+    },
+    {
+        id: "code-block:style",
+        layer: "mdast-geometry",
+        captures: ["code.sourceStyle"],
+        rationale: "fenced against indented; the same four lines of code, spelled two ways that CommonMark reads identically",
+        witnesses: [
+            {
+                a: "```\nx\n```",
+                b: "    x",
+                note: "a fence against a four-space indent — the widest spelling gap on the list",
+            },
+        ],
+    },
+    {
+        id: "code-fence:char",
+        layer: "mdast-geometry",
+        captures: ["code.sourceFenceChar"],
+        rationale: "backtick against tilde; the choice matters to an author whose body contains the other one",
+        witnesses: [
+            {
+                a: "```\nx\n```",
+                b: "~~~\nx\n~~~",
+                note: "the fence character",
+            },
+        ],
+    },
+    {
+        id: "code-fence:info-padding",
+        layer: "mdast-geometry",
+        captures: ["code.sourceInfoPadding"],
+        rationale: "spaces between the fence and its info string; CommonMark discards them and some authors line their fences up with them",
+        witnesses: [
+            {
+                a: "```ts\nx\n```",
+                b: "```   ts\nx\n```",
+                note: "the gap before the language",
+            },
+        ],
+    },
+    {
+        id: "code-fence:length",
+        layer: "mdast-geometry",
+        captures: ["code.sourceFenceLength"],
+        rationale: "how many fence characters; three is the minimum and more is a deliberate widening around a body that contains a closer-shaped line",
+        witnesses: [
+            {
+                a: "```\nx\n```",
+                b: "````\nx\n````",
+                note: "three characters against four",
+            },
+        ],
+    },
+    {
+        id: "definition:layout",
+        layer: "mdast-geometry",
+        captures: ["definition.sourceLayout"],
+        rationale: "whether a link reference definition sits on one line or wraps after the colon",
+        witnesses: [
+            {
+                a: "[a]: u",
+                b: "[a]:\n  u",
+                note: "the destination on the label's line against the line below it",
+            },
+        ],
+    },
+    {
+        id: "definition:title-quote",
+        layer: "mdast-geometry",
+        captures: ["definition.sourceTitleMarker"],
+        rationale: "which of the three title containers a definition used — `\"`, `'` or `()`",
+        witnesses: [
+            {
+                a: '[a]: u "t"',
+                b: "[a]: u 't'",
+                note: "double quotes against single",
+            },
+        ],
+    },
+    {
+        id: "emphasis:delimiter",
+        layer: "mdast-geometry",
+        captures: ["emphasis.sourceDelimiter"],
+        rationale: "`*` against `_` for a single-delimiter emphasis run",
+        witnesses: [
+            { a: "*em*", b: "_em_", note: "the delimiter character" },
+        ],
+    },
+    {
+        id: "hard-break:style",
+        layer: "mdast-geometry",
+        captures: ["break.sourceStyle"],
+        rationale: "a hard break spelled as a trailing backslash or as two trailing spaces; one of them is invisible in every editor ever written",
+        witnesses: [
+            {
+                a: "a\\\nb",
+                b: "a  \nb",
+                note: "backslash against two spaces",
+            },
+        ],
+    },
+    {
+        id: "heading:style",
+        layer: "mdast-geometry",
+        captures: ["heading.sourceStyle"],
+        rationale: "atx (`# T`) against setext (`T` over `=`); the same heading, and the only two spellings CommonMark has for depth one and two",
+        witnesses: [
+            {
+                a: "# T",
+                b: "T\n=",
+                note: "atx against setext — moves the underline length with it, which is the next axis's business",
+            },
+        ],
+    },
+    {
+        id: "heading:trailing-hashes",
+        layer: "mdast-geometry",
+        captures: ["heading.sourceTrailingHashes"],
+        rationale: "the optional closing run on an atx heading, which CommonMark discards entirely",
+        witnesses: [
+            { a: "# T", b: "# T #", note: "the closing hash" },
+        ],
+    },
+    {
+        id: "inline-code:fence-length",
+        layer: "mdast-geometry",
+        captures: ["inlineCode.sourceFenceLength"],
+        rationale: "how many backticks open and close a code span; more than one is how an author writes a backtick inside one",
+        witnesses: [
+            { a: "`x`", b: "``x``", note: "one backtick against two" },
+        ],
+    },
+    {
+        id: "inline-code:padding",
+        layer: "mdast-geometry",
+        captures: ["inlineCode.sourcePadded"],
+        rationale: "the one space either side that CommonMark strips, which is how `` ` `` itself is written inside a span",
+        witnesses: [
+            { a: "`x`", b: "` x `", note: "the stripped padding" },
+        ],
+    },
+    {
+        id: "lexical:backslash-escape",
+        layer: "mdast-geometry",
+        captures: ["text.escapedChars"],
+        rationale: "where the author put a backslash in front of a character that did not need one; the value is the same string either way, and re-deriving the escapes from it is how a serializer starts escaping things nobody typed",
+        witnesses: [
+            {
+                a: "a\\*b",
+                b: "a*b",
+                note: "an escaped asterisk against a bare one that opens nothing",
+            },
+        ],
+    },
+    {
+        id: "link:angle-url",
+        layer: "mdast-geometry",
+        captures: ["link.sourceUrlForm"],
+        rationale: "`[a](<u>)` against `[a](u)`; the angle form is what an author reaches for when the destination has a space in it, and keeping it is how the habit survives an edit elsewhere",
+        witnesses: [
+            { a: "[a](u)", b: "[a](<u>)", note: "the angle brackets around the destination" },
+        ],
+    },
+    {
+        id: "link:autolink-form",
+        layer: "mdast-geometry",
+        captures: ["link.sourceStyle"],
+        rationale: "a bare GFM literal autolink against the same url in angle brackets; both are one link node with the same href",
+        witnesses: [
+            {
+                a: "https://e.com",
+                b: "<https://e.com>",
+                note: "the gfm literal form against the angle form",
+            },
+        ],
+    },
+    {
+        id: "link:title-quote",
+        layer: "mdast-geometry",
+        captures: ["link.sourceTitleMarker"],
+        rationale: "which of the three title containers an inline link used — `\"`, `'` or `()`",
+        witnesses: [
+            { a: '[a](u "t")', b: "[a](u 't')", note: "double quotes against single" },
+        ],
+    },
+    {
+        id: "list:bullet-marker",
+        layer: "mdast-geometry",
+        captures: ["list.bulletMarker"],
+        rationale: "`-`, `*` or `+`; mdast-util-to-markdown picks one globally and alternates it between adjacent lists, so without the axis a document converges on whichever the serializer likes",
+        witnesses: [
+            { a: "- a", b: "* a", note: "the bullet character" },
+        ],
+    },
+    {
+        id: "list:item-marker-spacing",
+        layer: "mdast-geometry",
+        captures: ["listItem.sourceMarkerSpacing"],
+        rationale: "spaces between a list marker and its content; one to four all mean the same item and set different continuation indents",
+        witnesses: [
+            { a: "- a", b: "-   a", note: "one space after the bullet against three" },
+        ],
+    },
+    {
+        id: "list:item-ordinal",
+        layer: "mdast-geometry",
+        captures: ["listItem.sourceOrdinal"],
+        rationale: "the number each ordered item was actually typed with; mdast keeps the list's `start` and nothing else, so `1. 1. 1.` and `1. 2. 3.` are the same tree",
+        witnesses: [
+            {
+                a: "1. a\n1. b",
+                b: "1. a\n2. b",
+                note: "all-ones against counting up",
+            },
+        ],
+    },
+    {
+        id: "list:ordered-delimiter",
+        layer: "mdast-geometry",
+        captures: ["list.listMarkerDelimiter"],
+        rationale: "`1.` against `1)` for an ordered list",
+        witnesses: [
+            { a: "1. a", b: "1) a", note: "the character after the ordinal" },
+        ],
+    },
+    {
+        id: "setext:underline-length",
+        layer: "mdast-geometry",
+        captures: ["heading.sourceUnderlineLength"],
+        rationale: "how long the `=` or `-` run under a setext heading is; one character is enough and authors rule the whole width",
+        witnesses: [
+            { a: "T\n=", b: "T\n===", note: "the underline run" },
+        ],
+    },
+    {
+        id: "strike:delimiter",
+        layer: "mdast-geometry",
+        captures: ["delete.sourceDelimiter"],
+        rationale: "GFM's `~x~` against `~~x~~`; the single-tilde form is not in every reader's dialect, which is why the capture lands and the replay is a rendering decision",
+        witnesses: [
+            { a: "~~x~~", b: "~x~", note: "two tildes against one" },
+        ],
+    },
+    {
+        id: "strong:delimiter",
+        layer: "mdast-geometry",
+        captures: ["strong.sourceDelimiter"],
+        rationale: "`**` against `__` for a double-delimiter run",
+        witnesses: [
+            { a: "**s**", b: "__s__", note: "the delimiter run" },
+        ],
+    },
+    {
+        id: "structured-block:entry-bullet",
+        layer: "structured-block",
+        ledgerReasons: ["entry-bullet-marker"],
+        rationale: "a status or chat entry may be authored bare, with `-`, or with `*`; the block grammars strip the marker before reading the entry, so all three parse to the same timeline and which one was typed is a spelling rather than content",
+        witnesses: [
+            {
+                lang: "chat",
+                a: "- @mara: on it\n",
+                b: "* @mara: on it\n",
+                note: "the entry bullet — the id the wave-1 drop ledger already spends",
+            },
+            {
+                lang: "status",
+                a: "state: building\n- @ada: on it\n",
+                b: "state: building\n* @ada: on it\n",
+                note: "the same axis one block language over, because the ledger entry is not chat-specific",
+            },
+        ],
+    },
+    {
+        id: "table:cell-padding",
+        layer: "mdast-geometry",
+        captures: ["tableCell.sourcePadding"],
+        rationale: "the spaces inside each cell's pipes, which is what makes a hand-aligned table a grid instead of a wall",
+        witnesses: [
+            {
+                a: "| a |\n| - |",
+                b: "|a|\n| - |",
+                note: "padded cells against tight ones",
+            },
+        ],
+    },
+    {
+        id: "table:delimiter-dashes",
+        layer: "mdast-geometry",
+        captures: ["table.sourceDashCounts"],
+        rationale: "how many dashes each delimiter cell holds; one is legal and authors widen them to the column",
+        witnesses: [
+            {
+                a: "| a |\n| - |",
+                b: "| a |\n| --- |",
+                note: "a one-dash delimiter against a three-dash one",
+            },
+        ],
+    },
+    {
+        id: "table:delimiter-padding",
+        layer: "mdast-geometry",
+        captures: ["table.sourceAlignmentPadding"],
+        rationale: "the spaces around the dashes in the delimiter row, captured apart from the dash count because the two move independently",
+        witnesses: [
+            {
+                a: "| a |\n|---|",
+                b: "| a |\n| --- |",
+                note: "the same three dashes, padded and not",
+            },
+        ],
+    },
+    {
+        id: "table:outer-pipes",
+        layer: "mdast-geometry",
+        captures: ["table.sourceOuterPipes"],
+        rationale: "whether the rows carry their leading and trailing pipes; GFM makes both optional and the pipe-less form is a real authoring style",
+        witnesses: [
+            {
+                a: "| a | b |\n| :-- | --: |\n| 1 | 2 |",
+                b: "a | b\n:-- | --:\n1 | 2",
+                note: "edge pipes against none — the alignment colons keep the delimiter row from reading as a setext underline",
+            },
+        ],
+    },
+    {
+        id: "task-list:checkbox-char",
+        layer: "mdast-geometry",
+        captures: ["listItem.sourceCheckboxChar"],
+        rationale: "`[x]` against `[X]`; GFM checks the box either way and one of them is what the author's other tool writes",
+        witnesses: [
+            { a: "- [x] a", b: "- [X] a", note: "the checkbox character's case" },
+        ],
+    },
+    {
+        id: "thematic-break:marker",
+        layer: "mdast-geometry",
+        captures: ["thematicBreak.sourceRaw"],
+        rationale: "the whole authored run — `***`, `---`, `- - -`, `___` and every spaced variant are one thematic break",
+        witnesses: [
+            { a: "***", b: "- - -", note: "the marker run, verbatim" },
+        ],
+    },
+];
+/** Every id on the list, for a cheap membership test. */
+export const FORMAT_AXIS_IDS = new Set(FORMAT_AXES.map((axis) => axis.id));
+/**
+ * The axis with this id.
+ *
+ * Throws rather than returning undefined: every call site here is spending an
+ * id that the type system already proved exists, so a miss means the list and
+ * the union disagree, which is a bug in this file and not a case a caller
+ * should be writing a branch for.
+ */
+export function formatAxis(id) {
+    const found = FORMAT_AXES.find((axis) => axis.id === id);
+    if (!found)
+        throw new Error(`no format axis with id ${id}`);
+    return found;
+}
+/**
+ * The axis that owns a captured geometry key, spelled `<nodeType>.<dataKey>`.
+ *
+ * Undefined here is the interesting answer, not an error: it means the capture
+ * layer records a spelling the vocabulary cannot name, which is exactly what
+ * the closure test is looking for.
+ */
+export function axisForCapture(key) {
+    return FORMAT_AXES.find((axis) => axis.layer === "mdast-geometry" && axis.captures.includes(key));
+}
+/** Every `<nodeType>.<dataKey>` the vocabulary claims, deduplicated. */
+export function claimedCaptureKeys() {
+    const out = new Set();
+    for (const axis of FORMAT_AXES) {
+        if (axis.layer !== "mdast-geometry")
+            continue;
+        for (const key of axis.captures)
+            out.add(key);
+    }
+    return [...out].sort();
+}
+/**
+ * Every blank-run `data` key the vocabulary claims, deduplicated.
+ *
+ * Kept apart from {@link claimedCaptureKeys} rather than merged into it because
+ * the two are spelled differently on purpose — one carries a node type and the
+ * other cannot — and a single list would have to pick one spelling and be wrong
+ * about half its entries.
+ */
+export function claimedBlankRunKeys() {
+    const out = new Set();
+    for (const axis of FORMAT_AXES) {
+        if (axis.layer !== "blank-runs")
+            continue;
+        for (const key of axis.dataKeys)
+            out.add(key);
+    }
+    return [...out].sort();
+}