@agentshouse/mdmodel 0.1.0-cc9fa0498a7f7 → 0.1.0-cd90851746292

package/API.md

@@ -26,30 +26,34 @@ if (checked.verdict === 'accept') {
 }
 const parsed = parse(content);
 if (parsed.ok) {
-  for (const link of links(parsed.document, 'tasks/two.md')) {
-    if (link.target !== 'data/export.csv') continue;
-    const moved = apply(parsed.document, {
-      kind: 'link',
-      href: link.href,
-      to: '../files/export.csv',
-    });
-    if (moved.outcome === 'applied') serialize(moved.document);
-  }
+  const moved = apply(
+    parsed.document,
+    links(parsed.document, 'tasks/two.md')
+      .filter((link) => link.target === 'data/export.csv')
+      .map((link) => ({ kind: 'link' as const, href: link.href, to: '../files/export.csv' })),
+  );
+  if (moved.outcome === 'applied') serialize(moved.document);
 }
 ```
 
 `MARKDOWN_SYNTAX_PROFILE` is the Markdown every reader here applies, GitHub Flavored Markdown
 0.29, and `VERSION` is the MD model version.
 
-`verify(schema)` reads one schema text, the MD Types of one Room or App: one
-JSON Schema document (draft 2020-12) in the YAML subset. Every top-level key but
-`$defs` names an MD Type holding the JSON Schema of its frontmatter; `$defs`
-holds shapes the types reuse, reached by a `$ref` under `#/$defs/` alone. `type`
-and the built-in `any` are never declared. The model adds three keywords:
-`reference` on a string names one MD Type, several, `any` or `attachment`;
-`sections` declares a type's H2 sections, each `{ kind: text }` or `{ kind:
-list, items }` over one string, number, integer or boolean, optionally
-`required: true`; `member_document: true` marks at most one type.
+`verify(schema)` reads one schema text, the MD Types of one Room or App: one JSON Schema document
+(draft 2020-12) in the YAML subset. Every top-level key but `$defs` names an MD Type holding the
+JSON Schema of its frontmatter; `$defs` holds shapes the types reuse, reached by a `$ref` under
+`#/$defs/` alone. `type` and the built-in `any` are never declared. The model adds its keywords:
+`reference` on a string names one MD Type, several, `any` or `attachment`, and `cross_room: true`
+beside it admits a Document in another Room, of a type that Room declares; `sections` declares a
+type's H2 sections, each `{ kind: text }` or `{ kind: list, items }` over one string, number,
+integer or boolean, optionally `required: true`, which the `then` or `else` of an `if` on the type,
+in its `allOf` or in such a branch may repeat to ask a section for text, else `conditional-section`;
+`member_document: true` marks at most one type. A type alone takes `create_roles` and
+`delete_roles`, each a list of role names. `verify(schema, provided)` holds every type of a model
+`provided` as its own, named only for those two and only if `provided.nameable` lists it. On a field
+of a type's own `properties` alone, `change_roles` and `initial_roles` each list role names, and
+`transitions` lists `{ from, to, roles? }`, each end a scalar or `null`, on a field whose `type`,
+`enum`, `const`, `$ref`, `allOf`, `anyOf`, `oneOf`, `if` and `not` admit no list or object.
 
 ```yaml
 $defs:
@@ -64,70 +68,72 @@ plan:
 contact: { member_document: true }
 ```
 
-The model's `types` expose each MD Type's `name`, `memberDocument`, `sections`,
-`shape` (the type as written with every `$defs` shape it reaches), the facts of
-its `frontmatter` and `fields`, its top-level ones: JSON `type`, `format`,
-`values`, `reference` targets, `items`, `properties`, an open map's `patterns`
-and `additional`, and `prose`, a string with no `enum`, `const`, `format`,
-`pattern` or `reference`. Facts join `$ref` and `allOf` to their siblings and
-the variants of `anyOf`, `oneOf`, `then`, `else` and `dependentSchemas`, a
-variant's own keys `conditional`; a path without one `type`, `null` aside, has
-none. A malformed schema is refused whole, naming its `rule`, the keyword or the
-model's own, and the JSON Pointer `location` it concerns, such as
-`/plan/properties/status/enum`, or `line <n>` outside the YAML subset.
-`pointerFacts(definition, pointer)` reads the facts a frontmatter pointer reaches,
-`itemFacts` and `memberFacts` take one step and `crossedFacts` reads a list's item
-facts; `pointerSegments` and `escapePointer` split and write a pointer, `ARRAY_INDEX`
-matches an index segment, and `ATTACHMENT` is the `reference` that admits any path.
-
-`validate(library, context?, model?)` takes the resulting text Library and one
-`ValidationContext` of `binaryTargets: readonly string[]`, none when omitted.
-Library keys and binary targets share one canonical collection-relative path
-namespace; a path declared text and binary is refused. House declares every
-binary file of the Room as a binary target, and the model reads no bytes. Ordinary links and images
-in typed Documents may resolve to either membership, relative to the source
-Document, and one to a path neither holds is not refused: it dangles, as a
-deleted file's links do. A destination that is an absolute path or leaves the
-collection is `unresolved-link`. A binary target supplies no Document or MD Type. `model` is the
-verified schema that types the Library, and every Library member is content,
-whatever its Path. Without `model` the Library declares no type.
-
-A typed Document's frontmatter satisfies its type's schema, every keyword with
-its JSON Schema meaning, formats asserted and defaults never filled, except that
-an object schema declaring neither `additionalProperties` nor
-`unevaluatedProperties` is closed: an undeclared key is refused as `undeclared-
-field`. `TYPE_FIELD`, `SITE_FIELD` and `RECEIVED_FIELD` count as declared on every
-type, and the string `MEMBER_FIELD` on the `member_document` type; `HOUSE_FIELDS`
-lists the four. A refusal names the keyword, such as `required`, and the
-value's JSON Pointer, such as `/shipping/city`. Every occurrence of a declared
-list section is a list of values of its item schema, else `structured-section`;
-a reference item is one ordinary link alone, such as `- [Ada](ada.md)`. A
-reference, at any depth or a section item, reaching a URL, a binary target, or a
-Document of a type it does not name is `unresolved-reference`, and a frontmatter
-one reaching no member is `referenced-missing-file` naming its `referrers`;
-`attachment` admits any path, and no body link is a referrer.
-
-Validation returns acceptance or one refusal naming the file, the field or
-section, and the violated rule. `parse`, `apply`, and `serialize` operate on a single text
-Document; they do not validate collection membership or consume binary bytes.
-`apply` sets one typed field, including a list or an object as a whole, replaces
-one existing section body or the preamble against that value's prior content,
-or rewrites one Markdown link destination. A section need not be declared by
-the Document's MD Type to be replaced. A field change does not reformat the rest of the file:
-an inline value is replaced where it stands, and a value written after the key's line, an empty
-one or a block mapping or list, is written after the key's colon in its place.
-A field holds a `FrontmatterValue`: null, a string, a boolean, a finite number, a list
-without holes or a plain object of them, holding no cycle. `frontmatterValue(value)` is that test, and
-`apply` refuses any other value as `frontmatter-fence`; `equalValue(left, right)` compares
-two by value, whatever their key order, as `apply` compares a field with its prior.
-
-`Document.preamble` is a readonly string containing the exact source after
-frontmatter and before the first section heading recognized by the existing
-section parser. Without a section it is the entire post-frontmatter body;
-without frontmatter it starts at the beginning of the Document. An empty span
-is `''`. It includes whitespace, newlines, a leading H1, and any fenced text
-before the first section. The closing frontmatter delimiter and its line ending
-belong to frontmatter. Section recognition and the syntax profile are unchanged.
+The model's `types` expose each MD Type's `name`, `memberDocument`, `sections`, `shape` (the type as
+written, its rule keywords left out, with every `$defs` shape it reaches), `createRoles` and
+`deleteRoles` when written, its `rules` by field as `changeRoles`, `initialRoles` and `transitions`,
+the facts of its `frontmatter` and `fields`, its top-level ones: JSON `type`, `format`, `values`,
+`reference` targets, `items`, `properties`, an open map's `patterns` and `additional`, and `prose`,
+a string with no `enum`, `const`, `format`, `pattern` or `reference`. Facts join `$ref` and `allOf`
+to their siblings and the variants of `anyOf`, `oneOf`, `then`, `else` and `dependentSchemas`, a
+variant's own keys `conditional`; a path without one `type`, `null` aside, has none. A malformed
+schema is refused whole, naming its `rule`, the keyword or the model's own, and the JSON Pointer
+`location` it concerns, such as `/plan/properties/status/enum`, or `line <n>` outside the YAML
+subset. `pointerFacts(definition, pointer)` reads the facts a frontmatter pointer reaches, `itemFacts`
+and `memberFacts` take one step and `crossedFacts` reads a list's item facts; `JSON_POINTER` matches the
+JSON Pointer grammar, `''` included, `pointerSegments` and `escapePointer` split and write a pointer,
+`ARRAY_INDEX` matches an index segment, and `ATTACHMENT` is the `reference` that admits any path.
+
+`validate(library, context?, model?)` takes the resulting text Library and one `ValidationContext`
+of `binaryTargets: readonly string[]`, none when omitted. Library keys and binary targets share one
+canonical collection-relative path namespace; a path declared text and binary is refused. House
+declares every binary file of the Room as a binary target, and the model reads no bytes. Ordinary
+links and images in typed Documents may resolve to either membership, relative to the source
+Document, and one to a path neither holds is not refused: it dangles, as a deleted file's links do.
+One written `/rooms/<handle>/<path>` names a file in another Room, which the caller checks; any
+other absolute path, or one that leaves the collection, is `unresolved-link`. A binary target
+supplies no Document or MD Type. `model` is the verified schema that types the Library, and every
+Library member is content, whatever its Path. Without `model` the Library declares no type.
+
+A typed Document's frontmatter satisfies its type's schema, every keyword with its JSON Schema
+meaning, formats asserted and defaults never filled, except that an object schema declaring neither
+`additionalProperties` nor `unevaluatedProperties` is closed: an undeclared key is refused as
+`undeclared-field`. `TYPE_FIELD`, `SITE_FIELD` and `RECEIVED_FIELD` count as declared on every type,
+and the string `MEMBER_FIELD` on the `member_document` type; `HOUSE_FIELDS` lists the four. A refusal
+names the keyword, such as `required`, and the value's JSON Pointer, such as `/shipping/city`. Every
+occurrence of a declared list section is a list of values of its item schema, else
+`structured-section`; a reference item is one ordinary link alone, such as `- [Ada](ada.md)`. A
+reference, at any depth or a section item, reaching a URL, a binary target, another Room without
+`cross_room` or the context's `anyRoom: true`, or a Document of a type it does not name is
+`unresolved-reference`, and a frontmatter one reaching no member is `referenced-missing-file` naming
+its `referrers`; `attachment` admits any path, and no body link is a referrer. `validating` is
+`validate` over a `ReadonlyMap` Library, yielding each path read with the references it `resolved`;
+`assess(file, content, binary, model?, context?)` reads one member, naming its `type` and `targets`,
+and `verdicts()` `add`s assessments in any order, its `verdict()` the one `validate` gives.
+
+Validation returns acceptance or one refusal naming the file, the field or section, and the violated
+rule. `parse`, `apply`, and `serialize` operate on a single text Document; they do not validate
+collection membership or consume binary bytes. `apply(document, changes)` applies a list of changes
+in one pass however many: each sets one typed field, a list or an object as a whole included,
+replaces one existing section body or the preamble against its prior content, or rewrites one link
+destination or one reference, as the given Document holds it. All apply or none: one refused refuses
+all, and `{ outcome: 'stale', document, stale }` keeps the Document and lists the positions of each
+change it does not hold and of each later change of a part an earlier one changes, such as a link
+inside a replaced section. A section need not be declared by the Document's MD Type to be replaced.
+A section or preamble replacement that would hide a section the Document carried behind a repeated,
+merged or fenced heading is refused as `overlapping-section` naming that section. A field change
+does not reformat the rest of the file: an inline value is replaced where it stands, and a value
+written after the key's line, an empty one or a block mapping or list, is written after the key's
+colon in its place. A field holds a `FrontmatterValue`: null, a string, a boolean, a finite number,
+a list without holes or a plain object of them, no cycle, at most 63 deep. `frontmatterValue(value)`
+is that test; `apply` refuses any other value and `parse` a frontmatter past 64 deep as
+`frontmatter-fence`. `equalValue` compares two by value, whatever their key order, as `apply` does.
+
+`Document.preamble` is a readonly string holding the exact source after frontmatter and before the
+first section heading the section parser recognizes, `''` when empty, whitespace, newlines, a
+leading H1 and any fenced text before the first section included. The closing frontmatter
+delimiter and its line ending belong to frontmatter: `bodyAfterFrontmatter(content)` is the exact
+source after them, the whole text when it opens no closed fence, and the preamble begins it.
+`headingLines(content, from, to)` reads each ATX heading outside a fence, by `level` and `name`.
 
 ```ts
 const parsed = parse(content);
@@ -137,81 +143,88 @@ if (parsed.ok) {
     prior: parsed.document.preamble,
     value: 'Updated introduction.\n\n',
   };
-  const result = apply(parsed.document, change);
+  const result = apply(parsed.document, [change]);
   if (result.outcome === 'applied') serialize(result.document);
 }
 ```
 
-`apply(document, { kind: 'preamble', prior, value })` compares `prior` to the
-current preamble exactly. A mismatch returns `{ outcome: 'stale', document }`
-with the unchanged current Document. A match replaces only that source span,
-then parses a new immutable Document; the original is unchanged. Frontmatter
-and section source outside the span retain their exact bytes, including newline
-and heading syntax. As with section replacement, `value` is literal source:
-no newline or separator is inserted, and the resulting source is parsed again.
-An invalid frontmatter result receives the existing parse refusal. Independent
-field, section, and preamble edits keep their own prior values valid.
-
-For a moved link, use `links` and the `link` change below to obtain the rewritten Document, then
-compare its public `preamble` and `sections` to the original. Submit the changed values with their
-original priors. Consumers need no source span calculation or private Model import. Collection
-validation remains a separate call against the resulting Library.
+A `{ kind: 'preamble', prior, value }` change compares `prior` to the current preamble exactly. A
+mismatch is stale and leaves the current Document unchanged. A match replaces only that source span,
+then parses a new immutable Document; the original is unchanged. Frontmatter and section source
+outside the span retain their exact bytes, including newline and heading syntax. As with section
+replacement, `value` is literal source: no newline or separator is inserted, and the resulting
+source is parsed again. An invalid frontmatter result receives the existing parse refusal.
+Independent field, section, and preamble edits keep their own prior values valid.
+
+For moved links, use `links` and one `link` change per destination below, then compare the rewritten
+Document's `preamble` and `sections` to the original and submit the changed values with their
+original priors. Collection validation remains a separate call against the resulting Library.
 
 `links(document, file)` is the inventory of that Document's ordinary link and image destinations
 that resolve to a collection Path, each with the destination `href` exactly as written and the
 `target` Path it resolves to relative to `file`. A `?query` or `#fragment` the destination carries
 is part of `href` and not part of `target`. Entries are ordered by the position of the destination
 text a rewrite would change, and one destination written the same way appears once however many
-times it occurs. A destination that the syntax profile does not read as a link, one that is empty,
-a bare fragment, or carries a URL scheme, and one that leaves the collection are not in the
-inventory. Frontmatter field values, including typed references, are not link destinations and
-are never enumerated.
-
-`target(file, href)` is the collection Path one written destination resolves to relative to
-`file`, by the rule `links` and `validate` read: surrounding spaces and the `?query` or `#fragment`
-it carries are not part of that Path. A destination that is empty, a bare fragment, carries a URL
-scheme, starts at `/`, or leaves the collection resolves to `null`. It reads a destination written
-anywhere, such as a typed reference field value, which `links` never enumerates.
-`retarget(href, to)` is that destination rewritten to `to`, keeping the `?query` or `#fragment` it
-carried, which is the destination the `link` change below writes.
+times it occurs. A destination that the syntax profile does not read as a link, one that is empty, a
+bare fragment, or carries a URL scheme, and one that leaves the collection are not in the inventory.
+Frontmatter values are never link destinations.
+
+`target(file, href)` is the collection Path one written destination resolves to relative to `file`,
+by the rule `links` and `validate` read: surrounding spaces and the `?query` or `#fragment` it
+carries are not part of that Path. It reads backslash escapes and character references, strips as a
+URL does, then percent-encodings, so `plan\.md`, `pl&#97;n.md` and `pl%61n.md` all name `plan.md`;
+an escaped or referenced `?` or `#` starts the suffix. A destination that is empty, a bare fragment,
+carries a URL scheme, starts at `/` once read, encodes no UTF-8 text, or leaves the collection
+resolves to `null`. It reads a destination written anywhere, such as a typed reference field value,
+which `links` never enumerates. `roomTarget(href)` reads `/rooms/<handle>/<path>` as its `handle`
+and `path`, else `null`, and `roomPath(href)` is its canonical text, the target `facts` resolves for
+any reference written so. `retarget(href, to)` is that destination rewritten to `to`, keeping the
+`?query` or `#fragment` it carried, which is the destination the `link` change below writes. It
+percent-encodes in `to` what `target` would decode, so it resolves to `to` again.
+
+`checkDocument(file, content, model, context?)` reads one Document as `validate` does in that
+context: `ok: false` where it refuses, else its `type`, whether it is `typed`, and each reference's
+`target` and `to`, which `resolves(reference, typeOf, binary)` says reaches a member `validate`
+accepts. `sectionReferences(file, content, definition)` reads only its `list` sections' references;
+neither names another Room's, which `roomReferences(file, content, definition?)` lists by `kind`.
+`parseFrontmatter(content, file?)` is `parse` without the body: `content`, `type`, `fields`.
 
 `references(document, definition)` lists each frontmatter `reference` string at any depth as
-`pointer`, `value` and `to`; `apply(document, { kind: 'reference', pointer, value, prior })`
-rewrites one in place, keeping its quotes, or is `stale`. `relativize(document, file, handle,
-definition?)` rewrites each link destination and reference written `/rooms/<handle>/<path>` to
-`relative(file, path)`, keeping its `?query` or `#fragment`. A text without frontmatter, such as
-a captured body, is read from its first character: `relativizeBody(body, file, handle)` converts
-it so, refusing a destination left under `/rooms/` as `unresolved-link`; `bodyLinks(body, file)`
-is its `links`; `relinkBody(body, file, moves)` writes, in one pass, each destination whose Path
-is a key of `moves` as relative to its value, in the form the `link` change below writes.
+`pointer`, `value` and `to`; a `{ kind: 'reference', pointer, value, prior }` change rewrites one in
+place, keeping its quotes, or is stale. `relativize(document, file, handle, definition?, kept?)`
+rewrites, in one `apply`, each link destination and reference written `/rooms/<handle>/<path>` and
+not in `kept` to `relative(file, path)`, keeping its `?query` or `#fragment`. A text without
+frontmatter, such as a captured body, is read from its first character: `relativizeBody(body, file, handle)` converts it
+so, refusing a `/rooms/` one naming no Room's file as `unresolved-link`; `bodyLinks(body, file)`
+is its `links` and `bodyRoomLinks(body)` its other Rooms'. `relinkBody(body, file, moves)` writes,
+in the `link` change's form, each destination whose Path keys `moves` as that value relative to
+`file`, or as that value when it is a `/rooms/` path, and one whose `roomPath` does as that value.
 
 ```ts
 target('items/six.md', ' ../projects/atlas.md#overview '); // 'projects/atlas.md'
 retarget(' ../projects/atlas.md#overview ', '../archive/atlas.md'); // '../archive/atlas.md#overview'
 ```
 
-`apply(document, { kind: 'link', href, to })` rewrites every body occurrence of that written
-destination and changes nothing else, so a caller that moves a Path rewrites its referring
-Documents without a second Markdown parser. `to` is the Path the destination must resolve to
+A `{ kind: 'link', href, to }` change rewrites every body occurrence of that written destination
+and changes nothing else, so a caller that moves Paths rewrites each referring Document in one
+`apply` without a second Markdown parser. `to` is the Path the destination must resolve to
 instead; the `?query` or `#fragment` the written destination carried is not part of that Path and
-is kept. A Document that does not carry the destination is `stale`. A reference definition is
+is kept. A Document that does not carry the destination is stale. A reference definition is
 rewritten once however many uses name it. The destination is written in the form the syntax
 profile reads back: bare when it carries no whitespace, does not open with `<`, and balances its
 parentheses, otherwise inside angle brackets, which the original form is kept in when it still
 fits. A destination that neither form can carry is refused as `unwritable-link`.
 
-`project(document, file, context?)` is the Embedding projection input of one
-Document: the prose a consumer embeds, with every internal Markdown-document
-link reduced to the label a reader sees. A Document that declares no MD type
-contributes its body after the frontmatter fence. A typed one contributes its
-prose values, its preamble, and every H2 section its MD type does not declare as
-structure, each with its own heading line and every occurrence of a repeated
-name, in Document order: a declared text section, a list section of prose items
-and an undeclared section are in; a list section of any other item schema is
-not. A prose value is a frontmatter string at any depth whose schema is prose,
-projected bare and followed by a blank line, in the order the Document writes
-them, exactly as written. The Path and every other value stay outside, so
-changing them alone leaves the projection unchanged.
+`project(document, file, context?)` is the Embedding projection input of one Document: the prose a
+consumer embeds, with every internal Markdown-document link reduced to the label a reader sees. A
+Document that declares no MD type contributes its body after the frontmatter fence. A typed one
+contributes its prose values, its preamble, and every H2 section its MD type does not declare as
+structure, each with its own heading line and every occurrence of a repeated name, in Document
+order: a declared text section, a list section of prose items and an undeclared section are in; a
+list section of any other item schema is not. A prose value is a frontmatter string at any depth
+whose schema is prose, projected bare and followed by a blank line, in the order the Document writes
+them, exactly as written. The Path and every other value stay outside, so changing them alone leaves
+the projection unchanged.
 
 ```ts
 const parsed = parse(content);
@@ -224,36 +237,29 @@ if (parsed.ok) {
 }
 ```
 
-`ProjectionContext` carries `documentTargets`, the collection-relative Paths the
-caller holds as Markdown documents, and `definition`, the verified MD Type the
-Document's `type` names, taken from the model `verify` returned. A Document that
-declares a type the caller supplies no definition for is refused as
-`unknown-type`. `documentTargets` is what makes a destination internal: a
-destination resolving to a supplied Path is internal, and every other
-destination stays authored text, a binary target and a Path nobody
-supplied alike. The caller supplies that membership, as it does for `validate`;
-the model reads no bytes, resolves no authority, and classifies nothing it was
-not given.
-
-An internal link contributes its visible label and no destination, fragment, or
-title, in the inline `[label](path)`, full `[label][id]`, collapsed
-`[label][]`, and shortcut `[label]` forms, and the link reference definition an
-enumerated use resolves through leaves the projection with it. An image at a
-supplied document target contributes its alt text the same way, because a move
-rewrites that destination too. External URLs, binary targets,
-unsupplied destinations, bare fragments, code spans, fenced code, and escaped
-syntax stay exactly as authored, read through the same syntax profile `links`
-and `validate` read. `project` returns text and never rewrites the Document.
+`ProjectionContext` carries `documentTargets`, the collection-relative Paths the caller holds as
+Markdown documents, and `definition`, the verified MD Type the Document's `type` names, taken from
+the model `verify` returned. A Document that declares a type the caller supplies no definition for
+is refused as `unknown-type`. `documentTargets` is what makes a destination internal: a destination
+resolving to a supplied Path is internal, as is one into another Room, and every other stays text, a
+binary target and a Path nobody supplied alike. The caller supplies that membership, as it does for
+`validate`; the model reads no bytes, resolves no authority, and classifies nothing unsupplied.
+
+An internal link contributes its visible label and no destination, fragment, or title, in the inline
+`[label](path)`, full `[label][id]`, collapsed `[label][]`, and shortcut `[label]` forms, and the
+link reference definition an enumerated use resolves through leaves the projection with it. An image
+at a supplied document target contributes its alt text the same way, because a move rewrites that
+destination too. External URLs, binary targets, unsupplied destinations, bare fragments, code spans,
+fenced code, and escaped syntax stay exactly as authored, read through the same syntax profile
+`links` and `validate` read. `project` returns text and never rewrites the Document.
 
 A `link` change leaves the projection of every Document it rewrites
 byte-identical, while a changed visible label changes it.
 
-`divide(document, file, context)` divides that same Embedding projection input
-into ordered Spans under the budget the caller supplies. Each Span carries the
-`text` to embed and the `start` and `end` of the extent it addresses in the
-Document's own source, as UTF-8 byte offsets into that source and never string
-positions, so a caller stores the extent and hydrates the text it shows from
-the Document itself.
+`divide(document, file, context)` divides that same Embedding projection input into ordered Spans
+under the bound the caller supplies. Each Span carries the `text` to embed and the `start` and `end`
+of the extent it addresses in the Document's own source, as UTF-8 byte offsets and never string
+positions, so a caller stores the extent and hydrates the text it shows from the Document itself.
 
 ```ts
 const parsed = parse(content);
@@ -261,54 +267,45 @@ if (parsed.ok) {
   const divided = divide(parsed.document, 'tasks/two.md', {
     documentTargets: ['plans/q3.md', 'people/ana.md'],
     definition: model.types.task,
-    budget: 2000,
+    bound: 2000,
   });
   if (divided.ok) for (const span of divided.spans) embed(span.text);
 }
 ```
 
-`DivisionContext` is a `ProjectionContext` and one `budget`, the greatest number
-of characters a Span's embedded text may carry, counted the way JavaScript
-counts a string's length and not in the bytes the extent is measured in. No
-Span's text is ever longer than the budget. A bound cut never divides a
-character: where the budget ends between the two halves of a surrogate pair,
-the cut falls before the pair. A cut never falls inside link syntax either: it
-falls before an internal link, and a label longer than a whole piece's room is
-divided at the bound like any other text. The smallest budget is `2`, the
-length of one astral character; a budget that is not a whole number of at least
-`2` is refused with `division-budget` before any division. The budget is the
-caller's number alone: the model holds no tokenizer, counts nothing else, and
-refuses no input for being too long. A Document the caller supplies no
-definition for carries the same `unknown-type` refusal `project` returns.
-
-The division descends one level at a time, and only into the piece that does not
-fit: the whole input, a heading node, adjacent sibling nodes packed under one
-parent, adjacent blocks inside a node, sentences inside a block, and a bound cut
-inside a single sentence. A Span covers whole heading nodes or one contiguous
-piece inside a single node, never a partial node beside another. A Span that
-continues a node repeats that node's own heading line, projected like any other
-text and so carrying an internal link's label rather than its destination, and
-never an ancestor's heading, so renaming a heading costs what editing that
-heading's own body costs. A heading the budget cannot hold beside any text is
-not repeated, a continuation piece that begins with an astral character the
-room beside its heading cannot hold is carried without the heading, and a
-heading with no body of its own leads its first child node rather than standing
-as a Span of nothing but itself. Overlap exists at the
-sentence level alone, where each Span after the first repeats its predecessor's
-last sentence whenever the budget holds both; when it cannot hold both, the
-Span carries its own sentence alone. An untyped input arrives without a
-definition and descends the same ladder.
-
-Division is total: every authored character of the input belongs to at least one
-Span, and an input carrying no authored text divides into no Span at all. A
-Span's extent is the smallest range of the source that contains everything the
-Span covers, so it spans what the input excludes between two eligible pieces,
-and the extent addresses the authored source where the embedded text carries a
-repeated heading, an internal link's visible label, or a frontmatter value the
-author quoted. A Span divided out of one free-text field value addresses that
-whole value, because a YAML scalar's own positions are the only ones its source
-has, and a Span carrying part of one internal link's label addresses that whole
-link, because a cut never falls inside link syntax.
+`DivisionContext` is a `ProjectionContext` and one `bound`, the greatest number of characters a
+Span's embedded text may carry, counted the way JavaScript counts a string's length and not in the
+bytes the extent is measured in. No Span's text is ever longer than the bound. A bound cut never
+divides a character: where the bound ends between the two halves of a surrogate pair, the cut falls
+before the pair. A cut never falls inside link syntax either: it falls before an internal link, and
+a label longer than a whole piece's room is divided at the bound like any other text. The smallest
+bound is `2`, the length of one astral character; a bound that is not a whole number of at least
+`2` is refused with `division-bound` before any division. The bound is the caller's number alone:
+the model holds no tokenizer, counts nothing else, and refuses no input for being too long. A
+Document the caller supplies no definition for carries the `unknown-type` refusal of `project`.
+
+The division descends one level at a time, and only into the piece that does not fit: the whole
+input, a heading node, adjacent sibling nodes packed under one parent, adjacent blocks inside a
+node, sentences inside a block, and a bound cut inside a single sentence. A Span covers whole
+heading nodes or one contiguous piece inside a single node, never a partial node beside another. A
+Span that continues a node repeats that node's own heading line, projected like any other text and
+so carrying an internal link's label rather than its destination, and never an ancestor's heading,
+so renaming a heading costs what editing that heading's own body costs. A heading the bound cannot
+hold beside any text is not repeated, a continuation piece that begins with an astral character the
+room beside its heading cannot hold is carried without the heading, and a heading with no body of
+its own leads its first child node rather than standing as a Span of nothing but itself. Overlap
+exists at the sentence level alone, where each Span after the first repeats its predecessor's last
+sentence whenever the bound holds both; when it cannot hold both, the Span carries its own sentence
+alone. An untyped input arrives without a definition and descends the same ladder.
+
+Division is total: every authored character of the input belongs to at least one Span, and an input
+carrying no authored text divides into no Span at all. A Span's extent is the smallest range of the
+source that contains everything the Span covers, so it spans what the input excludes between two
+eligible pieces, and the extent addresses the authored source where the embedded text carries a
+repeated heading, an internal link's visible label, or a frontmatter value the author quoted. A Span
+divided out of one free-text field value addresses that whole value, because a YAML scalar's own
+positions are the only ones its source has, and a Span carrying part of one internal link's label
+addresses that whole link, because a cut never falls inside link syntax.
 
 ## Conditions
 
@@ -321,11 +318,11 @@ an object holds the facts of its members, and a value written `null` is `null`.
 not fit its kind, such as a string in a number field, a value outside its `enum` or a reference
 `resolve` maps to no `{ id, member? }`, is `unreadable`; a readable reference compares as that `id`.
 
-`holds(test, facts, owner)` is `true`, `false` or `undefined`, unknown; a Trigger fires on `true`
-alone. A test's `path` is `/path`, `/type`, `/sections/<heading>`, `/fields/<pointer>`, or
-`/members/<pointer>`, the member of the Document a reference field reaches. A pointer segment that
-is not an index reads every item of a list, an item lacking it unknown, and a test over several
-values holds as `any` does. The value `{ owner: true }` stands for `owner`, a User id.
+`holds(test, facts)` is `true`, `false` or `undefined`, unknown; a Trigger fires on `true` alone. A
+test's `path` is one of the `DOCUMENT_PATHS`, `/path` and `/type`, or `/sections/<heading>`,
+`/fields/<pointer>` or `/members/<pointer>`, the members a reference field's Documents bind, which
+`members(facts, path)` lists. A non-index pointer segment reads every item of a list, an item
+lacking it unknown, and a test over several values holds as `any` does.
 
 | The path reaches                        | `exists`    | `contains`     | any other test |
 | --------------------------------------- | ----------- | -------------- | -------------- |
@@ -343,21 +340,30 @@ and `in` are `undefined` against a value of another JSON type and compare a `tim
 `prefix` reads only a string, never a reference or a member; `gt`, `gte`, `lt` and `lte` order
 numbers, dates and times alone.
 
-`fires(condition, event, owner)` reads one change, `{ before, after, changed, by }`: the facts of
-each side, `null` on the side a creation or a deletion lacks, the `changed` paths whose prose
-changed, and the actor whose `/user`, `/site` and `/door`, one of `WRITE_DOORS`, a `by` test reads.
-`starts_matching` fires when the Document is in `scope` and holds `match` after the change and did
-not before; a creation fires only with `new_documents: true`, and a deletion never fires.
-`field_changes` fires when the side the Document has, after or for a deletion before, is in `scope`
-and holds `match`, and `field` changed: its facts differ, or with neither `from` nor `to` `changed`
-names it. A field absent on both sides never fires; `from: 'new'` admits only a creation and
-`to: 'deleted'` only a deletion, and listed values need that side to hold `field` `in` them.
-
-`itemFires(condition, item, owner)` compares `source` and `address` where given and holds `match`
-over `/delivery`, `/source`, `/provider`, `/container`, `/author`, `/has_attachment`,
-`/headers/<name>` and `/json/<pointer>`, unknown for an item with no JSON. `prosePaths(facts)` lists
-the prose and section paths whose text builds `changed`; `pointed(value, pointer)` reads a JSON
-Pointer, `undefined` where nothing is; `dated(value)` checks a `YYYY-MM-DD` date, and
-`instant(value)` reads an RFC 3339 time with its offset as epoch nanoseconds, a `bigint`.
-`datePointers(fields)` lists, sorted, the `/fields/<name>` pointers of a type's date fields, the
-only values a `DateCondition`'s `date` takes.
+`fires(condition, event)` reads one change, `{ before, after, changed, by, moved_between_rooms }`:
+each side's facts, `null` for a side the change lacks, the `changed` paths whose prose changed, the
+actor a `by` test reads at the `ACTOR_PATHS`, its `/door` one of `WRITE_DOORS`, and whether it
+moved between Rooms, which a set `moved_between_rooms` must match. `starts_matching` fires when the
+Document is in `scope` and holds `match` after the change but not before, a creation only with
+`new_documents: true`. `field_changes` fires when the Document's side, after or for a deletion
+before, is in `scope` and holds `match`, and `field` changed: its facts differ, or with neither
+`from` nor `to` `changed` names it, as it names `TEXT_FIELD`, `/text`, at any change of its text. A field on neither
+side never fires; `from: 'new'` admits only a creation and `to: 'deleted'` only a deletion, and listed values need that side's `field` `in` them.
+
+`itemFires(condition, item)` compares `source` and `address` where given and holds `match` over the
+`ITEM_PATHS`, a `/json` test unknown for an item with no JSON; `notificationFires(condition, notification)`
+holds `match`, where given, over a delivered notification's Document facts, and `messageFires(condition, message)`
+over a message's `MessageFacts` at the `MESSAGE_PATHS`, a `/routine` or `/path` test unknown where its request names none.
+`prosePaths(facts)` lists the prose and section paths whose text builds `changed`; `pointed(value, pointer)` reads a JSON Pointer,
+`undefined` where nothing is; `dated(value)` checks a `YYYY-MM-DD` date, and `instant(value)` reads an RFC 3339 time with its offset as epoch nanoseconds, a `bigint`.
+
+What a condition names of one MD Type: `datePointers(fields)` lists, sorted, its date fields as escaped
+`/fields/` pointers, the only `date`s a `DateCondition` takes; `changeFields(definition)` lists each
+declared field and section as a `field_changes` `field`, and `valuedChange(definition, field)` holds of
+a field read as one value, never prose, a list or reached through one, the only `field` that takes a
+`from` or a `to`; `memberPointers(definition, memberType)` lists the `/members/<name>` paths of the
+fields referring to `memberType`, alone or in a list. `testOperators(definitions, path, memberType)`
+lists the `CONDITION_OPERATORS` a test on a document `path` takes in any of those types, none where none
+declares it or, for a member path, `memberPointers` lacks it: `exists` alone on a section, prose or an
+object, `contains` and `exists` on a list, `prefix` on a string and `/path`, orderings on a number, date
+or time, and `contains` through a list. `actorOperators(path)`, `itemOperators(path)` and `messageOperators(path)` list the tests on a `by`, item or message path, `itemKind(path)` an item value's kind, or `null`, as on `/json/`.