@aburi/diff 0.3.0 → 0.5.0

package/dist/index.d.mts

@@ -1,19 +1,17 @@
-import { CallEdge, SerializeOptions } from "@aburi/core";
+import { CallEdge, IntegrityViolation, SerializeOptions } from "@aburi/core";
 import { Component, ComponentDiff, Dependency, DependencyDiff, DependencyEndpoint, DependencyUnknown, DiffResult, IR, IRRef, MatchRationale, NotComparedFile, RelativePath, Signature, SkipReason, SliceRecord, Summary, Symbol, SymbolChange, SymbolDelta, SymbolId } from "@aburi/types";
 
 //#region src/components.d.ts
 /**
- * `docs/design/diff-algorithm.md` §6.1 — Component diff. Assumes `components[].id` is unique on each side (ir-schema.md §14
- * #2) and does not check it: `buildDiff` establishes that before calling, and a caller
- * reaching this export directly owns the obligation, because the lookup map here is
- * last-write-wins and a repeat is lost rather than reported.
+ * `docs/design/diff-algorithm.md` — Component diff. Assumes `components[].id` is unique
+ * on each side (ir-schema.md #2) and does not check it: `buildDiff` establishes that, and
+ * a caller reaching this export directly owns the obligation, because the lookup map here is
+ * last-write-wins.
  *
- * Component identity is `id`; when a Component is present in both
- * base and head with the same id, the three delta booleans describe which axes actually
- * moved (roots reshuffle, public API globs, or framework hint list). `modified` deltas
- * are intentionally absent from this diff (diff-algorithm.md §5.2.3): fields are reported as before/after
- * pairs via the `changed[]` entries so consumers can render them without stringifying
- * arrays.
+ * Any field that differs makes a Component `changed` — the whole object is compared, not the
+ * three axes the delta names, so a `changed[]` entry with all three booleans `false` is a
+ * well-formed answer meaning "something else about this component moved". `modified` deltas
+ * are intentionally absent: fields are reported as before/after pairs.
  */
 declare function diffComponents(base: readonly Component[], head: readonly Component[]): ComponentDiff;
 /**
@@ -23,58 +21,40 @@ declare function diffComponents(base: readonly Component[], head: readonly Compo
  * `symbolFiles` is keyed on the endpoint id exactly as `dependencies[]` spells it, and its
  * values come from `symbols[].source.file` — the same space `stats.skippedFiles[].path` is in,
  * and the same space `buildDiff` classifies Symbols by. Reading the file out of the id's path
- * segment instead would introduce a second answer to "which file is this endpoint in", and
- * nothing in the schema forces the two to agree; a Symbol reported `unknown` while its edges
- * stayed `removed` is precisely the inconsistency this exists to remove.
- *
- * A Component endpoint is absent from the map, which is how it stays out of the reclassification
- * without a special case: a Component is an aggregate over roots and has no file to lose.
+ * segment instead would be a second answer to "which file is this endpoint in" that nothing
+ * forces to agree with the first. A Component endpoint is absent from the map, which keeps it
+ * out of the reclassification without a special case: an aggregate over roots has no file to
+ * lose.
  */
 interface DependencySideView {
   /**
-   * `source.file` of every Symbol this document holds, by the id an endpoint would name it
-   * with. Keyed by `DependencyEndpoint` rather than `SymbolId` because that is the question
-   * being asked: the lookup happens with an endpoint whose kind is not yet known, and "absent"
-   * is the answer for a Component id. Demanding the kind up front would move that decision to
-   * the caller, where it would need a second Symbol-id silhouette test to make.
+   * `source.file` of every Symbol this document holds, keyed by `DependencyEndpoint` rather
+   * than `SymbolId` because the lookup happens with an endpoint whose kind is not yet known,
+   * and "absent" is the answer for a Component id.
    */
   symbolFiles: ReadonlyMap<DependencyEndpoint, RelativePath>;
-  /**
-   * Files this document never analysed, by path, with the reason it gave. `RelativePath` on
-   * both sides of the file space is what states the bridge these two maps exist to make — one
-   * map's values are the other's keys — though the alias is unbranded, so it is intent rather
-   * than enforcement.
-   */
+  /** Files this document never analysed, by path, with the reason it gave. */
   lostFiles: ReadonlyMap<RelativePath, SkipReason>;
 }
 /**
- * Build a side view from a document. The only construction site there is.
- *
- * `buildDiff` reads `lostFiles` for its own Symbol classification too, from this same object,
- * so "a Symbol reported unknown and the edges it took with it cannot disagree about which file
- * went missing" is a property of the wiring rather than a claim in a comment.
- *
- * Exported because `DependencySideView` is public and `diffDependencies` requires one: without
- * a factory a caller would have to reproduce the `symbols[].source.file` keying and the
- * `stats.skippedFiles` read, and the likeliest outcome of that is a caller who gets it wrong.
+ * Build a side view from a document. Exported because `DependencySideView` is public and
+ * `diffDependencies` requires one; `buildDiff` reads `lostFiles` for its own Symbol
+ * classification from this same object.
  */
 declare function dependencySideView(ir: IR): DependencySideView;
 /**
- * `docs/design/diff-algorithm.md` §6.2 — Dependency diff. Identity is the composite
- * `(from, to, via)` triple; direction and effect changes are surfaced as an added + removed
- * pair so `modified` is not part of the schema (§6.2 tail). Uniqueness of that triple is the caller's obligation on the same
- * terms as `diffComponents` above — a repeat here is indistinguishable in the output from
- * the flip it encodes.
+ * `docs/design/diff-algorithm.md` — Dependency diff. Identity is the composite
+ * `(from, to, via)` triple; direction and effect changes surface as an added + removed pair so
+ * `modified` is not part of the schema. Uniqueness of the triple is the caller's
+ * obligation on the same terms as `diffComponents`.
  *
- * `sides` is what separates a deletion from a loss (`docs/design/diff-algorithm.md` §6.2.1).
- * It is required rather than optional: omitting it would classify every edge into a lost file
- * as a deletion again — silently, and while still writing `unknown: []`, which this schema
- * defines as "nothing was unknown" rather than "nobody looked". A caller with no skip list to
- * offer says so by passing a side view whose `lostFiles` is empty, which is the honest spelling
- * of an IR written before `stats.skippedFiles` existed.
+ * `sides` separates a deletion from a loss and is required rather than optional:
+ * omitting it would silently classify every edge into a lost file as a deletion while still
+ * writing `unknown: []`, which the schema defines as "nothing was unknown". A caller with no
+ * skip list passes a side view whose `lostFiles` is empty.
  *
- * The return type declares `unknown` present, where the schema leaves it optional for documents
- * that predate the field. A reader may find it missing; this function always writes it.
+ * The return type declares `unknown` present, where the schema leaves it optional for
+ * documents that predate the field.
  */
 declare function diffDependencies(base: readonly Dependency[], head: readonly Dependency[], sides: {
   base: DependencySideView;
@@ -89,18 +69,21 @@ declare const MAX_LINE_FUZZ = 10;
 declare const MIN_LINE_FUZZ = 0;
 interface DeltaOptions {
   /**
-   * §5.2.1 line fuzz for rule/call identity. Must be an integer in
+   * Line fuzz for pairing an edited rule/call/decorator with its predecessor
+   * (diff-algorithm.md); an unchanged one pairs however far it moved. Must be an integer in
    * `[MIN_LINE_FUZZ, MAX_LINE_FUZZ]` (`0..10`); anything outside — or a non-finite value —
-   * throws `DiffError({ code: "invalid-line-fuzz" })`. Setting `0` disables fuzz; omitting
-   * the field falls back to `DEFAULT_LINE_FUZZ` (2).
+   * throws `DiffError({ code: "invalid-line-fuzz" })`. Setting `0` pairs an edit only with an
+   * element on the same line, and does not stop an unchanged element from pairing; omitting the
+   * field falls back to `DEFAULT_LINE_FUZZ` (2).
    */
   lineFuzz?: number;
 }
 /**
- * §5 — Compute the full per-Symbol delta between two paired Symbols. The three axis
- * booleans (`apiChanged` / `logicChanged` / `syntaxChanged`) come from fingerprint
- * comparison; the array deltas (`rules` / `effects` / `calls` / `decorators`) come from
- * identity-preserving diff with configurable line fuzz.
+ * The full per-Symbol delta between two paired Symbols (diff-algorithm.md). The axis booleans
+ * come from fingerprint comparison; the array deltas from identity-preserving pairing, with line
+ * fuzz deciding only how far an edited element may sit from the one it replaced.
+ * `confidenceChanged` is always written, `false` included, so a reader can tell it from a diff
+ * that predates the field.
  */
 declare function computeSymbolDelta(base: Symbol, head: Symbol, options?: DeltaOptions): SymbolDelta;
 //#endregion
@@ -118,10 +101,10 @@ interface SymbolPair {
 /** Optional map returned by git for stage 2 (`old path` → `new path`). */
 type GitRenameMap = ReadonlyMap<string, string>;
 /**
- * §3.1 — segments Symbols with identical `id` into paired matches. Runs first because it
- * is the highest-confidence signal (no heuristics, just hash lookup).
+ * Stage 1 (diff-algorithm.md) — segments Symbols with identical `id` into paired matches.
+ * Runs first because it is the highest-confidence signal (no heuristics, just hash lookup).
  *
- * Assumes `id` is unique on each side (ir-schema.md §14 #1) and does not check it:
+ * Assumes `id` is unique on each side (ir-schema.md #1) and does not check it:
  * `buildDiff` establishes that before calling, and a caller reaching this export directly
  * owns the obligation. A repeat on the head side loses all but the last entry to the lookup
  * map and then removes every one of them from `remainingHead`; a repeat on the base side
@@ -134,10 +117,10 @@ declare function matchStageId(base: readonly Symbol[], head: readonly Symbol[]):
   remainingHead: Symbol[];
 };
 /**
- * §3.2 — git rename detection. When a rename map is available (typical of ref-driven
- * scans), rewrite the base id with the head-side file path and look for a hit. Two base
- * files renamed onto one target predict the same head id, so the claimants are collected
- * before one is chosen.
+ * Stage 2 (diff-algorithm.md) — git rename detection. When a rename map is available
+ * (typical of ref-driven scans), rewrite the base id with the head-side file path and look
+ * for a hit. Two base files renamed onto one target predict the same head id, so the
+ * claimants are collected before one is chosen.
  *
  * The rewriter only touches the `file` portion of the id — the `<language>:` prefix and
  * the trailing `#<qualified-name>` segment stay unchanged so a moved file with the same
@@ -150,14 +133,19 @@ declare function matchStageGitRename(remainingBase: readonly Symbol[], remaining
   remainingHead: Symbol[];
 };
 /**
- * §3.3 — group both sides by `fingerprint.logic` and pair within each group. Two branches:
+ * Stage 3 (diff-algorithm.md) — group both sides by kind and `fingerprint.logic` and pair
+ * within each group. Two branches:
  * - single base candidate → paired with `logic-fingerprint`, no similarity test
  * - several → `nameSimilarity` disambiguates at ≥ 0.85; a group that cannot reach it is left
  *   whole for stage 4 to re-evaluate
  *
- * Dropped symbols are excluded (§3.3 tail) — their logic fingerprint is the sentinel
+ * Dropped symbols are excluded — their logic fingerprint is the sentinel
  * `"000000000000"` and would collide with every other dropped Symbol in the workspace.
  * They flow to the stage-4.5 weak matcher instead.
+ *
+ * A logic value that names nothing (`logicNamesNothing`) is shared the same way, but by
+ * Symbols that are kept: its group forms and pairs, on names alone and without the first
+ * branch. See `pairWithinLogicGroup`.
  */
 declare function matchStageLogicFingerprint(remainingBase: readonly Symbol[], remainingHead: readonly Symbol[]): {
   matched: SymbolPair[];
@@ -165,39 +153,15 @@ declare function matchStageLogicFingerprint(remainingBase: readonly Symbol[], re
   remainingHead: Symbol[];
 };
 /**
- * §3.4 — name + signature similarity with `(kind, signatureNullness)` bucket pre-filter.
- * `signatureNullness` distinguishes symbols with no signature (`interface`, `type`,
- * `class`) from callable ones so that a class body is never paired against a function.
- *
- * §3.4.6's owner gate runs before the score: a pair whose owners are not the same class, or
- * one a rename of the other, is not the same Symbol however well the rest agrees, and no
- * weight small enough to leave the name axis meaning something can outvote a matching member
- * name and signature.
- *
- * The composite score, over the pairs that clear it:
+ * Stage 4 (diff-algorithm.md) — name + signature similarity with a `(kind, signatureNullness)`
+ * bucket pre-filter, so a class body is never paired against a function. The owner gate runs
+ * before the score; over the pairs that clear it:
  *   0.5 * memberSimilarity + 0.3 * signatureSimilarity + 0.2 * OWNER_AXIS_SATISFIED
- *
- * `memberSimilarity` reads the last segment, not the whole qualified name: the gate has
- * already settled the owner, and reading it again on the name axis charges for it twice —
- * which is what kept a renamed class from ever pairing (§3.4.6).
- *
- * §3.4.3 threshold table, applied per head by the token count of its last name segment:
- * - 1 token  → 1.0 (an exact match on both remaining axes)
- * - 2 tokens → 0.95 (strict, avoids `getUser` vs `getUsers`)
- * - default  → 0.85
- *
- * §3.4.3 also names two Symbols the stage does not read at all, because for them the score is
- * high on evidence it does not have rather than on a resemblance: one whose signature is
- * null, since `signatureSimilarity(null, null)` is 1.0 and the bucket key restricts it to
- * candidates that are equally signature-less; and one whose qualified name carries a single
- * distinct token, since that is the whole of what its identity amounts to. Neither can be
- * expressed as a threshold — the second wants a bar above the top of the scale.
- *
- * Every pairing in a bucket that clears its head’s threshold becomes a candidate, and the
- * candidates are settled in score order rather than one head at a time — see
- * `acceptInScoreOrder`. That doubles the pairings scored, from a bucket that shrank as heads
- * consumed it to the full bucket every time; `createNameScorer` more than pays for it by
- * tokenising each distinct name once per pass instead of once per comparison.
+ * `memberSimilarity` reads the last segment only, since the gate has settled the owner and
+ * reading it again would charge twice. Thresholds are per head, see `thresholdFor`; the two
+ * Symbols the stage does not read at all are explained at `saysEnoughToPair`. Candidates are
+ * settled in score order (`acceptInScoreOrder`), and `createNameScorer` pays for the full
+ * bucket being scored by tokenising each name once per pass.
  */
 declare function matchStageNameSignature(remainingBase: readonly Symbol[], remainingHead: readonly Symbol[]): {
   matched: SymbolPair[];
@@ -205,32 +169,18 @@ declare function matchStageNameSignature(remainingBase: readonly Symbol[], remai
   remainingHead: Symbol[];
 };
 /**
- * §3.4.5 — dropped-only weak matcher. For dropped Symbols the fingerprint is zeroed, so the
- * only signals left are the trailing segment of the qualified name and the file basename.
- * Either one alone is enough to pair, which is deliberately lax: §3.4.5 accepts a
- * false-positive risk because dropped Symbols sit outside the IR's main review surface.
+ * Stage 4.5 (diff-algorithm.md) — dropped-only weak matcher. With the fingerprint zeroed, the only
+ * signals left are the last segment of the qualified name and the file basename, and either alone
+ * pairs — deliberately lax, since dropped Symbols sit outside the main review surface.
  *
- * What it does not accept is a signal that identifies nothing. A basename hit on `index.ts`
- * — the most common filename in a TypeScript monorepo — paired every dropped Symbol of one
- * kind under one with every other, at a score they all tied on, and the results landed in
+ * What it does not accept is a signal that identifies nothing: a basename hit on `index.ts`
+ * paired every dropped Symbol of one kind with every other and landed them in
  * `summary.moved`, which `--fail-on moved` gates on. So a half counts only when exactly one
- * dropped base and one dropped head carry that key: such a key names a pair, where a key
- * several Symbols carry names a group, and a group is not a pairing.
- *
- * That makes the candidates the identifying keys themselves — at most one pairing each,
- * two axes, so at most `2 × min(base, head)` of them rather than the cross-product a shared
- * basename used to produce.
- *
- * "Exactly one" is counted over the Symbols this stage is handed, not over every dropped
- * Symbol in the Document — stages 1 and 2 have already taken theirs, and a key they emptied
- * out identifies again. That is the intended reading, since the question is which leftover a
- * key picks out, but it is also how a shared `index.ts` comes back: three dropped classes
- * under one, two of them matched by id, and the basename identifies the two that are left.
- *
- * It also leaves nothing for §3.4.5's 0.5-per-half scale to rank, so §3.8's sweep is the
- * wrong instrument: it settles conflicts by score, and with none to settle by it would drop
- * one identified pairing for another over nothing but the id it sorts under. `bestPairing`
- * takes a maximum matching instead.
+ * dropped base and one dropped head carry that key — counted over the Symbols this stage is
+ * handed, since a key the earlier stages emptied out identifies again. The candidates are
+ * then the identifying keys themselves, at most `2 × min(base, head)` of them, all worth the
+ * same; the score-ordered sweep would settle a conflict on nothing but id order, so
+ * `bestPairing` takes a maximum matching instead.
  */
 declare function matchStageDroppedWeak(remainingBase: readonly Symbol[], remainingHead: readonly Symbol[]): {
   matched: SymbolPair[];
@@ -240,12 +190,9 @@ declare function matchStageDroppedWeak(remainingBase: readonly Symbol[], remaini
 //#endregion
 //#region src/diff.d.ts
 /**
- * The two counters `buildDiff` always writes, narrowed off `Summary`'s optionals.
- *
- * They are optional on the wire so a diff written before they existed stays valid, and that
- * optionality is about *documents*, not about this function — a caller holding a value it just
- * built should not have to re-decide what "absent" would have meant. Same reason
- * `diffDependencies` declares its `unknown` array present.
+ * The two counters `buildDiff` always writes. They are optional on `Summary` only so a diff
+ * written before they existed stays valid; a caller holding a freshly built value should not
+ * have to re-decide what "absent" means.
  */
 interface UnknownCounters {
   unknown: number;
@@ -264,73 +211,64 @@ interface DiffInput {
   };
   /** Optional stage-2 rename table. When null/undefined stage 2 is skipped. */
   gitRenames?: GitRenameMap | null;
-  /** Passed through to computeSymbolDelta (§5.2.1 line fuzz). */
+  /** Passed through to computeSymbolDelta (line fuzz, diff-algorithm.md). */
   delta?: DeltaOptions;
 }
 /**
  * Top-level entry: run the 5-stage matcher, classify each pair, produce array deltas,
  * fold in Component / Dependency diffs, and assemble the `aburi.diff.v1` JSON projection.
- * The function is pure; write-to-disk is delegated to `writeCanonicalDiff` so callers can
- * pipe the result through additional steps (Markdown projection, `--fail-on` gate) before
- * serialisation.
+ * Pure; `writeCanonicalDiff` serialises, so callers can run the Markdown projection or the
+ * `--fail-on` gate over the result first.
  */
 declare function buildDiff(input: DiffInput): DiffResult & {
   notCompared: NotComparedFile[];
   summary: Summary & UnknownCounters;
 };
 /**
- * The files a document's scan never analysed, by path, with why.
- *
- * `stats.skippedFiles` is Class B: absent means the writer predates the field, not that
- * nothing was lost. An empty map is the honest answer either way — with no enumeration
- * there is no absence this function can explain, and guessing from
- * `totalFiles > parsedFiles` would attach a reason to whichever Symbols happened to be
- * missing. The CLI says so on stderr instead; a diff cannot invent the list.
- */
-/**
- * Byte-deterministic serialiser for a DiffResult. Delegates to `@aburi/core`
- * `serializeCanonical` so the sort order, NFC normalisation, and codepoint key sort are
- * shared with the IR side.
+ * Byte-deterministic serialiser for a DiffResult, sharing `@aburi/core`'s canonical sort
+ * order, NFC normalisation and key sort with the IR side.
  */
 declare function writeCanonicalDiff(diff: DiffResult, options?: SerializeOptions): string;
 //#endregion
 //#region src/errors.d.ts
+type DiffErrorCode = /** `base.$schema` and `head.$schema` disagree (diff-algorithm.md). */"schema-mismatch"
 /**
- * Coded error class for every @aburi/diff failure. Consumers branch on `code` without
- * parsing message text. Codes are stable and additions are non-breaking; renames are not.
+ * `DeltaOptions.lineFuzz` was not an integer in the documented [0, 10] range
+ * (diff-algorithm.md).
  */
-type DiffErrorCode = /** `base.$schema` and `head.$schema` disagree; a diff across schema versions cannot be trusted (§9.1). */"schema-mismatch" /** `config.diff.lineFuzz` was outside the documented [0, 10] range (diff-algorithm.md §5.2.1). */ | "invalid-line-fuzz"
+| "invalid-line-fuzz"
 /**
- * `baseIR` or `headIR` failed the top-level shape check that `buildDiff` enforces
- * before touching any Symbol. Missing / non-array `symbols`, `components`,
- * `dependencies`, or an absent `$schema` — anything that would otherwise crash with a
- * cryptic `TypeError: undefined is not iterable` deep in a matching stage.
+ * `baseIR` or `headIR` is not a Document of the shape `aburi.ir.v1` requires (invariant #20,
+ * `checkDocumentShape`), or its `$schema` is empty. Not the semantic invariants — see
+ * `assertDiffable`.
  */
 | "ir-shape-invalid"
 /**
  * `baseIR` or `headIR` repeats an identity the diff keys on: a `symbols[].id`, a
  * `components[].id`, or a `dependencies[]` `(from, to, via)` triple. Distinct from
  * `ir-shape-invalid` because the Document is well-formed — the collision changes the
- * answer rather than preventing one, and ir-schema.md §14 #1 / #2 / #13 already forbid it.
+ * answer rather than preventing one (ir-schema.md #1 / #2 / #13).
  */
 | "ir-identity-collision"
 /**
- * A `SliceRecord` broke the derivation invariant of slice-view.md §7.1 / §8.2:
- * `members[]` empty, `members[]` not in strictly ascending order, or `id` not
- * equal to `"slice:" + members[0]`. The last two compare one property against
- * another and so have no JSON Schema equivalent; the pass checks them itself
- * (§7.4). Raised only by an Aburi bug, never by user input — see
- * `SliceRecordViolation.kind` for which clause broke.
+ * A `SliceRecord` broke the derivation invariant of slice-view.md. Raised only by
+ * an Aburi bug, never by user input — see `SliceRecordViolation.kind` for which clause broke.
  */
 | "slice-invariant-violated";
 interface DiffErrorDetail {
   code: DiffErrorCode;
   /** Offending value (schema id, fuzz value, etc.) when applicable. */
   value?: string;
+  /**
+   * Populated only for `ir-shape-invalid`, one entry per breach of invariant #20, so a caller
+   * repairing a hand-assembled Document sees every breach rather than the first.
+   */
+  violations?: readonly IntegrityViolation[];
 }
 declare class DiffError extends Error {
   readonly code: DiffErrorCode;
   readonly value: string | undefined;
+  readonly violations: readonly IntegrityViolation[] | undefined;
   constructor(message: string, detail: DiffErrorDetail, options?: {
     cause?: unknown;
   });
@@ -338,117 +276,107 @@ declare class DiffError extends Error {
 //#endregion
 //#region src/signature.d.ts
 /**
- * §3.4.2 signatureSimilarity — averaged type-similarity across `inputs`, `outputs`, and
- * `throws`. Follows diff-algorithm.md §3.4.2:
- * - null + null    → 1.0 (both symbols are signature-less; nothing to compare)
- * - null + non-null (either side) → 0.0
- * - both non-null: mean of three subscores
- *
- * Subscores:
- * - inputs: ordered type-string equality rate (`i-th type equal` counts as 1, else 0)
- * - outputs: same shape, but outputs is an array of type strings
- * - throws: unordered set intersection over union
- *
- * Empty subarrays on both sides count as 1.0 (they are trivially equal); on one side only
- * they count as 0.0 (asymmetric information).
+ * signatureSimilarity (diff-algorithm.md) — mean of three subscores over `inputs` (ordered
+ * type equality rate), `outputs` (same) and `throws` (Jaccard). Two null signatures score
+ * 1.0, one null scores 0.0. Empty on both sides counts as 1.0, empty on one side as 0.0.
  */
 declare function signatureSimilarity(base: Signature | null | undefined, head: Signature | null | undefined): number;
 //#endregion
 //#region src/similarity.d.ts
 /**
  * String similarity utilities dedicated to the diff engine. Kept in one file so the
- * name/signature/owner formulas are auditable side by side against diff-algorithm.md §3.4.
+ * name/signature/owner formulas are auditable side by side against diff-algorithm.md.
  */
 /**
- * Tokenise a qualified name into normalised lowercase word segments. Splits on:
- * - camelCase / PascalCase boundaries (`fooBar` → `foo`, `Bar`)
- * - snake_case underscores (`foo_bar` → `foo`, `bar`)
- * - dotted namespaces (`ns.Class` → `ns`, `Class`)
- * - `::` static-method separator (`Class::method` → `Class`, `method`)
- * - runs of digits (`get2Users` → `get`, `2`, `Users`)
- *
- * The result is de-duplicated so `InvoiceService.createInvoice` collapses to
- * `["invoice", "service", "create"]` — Jaccard similarity relies on set semantics, so
- * repeated tokens must not double-count.
+ * Tokenise a qualified name into de-duplicated lowercase word segments, splitting on camel /
+ * Pascal boundaries, `_`, `.`, `::` and digit runs: `InvoiceService.createInvoice` →
+ * `["invoice", "service", "create"]`. De-duplicated because Jaccard relies on set semantics.
  *
- * **The camel boundary is ASCII.** `isCamelBoundary` compares code points against `a`–`z`,
- * `A`–`Z` and `0`–`9`, so a name written in a script with no ASCII case boundary and no
- * separator comes back whole: `获取用户信息` and `ユーザー情報を取得する` are one token each,
- * and so is `получитьПользователя` — its camel hump does not register. A separator still
- * splits (`ユーザー.取得` gives two), and a mixed name splits on its ASCII half.
+ * The input is normalised to NFC first, so a name spelled with combining marks tokenises as
+ * the name it is: `ガ` written U+30AB U+3099 is the same one token as U+30AC, and a Hangul
+ * syllable written as conjoining jamo is the same one token as its precomposed block.
  *
- * That makes the token count a poor measure of how much such a name says, which matters
- * wherever a count is read as a proxy for that — see §3.4.3's admissibility rule.
+ * The camel boundary is Unicode case, not the ASCII ranges, so `получитьПользователя` splits
+ * at its hump the way `getUser` does. What no case rule can split is a script that has no
+ * case at all: `ユーザー情報を取得する` and `获取用户信息` come back whole however much they
+ * say. That is a fact about the writing system rather than about the name, which is why the
+ * token count is not what diff-algorithm.md §3.4.3 asks how much a name says — see
+ * `nameEvidence`.
  */
 declare function tokenizeName(input: string): string[];
 /**
- * Jaccard similarity between two token multisets. Formula: |A ∩ B| / |A ∪ B|.
- * Empty on both sides returns 1.0 (both are "no tokens", which are equivalent for the
- * purpose of §3.4.1). Empty on one side only returns 0.0.
+ * How much a qualified name says, counted in words.
+ *
+ * The distinct-token count is this number for every name the tokeniser can segment, and every
+ * caller wanting *granularity* — how coarse a Jaccard over these tokens will be — should keep
+ * reading that count instead. This answers the different question diff-algorithm.md §3.4.3
+ * asks: could two unrelated Symbols carry this name by coincidence? One word could be carried
+ * twice; a phrase could not.
+ *
+ * A token of a cased or alphabetic script is one word, so it is worth 1 — `main`, `handler`
+ * and `initialize` alike, however long they run, because length is not what makes a word
+ * coincidental. A run of a script that writes no word boundary cannot be segmented, so it is
+ * counted instead: its characters divided by the longest a single word of that script runs,
+ * which is a floor on the number of words in it. `获取用户信息` is six Han characters, so it
+ * says at least two words and is admitted; `ハンドラー` is five kana, so it says at least
+ * one, which is what `handler` says and not enough to pair on.
+ *
+ * Characters are counted once each, the way tokens are — `取得取得` says what `取得` says,
+ * as `getGet` says what `get` says. Combining marks are not counted: after NFC they are the
+ * remainder of a character that did not compose, not a word.
+ *
+ * The number is a floor rather than a count, so it is fractional where a run is: `取得` is
+ * two thirds of a word, `getUser` is exactly 2.
  */
-declare function jaccard(a: readonly string[], b: readonly string[]): number;
+declare function nameEvidence(qname: string): number;
 /**
- * Jaccard over the tokens of two strings. Part of the module's surface rather than an
- * internal shortcut: the two formulas below go through a token table they share for one
- * matching pass, so this has no caller inside this file.
+ * Jaccard similarity |A ∩ B| / |A ∪ B| over two token lists. Empty on both sides is 1.0
+ * (both are "no tokens", equivalent for nameSimilarity); empty on one side only is 0.0.
  */
+declare function jaccard(a: readonly string[], b: readonly string[]): number;
+/** Jaccard over the tokens of two strings. Public surface only; nothing in this file calls it. */
 declare function jaccardTokens(a: string, b: string): number;
-/**
- * §3.4.1 — Jaccard over the tokens of the full qualified name. The last segment is not
- * given extra weight; the entire path contributes uniformly because the intersection is
- * what the score should reward.
- */
+/** The member name; read by the weak matcher and the threshold lookup. */
+declare function lastSegment(qname: string): string;
+/** Jaccard over the tokens of the full qualified name, every segment weighted alike. */
 declare function nameSimilarity(baseName: string, headName: string): number;
 /**
- * §3.4.1 — Jaccard over the tokens of the **last segment**: the member name with its owner
- * removed. What §3.4's composite reads, because §3.4.6 decides the owner separately and
- * reading it on both axes charges for it twice.
+ * Jaccard over the tokens of the **last segment** only. What stage 4's composite reads, because the
+ * owner gate decides the owner separately and reading it on both axes charges twice.
  */
 declare function memberSimilarity(baseName: string, headName: string): number;
 /**
- * §3.4.6 (R-8) — whether two Symbols are close enough in *scope* to be the same Symbol: the
- * same owner, or one whose owner was renamed. A gate rather than a score, because grading the
- * owner cannot do what R-8 asks. `UserRepo.findById` and `AdminRepo.findById` agree on their
- * member name and their signature, so a shared `Repo` token at weight 0.2 carries them to
- * 0.8667 against a 0.85 threshold — while `UsersRepository.findById`, which *is* the rename,
- * shares no owner token and scores 0.8000. The collision outscores the rename, and raising the
- * weight only moves both: at 0.3 two three-token class names sharing two tokens land on
- * exactly 0.85.
+ * The owner gate (diff-algorithm.md, R-8) — whether two Symbols are close enough in *scope* to be
+ * the same Symbol: the same owner, or one whose owner was renamed. A gate rather than a score,
+ * because grading the owner cannot do what R-8 asks. `UserRepo.findById` and `AdminRepo.findById`
+ * agree on their member name and their signature, so a shared `Repo` token at weight 0.2 carries
+ * them to 0.8667 against a 0.85 threshold — while `UsersRepository.findById`, which *is* the
+ * rename, shares no owner token and scores 0.8000. The collision outscores the rename, and
+ * raising the weight only moves both: at 0.3 two three-token class names sharing two tokens land
+ * on exactly 0.85.
  *
- * Owner extraction:
- * - `Class::method` → `Class`
- * - `A.B.C.method` → `A.B.C`
- * - `topLevel`     → `` (empty)
- *
- * Two empty owners are compatible: top-level Symbols share the one outer scope. One empty and
- * one not are never compatible — the two live at different depths.
- *
- * Otherwise the owners must correspond segment for segment, and within a segment every token
- * on each side must find a distinct partner on the other under `sameWord`. Both sides must be
- * covered, so `UserRepo` and `UserRepoV2` are two classes rather than one renamed — an added
- * token is as much evidence of a sibling as of a rename, and R-8's business is refusing the
- * collision.
+ * Two empty owners are compatible (top-level Symbols share the outer scope); one empty and one
+ * not never are. Otherwise the owners must correspond segment for segment, every token on each
+ * side finding a distinct partner under `sameWord`.
  */
 declare function ownersAreCompatible(baseName: string, headName: string): boolean;
-/** §3.4.1 tail note — used by drop-4.5 weak matcher and threshold lookup. */
-declare function lastSegment(qname: string): string;
 //#endregion
 //#region src/slice.d.ts
-/** The three inputs of the Slice View pass — docs/design/slice-view.md §3. */
+/** The three inputs of the Slice View pass — docs/design/slice-view.md. */
 interface SliceInput {
   /** SymbolChange records produced by `buildDiff` (pre-sort is not required). */
   changes: readonly SymbolChange[];
   /**
    * Resolved call edges from the base IR. Typically produced by
    * `reconstructCallEdgesFromIR(baseIR)` inside `buildDiff` — Slice View
-   * consumes only resolved edges (§5.4), never `Symbol.calls[]` directly.
+   * consumes only resolved edges, never `Symbol.calls[]` directly.
    */
   baseCallEdges: readonly CallEdge[];
   /** Resolved call edges from the head IR (same source rule as base). */
   headCallEdges: readonly CallEdge[];
 }
 /**
- * Slice View clustering pass — docs/design/slice-view.md §2, §4–§8.
+ * Slice View clustering pass — docs/design/slice-view.md.
  *
  * Groups the changed-Symbol set of a diff into weakly-connected components
  * over the union of base and head call edges. Emits a `SliceRecord[]` whose
@@ -459,16 +387,16 @@ interface SliceInput {
  * Determinism, idempotence, input-order insensitivity, and locality are all
  * guaranteed — see `computeWeaklyConnectedComponents` in `@aburi/core`.
  *
- * Returns `[]` when no SymbolChange is Node-eligible per §4.1; callers should
+ * Returns `[]` when no SymbolChange is Node-eligible; callers should
  * still serialise the empty array — the Markdown projection omits the section
- * (§9.4 / §12.5) but the JSON always emits the key so consumers never
- * distinguish "field absent" from "no slices" (§11.2).
+ * but the JSON always emits the key so consumers never
+ * distinguish "field absent" from "no slices".
  *
- * Every returned record has passed `assertSliceRecordInvariant` (§7.4).
+ * Every returned record has passed `assertSliceRecordInvariant`.
  */
 declare function computeSlices(input: SliceInput): SliceRecord[];
 /**
- * §7.1 — the anchor of a Slice: its lexicographically smallest member.
+ * The anchor of a Slice: its lexicographically smallest member.
  *
  * Answers from `members[0]`. `id` is derived from the anchor, so
  * reconstructing the anchor by stripping the `"slice:"` prefix is circular at
@@ -478,8 +406,8 @@ declare function computeSlices(input: SliceInput): SliceRecord[];
  * this; consumers that need a label keep using `id` directly.
  */
 declare function sliceAnchor(record: SliceRecord): SymbolId;
-/** Which clause of the §7.1 / §8.2 invariant a `SliceRecord` broke. */
-type SliceViolationKind = /** Not a `SliceRecord` at all: not an object, or `id` / `members` of the wrong type. */"malformed-shape" /** `members[]` has no entries, so the Slice has no anchor (§11.1). */ | "members-empty" /** `members[]` is not in strictly ascending order, so `members[0]` need not be the smallest (§8.2, §11.1). */ | "members-unordered" /** `id` is not `"slice:" + members[0]` (§7.1). */ | "id-not-derived" /** The anchor is itself in a reserved id namespace, so `id` would read as a doubled prefix (§7.5). */ | "anchor-in-reserved-namespace";
+/** Which clause of the slice-view.md derivation invariant a `SliceRecord` broke. */
+type SliceViolationKind = /** Not a `SliceRecord` at all: not an object, or `id` / `members` of the wrong type. */"malformed-shape" /** `members[]` has no entries, so the Slice has no anchor. */ | "members-empty" /** `members[]` is not in strictly ascending order, so `members[0]` need not be the smallest. */ | "members-unordered" /** `id` is not `"slice:" + members[0]`. */ | "id-not-derived" /** The anchor is itself in a reserved id namespace, so `id` would read as a doubled prefix. */ | "anchor-in-reserved-namespace";
 interface SliceRecordViolation {
   kind: SliceViolationKind;
   /** The offending record's `id` when it has a usable one, else a short stand-in. */
@@ -488,12 +416,12 @@ interface SliceRecordViolation {
   message: string;
 }
 /**
- * Report which clause of the §7.1 / §8.2 invariant a value breaks, or `null`
+ * Report which clause of the slice-view.md derivation invariant a value breaks, or `null`
  * when it is a well-formed `SliceRecord`. Non-throwing counterpart of
  * `assertSliceRecordInvariant`: the pass wants an exception, a schema
  * validator wants a verdict it can turn into its own error.
  *
- * Takes `unknown` rather than `SliceRecord` deliberately. §7.4 layer 2 points
+ * Takes `unknown` rather than `SliceRecord` deliberately. Enforcement layer 2 points
  * this at documents written by third-party or older producers — data that has
  * not been type-checked by definition — so anything that assumed a well-typed
  * argument would crash on exactly the input it exists to reject.
@@ -509,21 +437,21 @@ declare function assertSliceRecordInvariant(record: SliceRecord): void;
 //#region src/status.d.ts
 type SymbolStatus = "unchanged" | "moved" | "changed" | "moved+changed" | "dropped-toggled";
 /**
- * §4 — determine which of the five statuses a matched pair falls into. `dropped-toggled`
- * has absolute priority: whenever `dropped` differs between base and head, every other
- * signal is ignored because the underlying fingerprint change is a consequence of the
- * drop-rule flip, not a genuine code edit (§4.1 rationale).
+ * Which of the five statuses a matched pair falls into (diff-algorithm.md). `dropped-toggled`
+ * has absolute priority: a fingerprint change under a drop-rule flip is a consequence of the
+ * flip, not a code edit.
+ *
+ * `pathChanged` covers both file relocation and in-file rename: any id mismatch means the
+ * identifier moved, which DF9 ("method rename (same file, same logic) → moved") relies on.
+ * Stage-1 pairs match by exact id, so only stage-2+ pairings can take the id branch.
  *
- * "pathChanged" catches both physical file relocation AND in-file rename: any Symbol id
- * mismatch between base and head means the identifier moved, and DF9 ("method rename
- * (same file, same logic) → moved") relies on this. Stage-1 pairs cannot trigger the id branch
- * (they match by exact id), so this only affects stage-2+ pairings where the pair was
- * discovered via fingerprint or name/signature similarity.
+ * A confidence change is a change although no fingerprint reads it: the same code classified
+ * less surely is something a reviewer has to see, and `unchanged` is never reported. A dropped
+ * pair keeps its `unchanged`: a dropped Symbol is outside what the diff asks a reviewer to read.
  */
 declare function classifyStatus(base: Symbol, head: Symbol): SymbolStatus;
-/** Direction of a `dropped-toggled` transition (§7.1 SymbolDroppedToggled). */
+/** Direction of a `dropped-toggled` transition (SymbolDroppedToggled, diff-algorithm.md). */
 type DropDirection = "to-dropped" | "to-kept";
 declare function dropDirection(head: Symbol): DropDirection;
 //#endregion
-export { DEFAULT_LINE_FUZZ, type DeltaOptions, type DependencySideView, DiffError, type DiffErrorCode, type DiffErrorDetail, type DiffInput, type DropDirection, type GitRenameMap, MAX_LINE_FUZZ, MIN_LINE_FUZZ, type SliceInput, type SliceRecordViolation, type SliceViolationKind, type SymbolPair, type SymbolStatus, assertSliceRecordInvariant, buildDiff, classifyStatus, computeSlices, computeSymbolDelta, dependencySideView, diffComponents, diffDependencies, dropDirection, jaccard, jaccardTokens, lastSegment, matchStageDroppedWeak, matchStageGitRename, matchStageId, matchStageLogicFingerprint, matchStageNameSignature, memberSimilarity, nameSimilarity, ownersAreCompatible, signatureSimilarity, sliceAnchor, sliceRecordViolation, tokenizeName, writeCanonicalDiff };
-//# sourceMappingURL=index.d.mts.map
+export { DEFAULT_LINE_FUZZ, type DeltaOptions, type DependencySideView, DiffError, type DiffErrorCode, type DiffErrorDetail, type DiffInput, type DropDirection, type GitRenameMap, MAX_LINE_FUZZ, MIN_LINE_FUZZ, type SliceInput, type SliceRecordViolation, type SliceViolationKind, type SymbolPair, type SymbolStatus, assertSliceRecordInvariant, buildDiff, classifyStatus, computeSlices, computeSymbolDelta, dependencySideView, diffComponents, diffDependencies, dropDirection, jaccard, jaccardTokens, lastSegment, matchStageDroppedWeak, matchStageGitRename, matchStageId, matchStageLogicFingerprint, matchStageNameSignature, memberSimilarity, nameEvidence, nameSimilarity, ownersAreCompatible, signatureSimilarity, sliceAnchor, sliceRecordViolation, tokenizeName, writeCanonicalDiff };