punctilio 4.3.0 → 5.0.0

package/src/dashes.ts

@@ -1,24 +1,20 @@
-import { cachedRegExp, DEFAULT_SEPARATOR, getEscapedSeparator, LATIN_LETTERS, UNICODE_SYMBOLS, wordBoundaryEnd, wordBoundaryStart } from "./constants.js"
+import { cachedRegExp, LATIN_LETTERS, NBSP_CHARS, SPACE_CHARS, UNICODE_SYMBOLS } from "./constants.js"
+import { boundaryCountAt, overInput, type ProseView, replaceAllInView, type ReplaceAllOptions } from "./prose-view.js"
 import { namedGroups } from "./utils.js"
 
 export const DASH_STYLES = ["american", "british", "none"] as const
 export type DashStyle = (typeof DASH_STYLES)[number]
 
 export interface DashOptions {
-  /** Boundary marker for HTML element boundaries. Default: "\uE000\uE001" */
-  separator?: string
   /** "american" (unspaced em), "british" (spaced en), "none". Default: "american" */
   dashStyle?: DashStyle
 }
 
-const { EN_DASH, EM_DASH, MINUS, MULTIPLICATION, PLUS_MINUS, SUPERSCRIPT_ST, SUPERSCRIPT_ND, SUPERSCRIPT_RD, SUPERSCRIPT_TH, LEFT_DOUBLE_QUOTE, RIGHT_DOUBLE_QUOTE, LEFT_SINGLE_QUOTE, RIGHT_SINGLE_QUOTE } = UNICODE_SYMBOLS
-
-// Characters a later pass produces from a trailing *word* character: the symbol
-// pass turns "x" into "×", and the superscript pass turns the ordinal letters
-// "st"/"nd"/"rd"/"th" into superscripts. Before conversion those word chars block
-// a range via the trailing word boundary ("1-55x5", "5--1st"); rejecting their
-// non-word replacements keeps a second pass from en-dashing the same range.
-const SYMBOL_PASS_TRAILING = `${MULTIPLICATION}${SUPERSCRIPT_ST}${SUPERSCRIPT_ND}${SUPERSCRIPT_RD}${SUPERSCRIPT_TH}`
+const {
+  EN_DASH, EM_DASH, MINUS, MULTIPLICATION, PLUS_MINUS,
+  LEFT_DOUBLE_QUOTE, RIGHT_DOUBLE_QUOTE, LEFT_SINGLE_QUOTE, RIGHT_SINGLE_QUOTE,
+  SUPERSCRIPT_ST, SUPERSCRIPT_ND, SUPERSCRIPT_RD, SUPERSCRIPT_TH,
+} = UNICODE_SYMBOLS
 
 // Prevents false-positive ranges in model names like "Llama-2-7B".
 export const numberRangeDisallowedPrefixes = ["-", EN_DASH, EM_DASH, MINUS] as const
@@ -38,241 +34,1245 @@ const months: readonly string[] = [
 
 const monthPattern = months.join("|")
 
+const WORD_CHAR_RE = /\w/
+const LATIN_LETTER_RE = new RegExp(`[${LATIN_LETTERS}]`)
+
+/**
+ * Glyphs the symbol passes fold word characters into (multiplication `x` →
+ * `×`, ordinal suffixes → superscripts). The dash rules run before those
+ * passes, so their word-character tests must treat a folded glyph like the
+ * word character it folds from — otherwise a range blocked at `6x3` converts
+ * once a re-run sees `6×3`.
+ */
+const FOLDED_WORD_CHARS = new Set<string>([
+  MULTIPLICATION,
+  ...SUPERSCRIPT_ST, ...SUPERSCRIPT_ND, ...SUPERSCRIPT_RD, ...SUPERSCRIPT_TH,
+])
+
+/** `\w` extended with {@link FOLDED_WORD_CHARS} for fold-stable `\b` checks. */
+function isWordChar(ch: string | undefined): boolean {
+  return ch !== undefined && (WORD_CHAR_RE.test(ch) || FOLDED_WORD_CHARS.has(ch))
+}
+
+/**
+ * Glyphs the math-symbol pass folds two-character operators into (`=~` → `≈`,
+ * `!=` → `≠`, ...). The fold tolerates one node boundary at its junction, so
+ * it can consume the character that sat between a boundary and a dash.
+ */
+const FOLDED_MATH_OPERATORS = new Set<string>([
+  UNICODE_SYMBOLS.APPROXIMATE,
+  UNICODE_SYMBOLS.GREATER_EQUAL,
+  UNICODE_SYMBOLS.LESS_EQUAL,
+  UNICODE_SYMBOLS.NOT_EQUAL,
+  UNICODE_SYMBOLS.PLUS_MINUS,
+])
+
+/**
+ * Glyphs the symbol passes fold from a source ending in a dash char (`<-` →
+ * `←`, `+-` → `±`). Where a trailing hyphen blocks a dash rule, its folded
+ * glyph must keep blocking, or the rule fires only on the re-run.
+ */
+const FOLDED_FROM_TRAILING_DASH = new Set<string>([
+  UNICODE_SYMBOLS.ARROW_LEFT,
+  UNICODE_SYMBOLS.PLUS_MINUS,
+])
+
+/** Mirror of {@link FOLDED_FROM_TRAILING_DASH} for sources starting with a
+ * dash char (`->` → `→`). */
+const FOLDED_FROM_LEADING_DASH = new Set<string>([UNICODE_SYMBOLS.ARROW_RIGHT])
+
+/**
+ * Runs one boundary-gated regex pass over the view and commits it immediately.
+ * Committing after every pass sequences the passes so a later pass sees the
+ * earlier edits and never queues an overlapping edit.
+ */
+function pass(
+  view: ProseView,
+  regex: RegExp,
+  replacer: (match: RegExpExecArray, view: ProseView) => string | null,
+  options?: ReplaceAllOptions,
+): void {
+  replaceAllInView(view, regex, replacer, options)
+  view.commit()
+}
+
+// ---------------------------------------------------------------------------
+// Boundary-tolerance helpers over a ProseView
+//
+// The boundary-tolerance positions throughout this module reproduce the
+// element-boundary semantics of the pre-v5 sentinel-marked pipeline; they are
+// pinned by the golden corpus and the migration's differential fuzz. Over a
+// clean ProseView, `replaceAllInView` skips any match with an interior boundary
+// unless `allowBoundaries` opts in. Each pass below builds an `allowBoundaries`
+// callback that admits a match only when every interior boundary sits at a
+// tolerated slot offset.
+// ---------------------------------------------------------------------------
+
+/** Interior boundary offsets of a match, strictly inside (matchStart, matchEnd). */
+function interiorBoundaryOffsets(match: RegExpExecArray, view: ProseView): number[] {
+  const start = match.index
+  const end = match.index + match[0].length
+  const offsets: number[] = []
+  for (const boundary of view.boundaries) {
+    if (boundary > start && boundary < end) offsets.push(boundary)
+    else if (boundary >= end) break
+  }
+  return offsets
+}
+
+/**
+ * True iff every interior boundary of the match lands on a tolerated offset,
+ * with no offset carrying more boundaries than its budget. A tolerated slot
+ * admits exactly one boundary; two stacked boundaries (an empty fragment)
+ * exceed the single-boundary budget and block the match.
+ */
+function interiorBoundariesAllowed(
+  match: RegExpExecArray,
+  view: ProseView,
+  tolerated: Map<number, number>,
+): boolean {
+  const counts = new Map<number, number>()
+  for (const offset of interiorBoundaryOffsets(match, view)) {
+    counts.set(offset, (counts.get(offset) ?? 0) + 1)
+  }
+  for (const [offset, count] of counts) {
+    const budget = tolerated.get(offset)
+    if (budget === undefined || count > budget) return false
+  }
+  return true
+}
+
+/** Separators tolerated by the word-boundary helpers (mirrors constants.ts). */
+const MAX_BOUNDARY_SEPARATORS = 3
+
+/**
+ * Reproduces `wordBoundaryStart` = `(?<!\w${sep}{0,3})\b` at the marked position
+ * `matchStart` (the start char is a word char — a digit). The `\b` requires a
+ * non-word char to its left (a separator counts as non-word); the negative
+ * lookbehind additionally blocks when a word char sits within
+ * {@link MAX_BOUNDARY_SEPARATORS} stacked boundaries to the left.
+ */
+function wordBoundaryStartOk(view: ProseView, matchStart: number): boolean {
+  if (matchStart === 0) return true
+  const nb = boundaryCountAt(view, matchStart)
+  const prevWord = isWordChar(view.text[matchStart - 1])
+  if (nb === 0) return !prevWord // `\b` needs a non-word char immediately left
+  // A separator sits left (non-word) so `\b` holds; the lookbehind blocks only
+  // when a word char is within ≤MAX boundaries.
+  return !prevWord || nb > MAX_BOUNDARY_SEPARATORS
+}
+
+/**
+ * Reproduces `wordBoundaryEnd` = `\b(?!${sep}{0,3}\w)` at the marked position
+ * `pos`. The `\b` requires a word/non-word transition (a separator counts as
+ * non-word); the negative lookahead blocks when a word char sits within
+ * {@link MAX_BOUNDARY_SEPARATORS} stacked boundaries to the right.
+ */
+function wordBoundaryEndOk(view: ProseView, pos: number): boolean {
+  const text = view.text
+  // `wordBoundaryEndOk` is only ever called at the end of a digit, letter, or
+  // suffix run, so `pos` is always ≥ 1 and `text[pos - 1]` is a real char.
+  const leftWord = isWordChar(text[pos - 1])
+  // A multiplication sign one space to the right blocks like the word char
+  // it folds from: the multiplication pass renders `5X 8` as `5 × 8`, and a
+  // range blocked at `3-5X` must stay blocked at `3-5 ×`.
+  if (text[pos] !== undefined && SPACE_CHAR_RE.test(text[pos]) && text[pos + 1] === MULTIPLICATION) {
+    return false
+  }
+  const nb = boundaryCountAt(view, pos)
+  const nextCleanWord = isWordChar(text[pos])
+  // The marked char right after pos is a separator when nb > 0 (non-word).
+  const rightWord = nb > 0 ? false : nextCleanWord
+  if (leftWord === rightWord) return false // no `\b`
+  // `(?!${sep}{0,3}\w)`: a word char within ≤3 boundaries to the right blocks.
+  return !(nb <= MAX_BOUNDARY_SEPARATORS && nextCleanWord)
+}
+
+// ---------------------------------------------------------------------------
+// Number ranges
+// ---------------------------------------------------------------------------
+
+const CURRENCY_RE = /[$€£¥₹]/
+
 /** Convert number ranges to en-dash (e.g., "1-5" → "1–5"). */
-export function enDashNumberRange(text: string, options: DashOptions = {}): string {
-  const chr = getEscapedSeparator(options)
-  const wb = wordBoundaryStart(chr)
-  const wbe = wordBoundaryEnd(chr)
-
-  // Common currency symbols for price ranges
-  const currencies = "$€£¥₹"
-
-  // Build positive range pattern from readable components
-  const phoneAreaCode = `(?<precedingAreaCode>\\d{3}-|\\(\\d{3}\\) ?)?`  // 555- or (555)
-  // Lookbehind prevents matching after dashes, so Llama-2-7B and +44-20 don't en-dash.
-  // `${MULTIPLICATION}${PLUS_MINUS}` keep ranges stable across the symbol pass, which
-  // runs after this one: a range start preceded by `x` (a Latin letter) or `+`/`-` is
-  // already rejected here, and the symbol pass rewrites "5x10" → "5×10" and "+/-1" →
-  // "±1". Rejecting `×` and `±` too means a second pass over those outputs doesn't
-  // suddenly en-dash a tail the first pass left alone (e.g. "5x10-20" → "5×10-20").
-  const notAfterDash = `(?<![${DISALLOWED_PREFIX_CLASS_FRAGMENT}${LATIN_LETTERS}${MULTIPLICATION}${PLUS_MINUS}.+])`
-  const rangeStart = `(?<start>(?:p\\.?|[${currencies}])?\\d[\\d.,]*${chr}?)`  // p.10, $100, 1,000
-  const rangeEnd = `(?<end>${chr}?[${currencies}]?\\d[\\d.,]*)`          // 20, $200, 2,000
-  const moreSegments = `(?<following>(?:${chr}?[-${MINUS}]${chr}?\\d+)*)` // -4567 in phone numbers
-  const unitSuffix = `(?<suffix>${chr}?(?:[AaPp][Mm]|[xKBTM]))?`         // am/pm, K/M/B
-
-  // Positive ranges: 1-5, 5--10, $100-$200, €5-€10, p.10-15
-  const positiveRangePattern = [
-    phoneAreaCode, wb, notAfterDash, rangeStart, "-{1,3}", rangeEnd, moreSegments, unitSuffix, wbe
-  ].join("")
-
-  text = text.replace(
-    cachedRegExp(positiveRangePattern, "g"),
-    (...args) => {
-      const groups = namedGroups<{
-        precedingAreaCode?: string
-        start: string
-        end: string
-        following?: string
-        suffix?: string
-      }>(args)
-      if (groups.following) return args[0] as string
-      // A range whose end abuts a symbol-pass replacement (× or a superscript
-      // ordinal) is an operand, not a range (e.g. "1-55×5", "5--1ˢᵗ"). This
-      // mirrors `notAfterDash` on the trailing side; done in the callback rather
-      // than as a regex lookahead to keep the composed pattern ReDoS-safe.
-      const matchEnd = (args.at(-3) as number) + (args[0] as string).length
-      if (cachedRegExp(`^${chr}?[${SYMBOL_PASS_TRAILING}]`, "").test((args.at(-2) as string).slice(matchEnd))) {
-        return args[0] as string
+export function enDashNumberRange(input: string): string
+export function enDashNumberRange(input: ProseView): void
+export function enDashNumberRange(input: string | ProseView): string | void {
+  return overInput(input, (view) => {
+    convertPositiveRanges(view)
+    convertNegativeRanges(view)
+  })
+}
+
+/**
+ * The positive-range shape is `phoneAreaCode? \b (?:p\.?|[currency])?\d[\d.,]*
+ * -{1,3} [currency]?\d[\d.,]* (?:[-−]\d+)* (?:[AaPp][Mm]|[xKBTMCF])? \b`, scanned
+ * left to right. The consume/advance behaviour determines which sub-ranges a
+ * later pass sees. The scan below realizes it: at each offset it attempts the
+ * structural match (boundary-aware), consuming its span on success (whether or
+ * not the dash converts) and advancing one character otherwise. So a multi-segment
+ * number whose head fails to form a structural match (e.g. boundary-broken
+ * "2024-01-15") still exposes a later sub-range ("01-15"), while a structurally
+ * matching but non-converting number hides its inner dashes.
+ */
+function convertPositiveRanges(view: ProseView): void {
+  const text = view.text
+  let i = 0
+  while (i < text.length) {
+    const span = matchPositiveRangeAt(view, i)
+    i = span === null ? i + 1 : Math.max(span, i + 1)
+  }
+  view.commit()
+}
+
+/**
+ * Attempts the `phoneAreaCode? \b rangeStart -{1,3} rangeEnd following* suffix?
+ * \b` structural match beginning at clean offset `i`, converting the dash when
+ * the gates pass. Returns the consumed end offset on a structural match (so the
+ * scan skips it), or null when no structural match begins at `i`.
+ */
+function matchPositiveRangeAt(view: ProseView, i: number): number | null {
+  // Optional phoneAreaCode (`\d{3}-` or `(\d{3}) `); when it ends in a dash that
+  // dash can instead be the range dash, so both interpretations are tried.
+  // `rangeStart` carries no leading slot, so a boundary right where it would
+  // begin (just past the area code) invalidates the area-code reading: fall back
+  // to treating the area-code digits as the range start.
+  const areaCode = readPhoneAreaCode(view, i)
+  const startStarts = areaCode === null || view.hasBoundary(i + areaCode) ? [i] : [i + areaCode, i]
+  for (const startStart of startStarts) {
+    const span = matchRangeBody(view, startStart)
+    if (span !== null) return span
+  }
+  return null
+}
+
+const THREE_DIGITS_RE = /^\d\d\d/
+const THREE_DIGITS_IN_PARENS_RE = /^\(\d\d\d\)/
+
+/** `phoneAreaCode` = `\d{3}-|\(\d{3}\) ?` at `i`; returns its length or null. */
+function readPhoneAreaCode(view: ProseView, i: number): number | null {
+  const text = view.text
+  // `\d{3}-` with no boundary inside the digits or before the dash.
+  if (THREE_DIGITS_RE.test(text.slice(i, i + 3)) && text[i + 3] === "-"
+    && !view.hasBoundary(i + 1) && !view.hasBoundary(i + 2) && !view.hasBoundary(i + 3)) {
+    return 4
+  }
+  // `(\d{3})` with an optional trailing space.
+  if (THREE_DIGITS_IN_PARENS_RE.test(text.slice(i, i + 5))) {
+    return text[i + 5] === " " ? 6 : 5
+  }
+  return null
+}
+
+/**
+ * Matches `\b rangeStart -{1,3} rangeEnd following* suffix? \b` boundary-aware,
+ * beginning the start at `startStart`. Converts the dash on a converting match.
+ * Returns the consumed end offset (structural span end) or null.
+ */
+function matchRangeBody(view: ProseView, startStart: number): number | null {
+  const text = view.text
+  // rangeStart = `(?:p\.?|[currency])?\d[\d.,]*`, anchored at a `\b`.
+  const start = readStartRun(view, startStart)
+  if (start === null) return null
+  const startSpanEnd = startStart + start.length
+  // One slot between start and the dash tolerates a single boundary there.
+  if (boundaryCountAt(view, startSpanEnd) > 1) return null
+  // Dash run of 1..3 hyphens that never spans an interior boundary (a boundary
+  // between two hyphens truncates the run).
+  if (text[startSpanEnd] !== "-") return null
+  let dashEnd = startSpanEnd + 1
+  while (text[dashEnd] === "-" && dashEnd - startSpanEnd < 3 && !view.hasBoundary(dashEnd)) dashEnd++
+  if (boundaryCountAt(view, dashEnd) > 1) return null
+
+  // `readStartRun` already stops at the first interior boundary, so the start
+  // span is boundary-free and its clean digits are the whole run.
+  const startDigits = text.slice(startStart, startSpanEnd)
+  if (!wordBoundaryStartOk(view, startStart)) return null
+  if (!notAfterDashOk(view, startStart)) return null
+
+  const end = readEndRun(view, dashEnd)
+  if (end === null) return null
+  // The greedy end `[\d.,]*` and greedy `following*` backtrack together; the
+  // structural match commits at the longest end whose tail completes.
+  for (let len = end.length; len >= end.minLen; len--) {
+    const endPos = end.firstDigitStart + len
+    const tail = resolvePositiveTail(view, endPos)
+    if (tail === null) continue
+    const endStr = text.slice(dashEnd, endPos)
+    if (tail.followingEmpty && rangeNumbersConvert(startDigits, endStr)) {
+      view.replace(startSpanEnd, dashEnd, EN_DASH)
+    }
+    return tail.consumedEnd
+  }
+  return null
+}
+
+/**
+ * The `(?:p\.?|[currency])?\d[\d.,]*` start run, boundary-truncated. The `wb`
+ * (`(?<!\w)\b`) anchors the start at a word boundary, so the currency
+ * alternative is unreachable: a leading currency symbol is non-word, and the only
+ * `\b` sits after it, at the digit — so the currency is never part of `start`
+ * (e.g. "$100-2000" reads start "100", which the 3+4 phone gate then blocks). The
+ * `p`/`p.` prefix is a word char, so it can begin the run at a word boundary.
+ */
+function readStartRun(view: ProseView, pos: number): { length: number } | null {
+  const text = view.text
+  let i = pos
+  if (text[i] === "p") {
+    // `p\.?\d` is contiguous (no interior slot), so a boundary between `p`, an
+    // optional `.`, and the first digit breaks the prefix.
+    if (view.hasBoundary(i + 1)) return null
+    i++
+    if (text[i] === ".") {
+      if (view.hasBoundary(i + 1)) return null
+      i++
+    }
+  }
+  // `\d` is required after the optional `p\.?` prefix. The prefix chars are
+  // never digits, so a missing digit here means no start run begins.
+  if (!/\d/.test(text[i] ?? "")) return null
+  i++
+  while (/[\d.,]/.test(text[i] ?? "") && !view.hasBoundary(i)) i++
+  return { length: i - pos }
+}
+
+interface EndRun {
+  /** Clean offset of the first end digit (past an optional currency symbol). */
+  firstDigitStart: number
+  /** Length of the `[\d.,]` run from `firstDigitStart`, boundary-truncated. */
+  length: number
+  /** Minimum end length: 1 digit (`\d` is required). */
+  minLen: number
+}
+
+/** The `[currency]?\d[\d.,]*` end run starting at `pos`, boundary-truncated. */
+function readEndRun(view: ProseView, pos: number): EndRun | null {
+  const text = view.text
+  let i = pos
+  if (CURRENCY_RE.test(text[i] ?? "")) {
+    // No slot between the currency symbol and the digit.
+    if (view.hasBoundary(i + 1)) return null
+    i++
+  }
+  if (!/\d/.test(text[i] ?? "")) return null
+  const firstDigitStart = i
+  i++
+  while (/[\d.,]/.test(text[i] ?? "") && !view.hasBoundary(i)) i++
+  return { firstDigitStart, length: i - firstDigitStart, minLen: 1 }
+}
+
+interface ResolvedTail {
+  followingEmpty: boolean
+  /** End offset of the whole structural tail (following + optional suffix). */
+  consumedEnd: number
+}
+
+/**
+ * Resolves the tail at `endPos`: finds the longest `following` whose
+ * `suffix? wbe` completes (the greedy choice). `followingEmpty` reports whether
+ * that following is empty (the convert condition); `consumedEnd` is the tail's
+ * end including any suffix (for the scan's consume/advance). Returns null when
+ * no following completes.
+ */
+function resolvePositiveTail(view: ProseView, endPos: number): ResolvedTail | null {
+  const candidates = followingCandidates(view, endPos, true)
+  for (let i = candidates.length - 1; i >= 0; i--) {
+    const suffixEnd = suffixWbeEnd(view, candidates[i], true)
+    if (suffixEnd !== null) {
+      return { followingEmpty: i === 0, consumedEnd: suffixEnd }
+    }
+  }
+  return null
+}
+
+
+/**
+ * Negative ranges: −5-5 → −5–5, −5--2 → −5–−2. A separate hand scan (not a regex)
+ * because the end run (`\d[\d.,]*`, with one tolerated boundary at its head)
+ * must truncate at node boundaries before its tail is resolved: a regex would
+ * fuse boundary-split digits into one run (e.g. "8{b}{b}{b}{b}{b}0" reads as
+ * "80") and mis-place or miss the conversion. The scan mirrors `matchRangeBody`:
+ * at each MINUS-led start it resolves the longest end whose `following* suffix?
+ * wbe` tail completes, converting only when that tail leaves `following` empty.
+ */
+function convertNegativeRanges(view: ProseView): void {
+  const text = view.text
+  let i = 0
+  while (i < text.length) {
+    if (text[i] === MINUS) {
+      const span = matchNegativeRangeAt(view, i)
+      if (span !== null) {
+        i = Math.max(span, i + 1)
+        continue
       }
-      const startNum = groups.start.replace(cachedRegExp(chr, "g"), "")
-      const endNum = groups.end.replace(cachedRegExp(chr, "g"), "")
-      if (/^(?:19|20)\d{2}$/.test(startNum) && /^(?:0[1-9]|1[0-2])$/.test(endNum)) return args[0] as string
-      // Skip 3+4 digit phone-shaped patterns (555-1234, or the second half of
-      // 555-123-4567). Thousands-grouped endings (1,234 / 1.234) keep their
-      // internal separator and so fail /^\d{4}$/, falling through to convert
-      // as ranges. Space/NNBSP/thin-space grouping isn't captured by the outer
-      // range regex, so those variants currently aren't affected here.
-      if (/^\d{3}$/.test(startNum) && /^\d{4}$/.test(endNum)) return args[0] as string
-      // Skip US toll-free prefix pattern: 1-800, 1-888, 1-877, etc. Only the
-      // seven real toll-free codes (800, 888, 877, 866, 855, 844, 833) are
-      // blocked, so genuine ranges like 1-850 or 1-810 still en-dash.
-      if (/^1$/.test(startNum) && /^8(?:00|88|77|66|55|44|33)$/.test(endNum)) return args[0] as string
-      return `${groups.precedingAreaCode ?? ""}${groups.start}${EN_DASH}${groups.end}${groups.suffix ?? ""}`
     }
-  )
+    i++
+  }
+  view.commit()
+}
 
-  // Negative ranges: −5-5 → −5–5, −5--2 → −5–−2
-  // Separate regex because MINUS isn't a word char, so \b in ${wb} would match after it
-  text = text.replace(
-    cachedRegExp(
-      `(?<![${LATIN_LETTERS}])(?<start>${MINUS}\\d[\\d.,]*${chr}?)-{1,3}?(?<neg>-)?(?<end>${chr}?\\d[\\d.,]*)(?<following>(?:${chr}?-${chr}?\\d+)*)(?<suffix>${chr}?[xKBTM])?${wbe}`,
-      "g"
-    ),
-    (...args) => {
-      const groups = namedGroups<{
-        start: string
-        neg?: string
-        end: string
-        following?: string
-        suffix?: string
-      }>(args)
-      if (groups.following) return args[0] as string
-      return `${groups.start}${EN_DASH}${groups.neg ? MINUS : ""}${groups.end}${groups.suffix ?? ""}`
+/**
+ * Attempts `(?<![LAT]) ${MINUS}\d[\d.,]* -{1,3} -? end following* suffix? wbe`
+ * beginning the MINUS at clean offset `i`. Converts the dash run when the gates
+ * pass. Returns the consumed structural end offset, or null when no match begins.
+ */
+function matchNegativeRangeAt(view: ProseView, i: number): number | null {
+  const text = view.text
+  // `(?<![LAT])`: a Latin letter directly before the MINUS (no intervening
+  // boundary, which would hide it) blocks the match.
+  if (boundaryCountAt(view, i) === 0) {
+    const before = text[i - 1]
+    if (before !== undefined && LATIN_LETTER_RE.test(before)) return null
+  }
+
+  // start = `${MINUS}\d[\d.,]*`: the digits after MINUS must not cross a
+  // boundary (the `start` body has no interior slot before its trailing one).
+  if (!/\d/.test(text[i + 1] ?? "")) return null
+  let startEnd = i + 1
+  while (/[\d.,]/.test(text[startEnd] ?? "") && !view.hasBoundary(startEnd)) startEnd++
+  // One slot between start and the dash tolerates a single boundary there.
+  if (boundaryCountAt(view, startEnd) > 1) return null
+  if (text[startEnd] !== "-") return null
+
+  // `-{1,3}?` then optional `(?<neg>-)`: a 1..3 hyphen run with no interior
+  // boundary, the last of which becomes `neg` when two or more hyphens are present.
+  let dashRunEnd = startEnd + 1
+  while (text[dashRunEnd] === "-" && dashRunEnd - startEnd < 3 && !view.hasBoundary(dashRunEnd)) dashRunEnd++
+  const dashRunLen = dashRunEnd - startEnd
+  const hasNeg = dashRunLen >= 2
+  const negEnd = dashRunEnd
+  // The `end` head slot tolerates one boundary.
+  if (boundaryCountAt(view, negEnd) > 1) return null
+
+  const end = readNegEndRun(view, negEnd)
+  if (end === null) return null
+  // The greedy end and `following*` backtrack together; commit at the longest end
+  // whose tail completes, converting only when that tail's `following` is empty.
+  for (let len = end.length; len >= 1; len--) {
+    const endPos = end.firstDigitStart + len
+    const candidates = followingCandidates(view, endPos, false)
+    for (let c = candidates.length - 1; c >= 0; c--) {
+      const suffixEnd = suffixWbeEnd(view, candidates[c], false)
+      if (suffixEnd === null) continue
+      if (c === 0) {
+        view.replace(startEnd, negEnd, `${EN_DASH}${hasNeg ? MINUS : ""}`)
+      }
+      return suffixEnd
     }
-  )
+  }
+  return null
+}
+
+interface NegEndRun {
+  firstDigitStart: number
+  length: number
+}
+
+/** The negative `end` body `\d[\d.,]*` starting at `pos`, boundary-truncated. */
+function readNegEndRun(view: ProseView, pos: number): NegEndRun | null {
+  const text = view.text
+  if (!/\d/.test(text[pos] ?? "")) return null
+  let i = pos + 1
+  while (/[\d.,]/.test(text[i] ?? "") && !view.hasBoundary(i)) i++
+  return { firstDigitStart: pos, length: i - pos }
+}
+
+/**
+ * Boundary-aware `following` end positions, shortest (empty) first, walking
+ * segments `[-(−)]\d+`. Each segment tolerates one boundary on each side of its
+ * dash, and each `\d+` stops at a boundary. `allowMinus` admits a MINUS as a
+ * segment's dash (positive ranges).
+ */
+function followingCandidates(view: ProseView, endSpanEnd: number, allowMinus: boolean): number[] {
+  const candidates = [endSpanEnd]
+  let pos = endSpanEnd
+  for (;;) {
+    if (boundaryCountAt(view, pos) > 1) break
+    const dashCh = view.text[pos]
+    if (dashCh !== "-" && !(allowMinus && dashCh === MINUS)) break
+    const digitStart = pos + 1
+    if (boundaryCountAt(view, digitStart) > 1) break
+    if (!/\d/.test(view.text[digitStart] ?? "")) break
+    let end = digitStart + 1
+    while (/\d/.test(view.text[end] ?? "") && !view.hasBoundary(end)) end++
+    pos = end
+    candidates.push(pos)
+  }
+  return candidates
+}
+
+/**
+ * Single-letter range suffixes: the multipliers `x`/`K`/`B`/`T`/`M` and the
+ * temperature units `C`/`F`. The degree pass rewrites `5C` to `5 °C` — a
+ * space a re-run's word boundary sees through — so the range must already
+ * convert with the unit attached. `X` and `×` are deliberately absent: an
+ * unconverted `1-10X` folds to `1-10×`, which the word-boundary check (which
+ * treats `×` as the word char it folds from) keeps blocking on re-runs.
+ */
+const RANGE_SUFFIX_RE = /[xKBTMCF]/
+
+/**
+ * The `(?:[AaPp][Mm]|[xKBTMCF])?${wbe}` suffix at `pos`, with one tolerated
+ * boundary before the suffix. Returns the end offset after the optional suffix
+ * when the tail completes (wbe passes), or null when it does not.
+ */
+function suffixWbeEnd(view: ProseView, pos: number, allowAmPm: boolean): number | null {
+  if (wordBoundaryEndOk(view, pos)) return pos
+  // An optional suffix: one tolerated boundary then am/pm or one of
+  // {@link RANGE_SUFFIX_RE}'s letters, followed by wbe. The suffix letters sit
+  // at the clean `pos`.
+  if (boundaryCountAt(view, pos) > 1) return null
+  const text = view.text
+  // `[AaPp][Mm]` is contiguous — a boundary between the two letters breaks the
+  // suffix.
+  if (allowAmPm && /[ap]/i.test(text[pos] ?? "") && /m/i.test(text[pos + 1] ?? "") && !view.hasBoundary(pos + 1)) {
+    return wordBoundaryEndOk(view, pos + 2) ? pos + 2 : null
+  }
+  if (RANGE_SUFFIX_RE.test(text[pos] ?? "")) {
+    return wordBoundaryEndOk(view, pos + 1) ? pos + 1 : null
+  }
+  return null
+}
+
+const NOT_AFTER_DASH_RE = new RegExp(`[${DISALLOWED_PREFIX_CLASS_FRAGMENT}${LATIN_LETTERS}${MULTIPLICATION}${PLUS_MINUS}.+]`)
+
+const SPACE_CHAR_RE = new RegExp(`[${SPACE_CHARS}]`)
+
+/**
+ * Offset of the dot reachable backwards across one inter-dot gap ending at the
+ * dot at `dotIndex` (the mirror of the ellipsis pass's forward gap rule: the
+ * gap is nothing, one space character, or one node boundary). -1 when none.
+ */
+function ellipsisPrevDot(view: ProseView, dotIndex: number): number {
+  const text = view.text
+  if (text[dotIndex - 1] === ".") {
+    // Empty gap: at most one boundary between the two dots.
+    return boundaryCountAt(view, dotIndex) <= 1 ? dotIndex - 1 : -1
+  }
+  // Space gap: one space character with no boundary on either side of it.
+  if (boundaryCountAt(view, dotIndex) === 0 && dotIndex >= 2 && SPACE_CHAR_RE.test(text[dotIndex - 1])
+    && text[dotIndex - 2] === "." && boundaryCountAt(view, dotIndex - 1) === 0) {
+    return dotIndex - 2
+  }
+  return -1
+}
+
+/** True iff the dot at `dotIndex` can close a `...`/`. . .` triple. */
+function endsFoldableEllipsisDots(view: ProseView, dotIndex: number): boolean {
+  const second = ellipsisPrevDot(view, dotIndex)
+  return second >= 0 && ellipsisPrevDot(view, second) >= 0
+}
+
+/**
+ * `notAfterDash` is a single-char negative lookbehind. A node boundary
+ * immediately before the start hides the clean char (a boundary is not the
+ * dash/letter), so the guard only fires when a disallowed clean char sits
+ * directly before the start with no intervening boundary.
+ *
+ * A `.` that closes a dot triple does not block: the ellipsis pass (which
+ * runs after the dash rules) folds the triple to `…` and pads a following
+ * digit with a space, so a re-run would see a plain space here and convert —
+ * converting now reaches that fixed point in one pass. A line-leading em dash
+ * does not block for the same reason: the em-dash spacing normalizer pads it
+ * before a digit (`^—5` → `— 5`), mirroring that rule's one tolerated
+ * boundary at line start.
+ */
+function notAfterDashOk(view: ProseView, startOffset: number): boolean {
+  if (startOffset === 0) return true
+  if (boundaryCountAt(view, startOffset) > 0) return true
+  const prev = view.text[startOffset - 1]
+  // A multiplication sign one space slot further left blocks like the `x` it
+  // folds from: the multiplication pass renders `5x8` as `5 × 8`, so a range
+  // blocked at `x8-6` must stay blocked at `× 8-6`.
+  if (prev !== undefined && SPACE_CHAR_RE.test(prev) && view.text[startOffset - 2] === MULTIPLICATION) {
+    return false
+  }
+  if (prev === undefined || !NOT_AFTER_DASH_RE.test(prev)) return true
+  if (prev === "." && endsFoldableEllipsisDots(view, startOffset - 1)) return true
+  const dashIndex = startOffset - 1
+  if (prev === EM_DASH && boundaryCountAt(view, dashIndex) <= 1
+    && (dashIndex === 0 || view.text[dashIndex - 1] === "\n" || emDashBehindLeadingSpaces(view, dashIndex))) {
+    return true
+  }
+  // An en dash behind leading spaces converts the same way: the spaced-dash
+  // rule renders it as a line-leading em dash, which the normalizer then pads.
+  return prev === EN_DASH && boundaryCountAt(view, dashIndex) <= 1
+    && emDashBehindLeadingSpaces(view, dashIndex)
+}
 
-  return text
+/**
+ * True iff only plain spaces sit between text start and the em dash at
+ * `emIndex`. The spaced-dash rule collapses that run, leaving the em dash
+ * line-leading, which the spacing normalizer then pads before a digit — the
+ * same state the `\n` arm of {@link notAfterDashOk} accepts.
+ */
+function emDashBehindLeadingSpaces(view: ProseView, emIndex: number): boolean {
+  let j = emIndex
+  while (j > 0 && view.text[j - 1] === " ") j--
+  return j < emIndex && j === 0
 }
 
+/** Range-number gates: reject year-month, phone-shaped, and toll-free pairs. */
+function rangeNumbersConvert(startNum: string, endNum: string): boolean {
+  if (/^(?:19|20)\d{2}$/.test(startNum) && /^(?:0[1-9]|1[0-2])$/.test(endNum)) return false
+  // Skip 3+4 digit phone-shaped patterns (555-1234, or the second half of
+  // 555-123-4567). Thousands-grouped endings (1,234 / 1.234) keep their
+  // internal separator and so fail /^\d{4}$/, falling through to convert as
+  // ranges.
+  if (/^\d{3}$/.test(startNum) && /^\d{4}$/.test(endNum)) return false
+  // Skip US toll-free prefix pattern: 1-800, 1-888, 1-877, etc. Only the seven
+  // real toll-free codes (800, 888, 877, 866, 855, 844, 833) are blocked, so
+  // genuine ranges like 1-850 or 1-810 still en-dash.
+  if (/^1$/.test(startNum) && /^8(?:00|88|77|66|55|44|33)$/.test(endNum)) return false
+  return true
+}
+
+// ---------------------------------------------------------------------------
+// Date ranges
+// ---------------------------------------------------------------------------
+
 /** Convert month ranges to en-dash (e.g., "January-March" → "January–March"). */
-export function enDashDateRange(text: string, options: DashOptions = {}): string {
-  const dashStyle = options.dashStyle ?? "american"
-  if (dashStyle === "none") return text
-  const chr = getEscapedSeparator(options)
-  const wb = wordBoundaryStart(chr)
-  const wbe = wordBoundaryEnd(chr)
-
-  // Atomic-optional year groups (lookahead + backref). The capture inside
-  // the lookahead is locked in once matched, so the year group commits
-  // without backtracking.
-  const startYear = `(?=(?<startYear>${chr}? \\d{4})?)\\k<startYear>`
-  const endYear = `(?=(?<endYear> \\d{4})?)\\k<endYear>`
-  return text.replace(
-    cachedRegExp(`${wb}(?<startMonth>${monthPattern})${startYear}(?<preSep>${chr}?)(?<preSpace> ?)-(?<postSpace> ?)(?<postSep>${chr}?)(?<endMonth>${monthPattern})${endYear}${wbe}`, "g"),
-    (...args) => {
+export function enDashDateRange(input: string, options?: DashOptions): string
+export function enDashDateRange(input: ProseView, options?: DashOptions): void
+export function enDashDateRange(input: string | ProseView, options: DashOptions = {}): string | void {
+  return overInput(input, (view) => enDashDateRangeOverView(view, options.dashStyle ?? "american"))
+}
+
+function enDashDateRangeOverView(view: ProseView, dashStyle: DashStyle): void {
+  if (dashStyle === "none") return
+  const [pre, post] = dashStyle === "british" ? [" ", " "] : ["", ""]
+
+  // Atomic-optional year groups (lookahead + backref). The capture inside the
+  // lookahead is locked in once matched, so the year group commits without
+  // backtracking. The year space also matches NBSP/NNBSP: the nbsp pass (which
+  // runs after the dash rules) glues "March 2025" with an NBSP, and the year
+  // must still capture on a re-run so the trailing-boundary gate sees the same
+  // end position.
+  const startYear = `(?=(?<startYear>[ ${NBSP_CHARS}]\\d{4})?)\\k<startYear>`
+  const endYear = `(?=(?<endYear>[ ${NBSP_CHARS}]\\d{4})?)\\k<endYear>`
+  const pattern = `\\b(?<startMonth>${monthPattern})${startYear}(?<preSpace> ?)-(?<postSpace> ?)(?<endMonth>${monthPattern})${endYear}\\b`
+  pass(
+    view,
+    cachedRegExp(pattern, "g"),
+    (match, v) => {
+      // The pattern's trailing `\b` sees a folded glyph (`×`, a superscript
+      // ordinal) as a non-word char, but the source character it folds from is
+      // a word char that blocks the match ("March 2025x3"); re-check the end
+      // boundary fold-stably.
+      if (!wordBoundaryEndOk(v, match.index + match[0].length)) return null
       const groups = namedGroups<{
         startMonth: string
         startYear?: string
-        preSep: string
+        preSpace: string
+        postSpace: string
         endMonth: string
         endYear?: string
-        postSep: string
-      }>(args)
-      const [pre, post] = dashStyle === "british" ? [" ", " "] : ["", ""]
-      return `${groups.startMonth}${groups.startYear || ""}${groups.preSep}${pre}${EN_DASH}${post}${groups.postSep}${groups.endMonth}${groups.endYear || ""}`
-    }
+      }>(matchArgs(match))
+      // Replace `[preSpace]-[postSpace]` (the dash and its spaces) with the
+      // styled en-dash, leaving the months and any boundaries around them.
+      const startLen = groups.startMonth.length + (groups.startYear?.length ?? 0)
+      const dashSpan = groups.preSpace.length + 1 + groups.postSpace.length
+      const dashStart = match.index + startLen
+      v.replace(dashStart, dashStart + dashSpan, `${pre}${EN_DASH}${post}`)
+      return null
+    },
+    {
+      allowBoundaries: (match, v) => {
+        const groups = namedGroups<{
+          startMonth: string
+          startYear?: string
+          preSpace: string
+          postSpace: string
+          endMonth: string
+          endYear?: string
+        }>(matchArgs(match))
+        const startLen = groups.startMonth.length + (groups.startYear?.length ?? 0)
+        // One boundary is tolerated after the start year and one before the end
+        // month, bracketing the dash and spaces.
+        const dashStart = match.index + startLen
+        const dashEnd = dashStart + groups.preSpace.length + 1 + groups.postSpace.length
+        const tolerated = new Map<number, number>()
+        tolerated.set(dashStart, 1)
+        tolerated.set(dashEnd, 1)
+        if (!interiorBoundariesAllowed(match, v, tolerated)) return false
+        return wordBoundaryStartOk(v, match.index) && wordBoundaryEndOk(v, match.index + match[0].length)
+      },
+    },
   )
 }
 
+// ---------------------------------------------------------------------------
+// Minus signs
+// ---------------------------------------------------------------------------
+
 /** Convert hyphens to minus signs in numeric contexts (e.g., "-5" → "−5"). */
-export function minusReplace(text: string, options: DashOptions = {}): string {
-  const chr = getEscapedSeparator(options)
-
-  // Pattern 1a: Subtraction of negative number (e.g., "5 - -3" → "5 − −3")
-  // Must come before Pattern 1b so the negative hyphen is consumed first
-  text = text.replace(
-    cachedRegExp(`(?<=\\d${chr}?) - -(?<num>${chr}?\\d*\\.?\\d+)`, "g"),
-    ` ${MINUS} ${MINUS}$<num>`
-  )
+export function minusReplace(input: string): string
+export function minusReplace(input: ProseView): void
+export function minusReplace(input: string | ProseView): string | void {
+  return overInput(input, (view) => {
+    minusSubtractionOfNegative(view)
+    minusSpacedSubtraction(view)
+    minusDirectNegative(view)
+  })
+}
 
-  // Pattern 1b: Spaced math subtraction (e.g., "5 - 3" → "5 − 3")
-  // Only when preceded by a digit - this distinguishes "5 - 3" from "Safari) - 9"
-  text = text.replace(
-    cachedRegExp(`(?<=\\d${chr}?) - (?<num>${chr}?\\d*\\.?\\d+)`, "g"),
-    ` ${MINUS} $<num>`
+/** Pattern 1a/1b: spaced math subtraction after a digit (`5 - 3`, `5 - -3`). */
+function minusSubtractionOfNegative(view: ProseView): void {
+  // "5 - -3" → "5 − −3"; the negative hyphen is consumed first. The slot before
+  // `num` tolerates one boundary there (the `- -` prefix has 4 chars).
+  pass(
+    view,
+    cachedRegExp(`(?<=\\d) - -(?<num>\\d*\\.?\\d+)`, "g"),
+    (match, v) => {
+      // Replace only the ` - -` prefix; `num` (which may carry interior
+      // boundaries that truncate it) stays untouched.
+      if (!minusSubtractionPrefixOk(match, v, 4)) return null
+      v.replace(match.index, match.index + 4, ` ${MINUS} ${MINUS}`)
+      return null
+    },
+    { allowBoundaries: () => true },
   )
+}
 
-  // Pattern 2: Direct negative numbers (e.g., "-5" → "−5", "(-3)" → "(−3)")
-  // Match after: start of line, whitespace, (, or quotes (straight or curly)
-  // No space allowed between hyphen and digit
-  text = text.replace(
-    cachedRegExp(`(?<before>^|[\\s\\("'${LEFT_DOUBLE_QUOTE}${RIGHT_DOUBLE_QUOTE}${LEFT_SINGLE_QUOTE}${RIGHT_SINGLE_QUOTE}])-(?<num>\\d*\\.?\\d+)`, "gm"),
-    `$<before>${MINUS}$<num>`
+function minusSpacedSubtraction(view: ProseView): void {
+  // "5 - 3" → "5 − 3"; only after a digit ("5 - 3" not "Safari) - 9").
+  pass(
+    view,
+    cachedRegExp(`(?<=\\d) - (?<num>\\d*\\.?\\d+)`, "g"),
+    (match, v) => {
+      if (!minusSubtractionPrefixOk(match, v, 3)) return null
+      v.replace(match.index, match.index + 3, ` ${MINUS} `)
+      return null
+    },
+    { allowBoundaries: () => true },
   )
+}
 
-  // Pattern 2b: Direct negative numbers after separator boundary (e.g., <em>-5</em>)
-  // Only when no word character precedes the separator — prevents both
-  // "2{SEP}-3" in "1-2-3" and "GPT{SEP}-3" from being misidentified as negative 3
-  text = text.replace(
-    cachedRegExp(`(?<![\\d.,${LATIN_LETTERS}])(?<before>${chr})-(?<num>\\d*\\.?\\d+)`, "gm"),
-    `$<before>${MINUS}$<num>`
-  )
+const NUM_BODY_RE = /^\d*\.?\d+/
 
-  return text
+/**
+ * The subtraction patterns tolerate one boundary after the leading digit (at
+ * the match start) and one before `num`. The `-` prefix span ([match.index,
+ * match.index + prefixLen)) must stay boundary-free; one boundary may sit at
+ * `num`'s head. The `num` body (`\d*\.?\d+`) carries no interior slot, so it
+ * stops at the first interior boundary: a boundary that truncates `num` to a
+ * still-valid number (e.g. "12" in "5 - 12{b}5") leaves the match standing, but
+ * one that strips the mandatory `\d+` (e.g. "." in "5 - .{b}5") breaks it. The
+ * targeted edit only rewrites the ` - ` prefix, so a converting match must keep
+ * `num` satisfiable.
+ */
+function minusSubtractionPrefixOk(match: RegExpExecArray, view: ProseView, prefixLen: number): boolean {
+  // The slot in the lookbehind (`(?<=\d)`) sits at the match start between the
+  // leading digit and the space; it tolerates at most one boundary.
+  if (boundaryCountAt(view, match.index) > 1) return false
+  const numStart = match.index + prefixLen
+  for (const b of view.boundaries) {
+    if (b > match.index && b < numStart) return false
+  }
+  // The slot before `num` tolerates at most one boundary at the num head.
+  if (boundaryCountAt(view, numStart) > 1) return false
+  // `num` stops at the first interior boundary; the clean run up to it must
+  // still satisfy `\d*\.?\d+` (the mandatory trailing `\d+`).
+  const numEnd = match.index + match[0].length
+  let truncatedEnd = numEnd
+  for (const b of view.boundaries) {
+    if (b > numStart && b < numEnd) { truncatedEnd = b; break }
+    if (b >= numEnd) break
+  }
+  return NUM_BODY_RE.test(view.text.slice(numStart, truncatedEnd))
 }
 
-function convertParentheticalDashes(text: string, sep: string, style: DashStyle): string {
+const MINUS_BEFORE_CLASS_RE = new RegExp(`[\\s("'${LEFT_DOUBLE_QUOTE}${RIGHT_DOUBLE_QUOTE}${LEFT_SINGLE_QUOTE}${RIGHT_SINGLE_QUOTE}]`)
+// The folded word glyphs block like the word chars they fold from ("1x|-0"
+// and "1×|-0" must agree), the folded math operators block like the operator
+// chars they fold from (the `!=` → `≠` fold consumes the `=` that blocked
+// Pattern 2 in "!=-1"), and the folded terminals — plus the bare `!` the
+// `!!` normalization leaves — block like the `!`/`?`/`.` runs they fold from
+// ("?|!-0" folds to "⁈|-0" with the hyphen stranded boundary-adjacent).
+const FOLDED_WORD_CLASS_FRAGMENT = [...FOLDED_WORD_CHARS].join("")
+const FOLDED_MATH_CLASS_FRAGMENT = [...FOLDED_MATH_OPERATORS].join("")
+const FOLDED_TERMINAL_CLASS_FRAGMENT = [
+  UNICODE_SYMBOLS.ELLIPSIS, UNICODE_SYMBOLS.DOUBLE_QUESTION, UNICODE_SYMBOLS.QUESTION_EXCLAMATION,
+  UNICODE_SYMBOLS.EXCLAMATION_QUESTION, UNICODE_SYMBOLS.DOUBLE_EXCLAMATION, UNICODE_SYMBOLS.INTERROBANG, "!",
+].join("")
+const NUM_BEFORE_2B_BLOCK_RE = new RegExp(`[\\d.,${LATIN_LETTERS}${FOLDED_WORD_CLASS_FRAGMENT}${FOLDED_MATH_CLASS_FRAGMENT}${FOLDED_TERMINAL_CLASS_FRAGMENT}]`)
+
+/**
+ * Direct negative numbers, in two cases: Pattern 2 converts `-\d` after line
+ * start, whitespace, `(`, or a quote; Pattern 2b converts a boundary-led hyphen
+ * `-\d` when no digit/`.`/`,`/Latin precedes the boundary — which prevents
+ * `1{b}-2` and `GPT{b}-3` from reading as negatives. Pattern 2 fires when the
+ * clean char before the hyphen qualifies and no boundary intrudes; Pattern 2b
+ * fires when a boundary leads the hyphen.
+ */
+function minusDirectNegative(view: ProseView): void {
+  pass(
+    view,
+    cachedRegExp(`-(?<num>\\d*\\.?\\d+)`, "gm"),
+    (match, v) => {
+      const start = match.index
+      // `num` (`\d*\.?\d+`) carries no interior slot, so it stops at the first
+      // interior boundary; the clean run up to it must still satisfy
+      // `\d*\.?\d+`. A boundary right after the hyphen (head) leaves no digits
+      // ("-{b}5"), and one that strips the mandatory `\d+` ("-.{b}1") breaks the
+      // match, but one past a valid number ("-5{b}5") leaves it intact.
+      const numEnd = start + match[0].length
+      let numTruncEnd = numEnd
+      for (const b of v.boundaries) {
+        if (b > start && b < numEnd) { numTruncEnd = b; break }
+        if (b >= numEnd) break
+      }
+      if (!NUM_BODY_RE.test(v.text.slice(start + 1, numTruncEnd))) return null
+      // A boundary directly before the hyphen routes to Pattern 2b; otherwise
+      // Pattern 2 examines the clean preceding char. (The two never overlap: a
+      // hyphen has either a boundary or a clean char immediately before it.)
+      const nb = boundaryCountAt(v, start)
+      if (nb > 0) {
+        // Pattern 2b: the char before the boundary run must not be a digit,
+        // `.`, `,`, or Latin letter. With ≥2 stacked boundaries that char is
+        // another boundary (admitted); with one, it is the clean char left.
+        if (nb < 2) {
+          const before = v.text[start - 1]
+          if (before !== undefined && NUM_BEFORE_2B_BLOCK_RE.test(before)) return null
+        }
+      } else {
+        // Pattern 2: line start, whitespace, `(`, or a quote immediately left.
+        const before = v.text[start - 1]
+        const atLineStart = before === undefined || before === "\n"
+        if (!atLineStart && !MINUS_BEFORE_CLASS_RE.test(before)) return null
+      }
+      v.replace(start, start + 1, MINUS)
+      return null
+    },
+    { allowBoundaries: () => true },
+  )
+}
+
+// ---------------------------------------------------------------------------
+// Parenthetical dashes
+// ---------------------------------------------------------------------------
+
+function convertParentheticalDashes(view: ProseView, style: DashStyle): void {
   const localizedDash = style === "british" ? EN_DASH : EM_DASH
   const maybeSpace = style === "british" ? " " : ""
-  const escapedSep = getEscapedSeparator({ separator: sep })
-
-  // Convert spaced dashes: "word - word" or "word — word"
-  // When a separator follows the dash, preserve trailing spaces (they belong to the next text segment).
-  // A single plain hyphen directly before a word character (through optional separators) is a
-  // suspended/hanging hyphen ("Yes-men and -women"), not a parenthetical dash — skip it.
-  // Em/en dashes and multiple hyphens can have zero trailing spaces.
-  // `sepAfter` and `trailing` share one optional unit so a whitespace run
-  // can only be assigned to it as a whole, keeping the match unambiguous.
-  const spacedDashPattern = cachedRegExp(
-    `(?<=[^\\s]|^)(?<sepBefore>${escapedSep}?)[ ]+(?:[${EN_DASH}${EM_DASH}][-${EN_DASH}${EM_DASH}]*|-{2,}|-(?!${escapedSep}*[${LATIN_LETTERS}\\d]))(?!${escapedSep}?-*${escapedSep}?>)[ ]*(?:(?<sepAfter>${escapedSep})(?<trailing>[ ]*))?(?=\\S|$)`, "g"
-  )
-  text = text.replace(spacedDashPattern, (_match, sepBefore, sepAfter, trailing) => {
-    // sepAfter and trailing are undefined when the optional group didn't match.
-    sepAfter = sepAfter ?? ""
-    trailing = trailing ?? ""
-    // Trailing space after a separator belongs to the next text node — keep it
-    // regardless of style. Without a separator, the trailing space is part of
-    // the current node's whitespace around the dash and gets consumed.
-    const keepTrailing = sepAfter ? trailing : ""
-    return `${sepBefore}${maybeSpace}${localizedDash}${maybeSpace}${sepAfter}${keepTrailing}`
-  })
-  // Convert dashes at text node boundaries: separator alone precedes dash (e.g., "word{sep}– rest")
-  // Requires space after the dash to avoid matching number ranges like "1{sep}-{sep}5"
-  text = text.replace(
-    cachedRegExp(`(?<=[^\\s])(?<sepBefore>${escapedSep})[${EN_DASH}${EM_DASH}-]+[ ]+(?<sepAfter>${escapedSep}?)`, "g"),
-    `$<sepBefore>${maybeSpace}${localizedDash}${maybeSpace}$<sepAfter>`
+  const rendered = `${maybeSpace}${localizedDash}${maybeSpace}`
+
+  convertSpacedDashes(view, rendered)
+  convertBoundaryLedDashes(view, rendered)
+  convertMultipleDashes(view, rendered)
+  convertLeadingLineDashes(view, localizedDash)
+  if (style === "british") convertUnspacedEmDashes(view, rendered)
+}
+
+const QUOTE_CHARS = `"'${LEFT_DOUBLE_QUOTE}${RIGHT_DOUBLE_QUOTE}${LEFT_SINGLE_QUOTE}${RIGHT_SINGLE_QUOTE}`
+
+const DASH_RUN_CHARS = new Set(["-", EN_DASH, EM_DASH])
+
+function isDashChar(ch: string | undefined): boolean {
+  return ch !== undefined && DASH_RUN_CHARS.has(ch)
+}
+
+/**
+ * Spaced dashes: "word - word" / "word — word". A single plain hyphen directly
+ * before a word char (the suspended/hanging hyphen, "Yes-men and -women") is
+ * skipped. The arrow guard rejects shapes like "- >" so `foo -> bar` stays.
+ *
+ * The dash core (`[${EN}${EM}][-${EN}${EM}]*|-{2,}|-(?![LAT\d])`) never spans a
+ * boundary, so a node boundary inside the dash run truncates the run — the
+ * boundary becomes the trailing slot and any dashes past it are left untouched.
+ * The replacer realizes this by truncating the run at the first interior
+ * boundary and re-validating the dash core on the truncated form.
+ */
+function convertSpacedDashes(view: ProseView, rendered: string): void {
+  // The `(?<=[^\s]|^)` lookbehind is validated in the replacer instead, since a
+  // boundary (non-space) also satisfies it, letting a match begin even when the
+  // clean char to the left is whitespace. The cheap `(?<![ ])` anchor keeps each
+  // `[ ]+` run starting at its leftmost space, avoiding quadratic backtracking.
+  const pattern = `(?<![ ])[ ]+(?<dashStart>[-${EN_DASH}${EM_DASH}])`
+  // The full shape consumes the trailing `[ ]*`, advancing past it; this pattern
+  // does not, so a later match could begin inside a previous match's consumed
+  // trailing spaces. Track the last consumed offset and skip.
+  let consumedThrough = -1
+  pass(
+    view,
+    cachedRegExp(pattern, "g"),
+    (match, v) => {
+      const text = v.text
+      const groups = namedGroups<{ dashStart: string }>(matchArgs(match))
+      const firstDash = match.index + match[0].length - groups.dashStart.length
+      // The greedy `[ ]+` starts at the leftmost position whose `(?<=[^\s]|^)`
+      // holds: line/text start, a non-whitespace clean char, or a boundary. A
+      // boundary inside the leading spaces restarts the match just after it (the
+      // spaces before it stay with the previous node). The previous match's
+      // consumed span is also a floor: leading spaces it already consumed are not
+      // re-used.
+      let matchStart = Math.max(match.index, consumedThrough)
+      for (const b of view.boundaries) {
+        if (b > matchStart && b <= firstDash - 1) matchStart = b
+        else if (b > firstDash) break
+      }
+      /* istanbul ignore if -- defensive: the previous match's trailing `(?=\S|$)`
+         stops at this dash, so `consumedThrough` never reaches past `firstDash`. */
+      if (matchStart >= firstDash + 1) return null
+      // Validate the lookbehind at matchStart: start of text/line, a boundary
+      // immediately left, or a non-whitespace clean char.
+      const leftOk = matchStart === 0
+        || view.hasBoundary(matchStart)
+        || (text[matchStart - 1] !== undefined && !/\s/.test(text[matchStart - 1]))
+      if (!leftOk) return null
+      // A dash char left of the leading spaces (boundaries are zero-width, so
+      // this reads through them) marks this run as the tail of an
+      // already-spaced dash ("x – -"): the overlap consumption above leaves
+      // it on the pass that converts the head, and the head's edits can
+      // rewrite the linking context so the consumption never re-links; block
+      // the tail deterministically. Glyphs folded from a trailing dash (`←`,
+      // `±`) block like the hyphen that sat in this slot before the fold.
+      const charBeforeSpaces = text[matchStart - 1]
+      if (matchStart > 0
+        && (isDashChar(charBeforeSpaces) || FOLDED_FROM_TRAILING_DASH.has(charBeforeSpaces))) return null
+      // The leading spaces between matchStart and the dash must be at least one
+      // (`[ ]+` is one-or-more) — a boundary directly before the dash leaves
+      // none, which is the boundary-led pattern's job, not this one.
+      if (matchStart === firstDash) return null
+      // A boundary between the leading spaces and the dash breaks `[ ]+dashCore`
+      // (no slot there); neither spaced nor boundary-led matches, so skip.
+      if (view.hasBoundary(firstDash)) return null
+      // Maximal dash run from firstDash that does not cross a boundary (the core
+      // never spans a boundary).
+      let fullRunEnd = firstDash + 1
+      while (isDashChar(text[fullRunEnd]) && !v.hasBoundary(fullRunEnd)) fullRunEnd++
+      // The dashCore (`-{2,}` greedy, then single `-`) backtracks past the arrow
+      // guard AND the trailing `[ ]*(?=\S|$)`, so try run ends longest-first and
+      // take the first one whose whole tail validates.
+      const chosen = chooseSpacedDashRun(v, firstDash, fullRunEnd)
+      if (chosen === null) return null
+      consumedThrough = chosen.consumedEnd
+      v.replace(matchStart, chosen.editEnd, rendered)
+      return null
+    },
+    { allowBoundaries: () => true },
   )
-  // Convert multiple dashes: "word--word" or "word---word" or "quote"--"quote"
+}
+
+/**
+ * Picks the dashCore from the maximal dash run [firstDash, fullRunEnd). Tries
+ * `[EN EM][-EN EM]*`, then `-{2,}` (greedy), then a single `-`, backtracking past
+ * the arrow guard and the trailing `[ ]*(?=\S|$)`. Returns the chosen run end and
+ * the trailing-spaces end (`scan`) to replace, or null when no branch validates.
+ */
+function chooseSpacedDashRun(view: ProseView, firstDash: number, fullRunEnd: number): SpacedTrailing | null {
+  const text = view.text
+  const first = text[firstDash]
+  if (first === EN_DASH || first === EM_DASH) {
+    // `[EN EM][-EN EM]*`: the leading EN/EM is mandatory, the `[-EN EM]*` tail is
+    // greedy but backtracks past the arrow guard — e.g. "–-{sep}{sep}>" gives the
+    // `-` back so the run is just "–" (not an arrow) and converts.
+    for (let end = fullRunEnd; end > firstDash; end--) {
+      if (spacedArrowAhead(view, end)) continue
+      const trailing = spacedTrailingEnd(view, end)
+      if (trailing !== null) return trailing
+    }
+    return null
+  }
+  // The `-{2,}` and single-`-` branches match hyphens only, so a leading-hyphen
+  // run stops at the first EN/EM dash (e.g. "-–" is a single `-` core, not a
+  // two-dash run that would swallow the EN dash).
+  let hyphenRunEnd = firstDash
+  while (text[hyphenRunEnd] === "-" && hyphenRunEnd < fullRunEnd) hyphenRunEnd++
+  // All hyphens. Try `-{2,}` lengths longest-first, then a single `-`.
+  for (let end = hyphenRunEnd; end > firstDash; end--) {
+    if (end - firstDash === 1) {
+      // Single `-(?!${sep}*[LAT\d])`: the next clean char must not be a letter or
+      // digit (the suspended-hyphen guard).
+      const next = text[end]
+      if (next !== undefined && /[a-z\d]/i.test(next)) return null
+    }
+    if (spacedArrowAhead(view, end)) continue
+    const trailing = spacedTrailingEnd(view, end)
+    if (trailing !== null) return trailing
+  }
+  return null
+}
+
+interface SpacedTrailing {
+  /** Offset through which the edit deletes (kept post-boundary spaces excluded). */
+  editEnd: number
+  /** Offset past all consumed trailing chars (for overlap/`lastIndex` tracking). */
+  consumedEnd: number
+}
+
+/**
+ * The trailing `[ ]*(?:[ ]*)?(?=\S|$)` with one tolerated boundary. The leading
+ * `[ ]*` (before any boundary) is deleted; spaces after a boundary are kept
+ * (they belong to the next node) but still consumed. The lookahead then requires
+ * a non-space char, a boundary, or end of text. Returns the edit/consumed
+ * offsets, or null on failure.
+ */
+function spacedTrailingEnd(view: ProseView, runEnd: number): SpacedTrailing | null {
+  const text = view.text
+  let editEnd = runEnd
+  while (text[editEnd] === " " && !view.hasBoundary(editEnd)) editEnd++
+  // The trailing optional group is greedy: try consuming a single boundary and
+  // its kept trailing spaces first, then fall back to no boundary. `(?=\S|$)` is
+  // satisfied by a boundary (non-space), a non-space clean char, or end of text.
+  if (boundaryCountAt(view, editEnd) === 1) {
+    let consumedEnd = editEnd
+    while (text[consumedEnd] === " " && (consumedEnd === editEnd || !view.hasBoundary(consumedEnd))) consumedEnd++
+    if (lookaheadNonSpace(view, consumedEnd)) return { editEnd, consumedEnd }
+  }
+  if (lookaheadNonSpace(view, editEnd)) return { editEnd, consumedEnd: editEnd }
+  return null
+}
+
+/** `(?=\S|$)`: end of text, a node boundary, or a non-space clean char. */
+function lookaheadNonSpace(view: ProseView, pos: number): boolean {
+  return pos === view.text.length || view.hasBoundary(pos) || !/\s/.test(view.text[pos])
+}
+
+/**
+ * The `(?!-*>)` arrow guard, boundary-aware. Two slots bracket the `-*` hyphen
+ * run: one boundary may sit before the run and one after. With no hyphens both
+ * slots collapse to the run-end offset (up to two boundaries there); a boundary
+ * inside the hyphen run, or more than the two slots allow, hides the `>` (so the
+ * dash is not an arrow and converts).
+ */
+function spacedArrowAhead(view: ProseView, runEnd: number): boolean {
+  const text = view.text
+  let h = runEnd
+  while (text[h] === "-") {
+    if (h > runEnd && view.hasBoundary(h)) return false // boundary inside -*
+    h++
+  }
+  if (h === runEnd) {
+    // No hyphens: both slots fall at this offset (≤2 boundaries total).
+    if (boundaryCountAt(view, runEnd) > 2) return false
+  } else {
+    if (boundaryCountAt(view, runEnd) > 1) return false // first slot
+    if (boundaryCountAt(view, h) > 1) return false // second slot
+  }
+  // `≥` guards like the `>=` it folds from: the math pass folds the `>` this
+  // guard keys on, and a dash blocked before `–>=` must stay blocked at `–≥`.
+  return text[h] === ">" || text[h] === UNICODE_SYMBOLS.GREATER_EQUAL
+}
+
+/**
+ * Dashes where a boundary alone precedes the dash run: "word{b}– rest". The
+ * shape `(?<=[^\s])[dash]+[ ]+` with a leading boundary requires a boundary
+ * immediately before the dash run, a non-space char before that boundary, and
+ * at least one trailing space. Because the leading boundary is required, this
+ * only fires where a node boundary opens a dash run, so the pass scans
+ * boundaries rather than the clean text. The "non-space before the boundary" is
+ * satisfied by a real non-space clean char (single boundary) or by another
+ * stacked boundary (two or more boundaries — a boundary is itself non-space).
+ */
+function convertBoundaryLedDashes(view: ProseView, rendered: string): void {
+  const text = view.text
+  const edits: Array<[number, number]> = []
+  let lastProcessedEnd = -1
+  let prevOffset = -1
+  let stackedHere = 0
+  for (const b of view.boundaries) {
+    stackedHere = b === prevOffset ? stackedHere + 1 : 1
+    prevOffset = b
+    if (b < lastProcessedEnd) continue
+    if (!isDashChar(text[b])) continue
+    // Non-space before the boundary: ≥2 stacked boundaries, or a non-space
+    // clean char immediately left. A folded math operator there marks the
+    // fold-stranded configuration: the math pass consumes one character
+    // across this junction (`=` + `~—` folds to `≈` + `—`), so the character
+    // that blocked this rule on the source text is gone, and converting now
+    // would diverge from that blocked first pass.
+    if (stackedHere < 2) {
+      const before = text[b - 1]
+      if (before === undefined || /\s/.test(before) || FOLDED_MATH_OPERATORS.has(before)) continue
+    }
+    // Dash run from the boundary, not crossing a further boundary.
+    let runEnd = b + 1
+    while (isDashChar(text[runEnd]) && !view.hasBoundary(runEnd)) runEnd++
+    // Required trailing space run; stop at a boundary.
+    let spaceEnd = runEnd
+    while (text[spaceEnd] === " " && !view.hasBoundary(spaceEnd)) spaceEnd++
+    if (spaceEnd === runEnd) continue // no trailing space
+    // A dash (or a glyph folded from a dash-led source) after the spaces
+    // blocks when the rendered form ends in a dash: consuming the spaces
+    // would fuse the rendered dash with it into a run the next pass
+    // collapses (see the same guard in the multiple-dash rule). A boundary
+    // at the junction prevents the fusion, so only a same-node follower
+    // blocks.
+    const afterSpaces = text[spaceEnd]
+    if (afterSpaces !== undefined && !view.hasBoundary(spaceEnd)
+      && isDashChar(rendered[rendered.length - 1])
+      && (isDashChar(afterSpaces) || FOLDED_FROM_LEADING_DASH.has(afterSpaces))) continue
+    edits.push([b, spaceEnd])
+    lastProcessedEnd = spaceEnd
+  }
+  for (const [from, to] of edits) view.replace(from, to, rendered)
+  view.commit()
+}
+
+/**
+ * Multiple dashes: "word--word", "word---word", `"quote"--"quote"`. The shape
+ * `(?<=[LAT\d quote])[dash]{2,50}(?=[LAT quote space])` tolerates one boundary
+ * on each side of the run, but its `[dash]{2,50}` never spans a boundary. A
+ * boundary inside the run truncates it (and the 2..50 count must still hold for
+ * the truncated run); two stacked boundaries on either edge exceed the single
+ * slot and block the match.
+ */
+function convertMultipleDashes(view: ProseView, rendered: string): void {
+  const before = `[${LATIN_LETTERS}\\d${QUOTE_CHARS}]`
+  const after = `[${LATIN_LETTERS}${QUOTE_CHARS} ]`
   // Upper bound of 50 prevents ReDoS on pathological runs of dashes.
-  const quoteChars = `"'${LEFT_DOUBLE_QUOTE}${RIGHT_DOUBLE_QUOTE}${LEFT_SINGLE_QUOTE}${RIGHT_SINGLE_QUOTE}`
-  text = text.replace(
-    cachedRegExp(`(?<=[${LATIN_LETTERS}\\d${quoteChars}])(?<sepBefore>${escapedSep}?)[${EN_DASH}${EM_DASH}-]{2,50}(?<sepAfter>${escapedSep}?)(?=[${LATIN_LETTERS}${quoteChars} ])`, "g"),
-    `$<sepBefore>${maybeSpace}${localizedDash}${maybeSpace}$<sepAfter>`
+  const pattern = `(?<=${before})[${EN_DASH}${EM_DASH}-]{2,50}(?=${after})`
+  const afterRe = new RegExp(after)
+  pass(
+    view,
+    cachedRegExp(pattern, "g"),
+    (match, v) => {
+      const text = v.text
+      const start = match.index
+      // Two stacked boundaries before the run exceed the leading slot's budget.
+      if (boundaryCountAt(v, start) >= 2) return null
+      // Truncate the run at the first interior boundary (the class stops there).
+      let runEnd = start + 1
+      while (isDashChar(text[runEnd]) && !v.hasBoundary(runEnd)) runEnd++
+      if (runEnd - start < 2 || runEnd - start > 50) return null
+      // The trailing slot tolerates one boundary; the trailing lookahead then
+      // checks the clean char after the run is [LAT quote space].
+      if (boundaryCountAt(v, runEnd) >= 2) return null
+      const next = text[runEnd]
+      if (next === undefined || !afterRe.test(next)) return null
+      // Consume the trailing space run (not crossing a boundary): the rendered
+      // dash carries its own spacing, and a re-run's spaced-dash rule would
+      // otherwise collapse the leftover spaces into it. When the rendered
+      // form ends in a dash (the unspaced American em dash), a dash (or a
+      // glyph folded from a dash-led source) after the spaces keeps them:
+      // consuming would fuse the rendered dash with it into a fresh run, and
+      // the spaced-dash rule's dash-tail guard already holds the leftover
+      // spaces still.
+      let editEnd = runEnd
+      while (text[editEnd] === " " && !v.hasBoundary(editEnd)) editEnd++
+      // A boundary at the junction prevents the fusion (dash runs never span
+      // boundaries), so only a same-node follower blocks.
+      const afterSpaces = text[editEnd]
+      if (editEnd > runEnd && afterSpaces !== undefined && !v.hasBoundary(editEnd)
+        && isDashChar(rendered[rendered.length - 1])
+        && (isDashChar(afterSpaces) || FOLDED_FROM_LEADING_DASH.has(afterSpaces))) {
+        editEnd = runEnd
+      }
+      v.replace(start, editEnd, rendered)
+      return null
+    },
+    { allowBoundaries: () => true },
   )
-  // Convert dashes at start of line
-  text = text.replace(cachedRegExp(`^(?<leadingSep>${escapedSep})?[-]+ `, "gm"), `$<leadingSep>${localizedDash} `)
-  // British: convert unspaced em-dashes to spaced en-dashes (word—word → word – word)
-  if (style === "british") {
-    text = text.replace(
-      cachedRegExp(`(?<=[${LATIN_LETTERS}.!?'"])(?<sepBefore>${escapedSep}?)${EM_DASH}(?<sepAfter>${escapedSep}?)(?=[${LATIN_LETTERS}])`, "g"),
-      `$<sepBefore>${maybeSpace}${localizedDash}${maybeSpace}$<sepAfter>`
-    )
-  }
-  return text
 }
 
+/**
+ * Dashes at start of line: "^- " → styled dash. One boundary is tolerated at
+ * line start (before the dashes); it sits at the match start, never interior.
+ * The dash run has no interior slot, so a boundary inside the run blocks the
+ * match — the default interior-boundary rejection, which is what we want here.
+ */
+function convertLeadingLineDashes(view: ProseView, localizedDash: string): void {
+  pass(view, cachedRegExp(`^-+ `, "gm"), () => `${localizedDash} `)
+}
+
+/** British: unspaced em-dashes between letters become spaced en-dashes. */
+function convertUnspacedEmDashes(view: ProseView, rendered: string): void {
+  const pattern = `(?<=[${LATIN_LETTERS}.!?'"])${EM_DASH}(?=[${LATIN_LETTERS}])`
+  pass(
+    view,
+    cachedRegExp(pattern, "g"),
+    (match, v) => {
+      // One boundary is tolerated on each side of the em-dash; two or more
+      // exceed the slot and block the match.
+      if (boundaryCountAt(v, match.index) > 1) return null
+      if (boundaryCountAt(v, match.index + match[0].length) > 1) return null
+      v.replace(match.index, match.index + match[0].length, rendered)
+      return null
+    },
+  )
+}
+
+// ---------------------------------------------------------------------------
+// Em-dash spacing normalization (american)
+// ---------------------------------------------------------------------------
+
 // TODO: Handle interrupted-then-resumed speech within quotes, where Chicago
 // allows a space after the dash: "Don't inter— Hey! Who threw that?"
-function normalizeEmDashSpacing(text: string, sep: string): string {
-  const escapedSep = getEscapedSeparator({ separator: sep })
-
-  // Remove all spaces around em-dashes. The `\S|^` / `\S|$` boundary
-  // anchors restrict the match to dashes adjacent to non-whitespace.
-  text = text.replace(
-    cachedRegExp(`(?<=\\S|^)(?<before>${escapedSep}?)[ ]*${EM_DASH}[ ]*(?<after>${escapedSep}?)(?=\\S|$)`, "g"),
-    `$<before>${EM_DASH}$<after>`
-  )
+function normalizeEmDashSpacing(view: ProseView): void {
+  removeSpacesAroundEmDash(view)
+  preserveLineLeadingEmDashSpace(view)
+}
 
-  // Preserve space after em-dash at start of line (e.g., attribution)
-  text = text.replace(
-    cachedRegExp(`^(?<sep>${escapedSep}?)${EM_DASH}(?<after>[A-Z0-9])`, "gm"),
-    `$<sep>${EM_DASH} $<after>`
+/**
+ * Remove spaces directly around an em-dash that has non-whitespace on both
+ * sides. Only the spaces immediately adjacent to the dash are consumed; a
+ * boundary between a space and the dash keeps that space (space consumption
+ * cannot reach across a boundary). The walk therefore stops space consumption
+ * at node boundaries, and the `\S|^` / `\S|$` anchors treat a boundary (a
+ * non-space) as satisfying the anchor.
+ */
+function removeSpacesAroundEmDash(view: ProseView): void {
+  pass(
+    view,
+    cachedRegExp(EM_DASH, "g"),
+    (match, v) => {
+      const d = match.index
+      const text = v.text
+      // Leading spaces, stopping at a boundary or non-space.
+      let left = d
+      while (left > 0 && text[left - 1] === " " && !v.hasBoundary(left)) left--
+      // Trailing spaces, stopping at a boundary or non-space.
+      let right = d + 1
+      while (right < text.length && text[right] === " " && !v.hasBoundary(right)) right++
+      const leftAnchored = left === 0 || v.hasBoundary(left) || !/\s/.test(text[left - 1])
+      const rightAnchored = right === text.length || v.hasBoundary(right) || !/\s/.test(text[right])
+      if (!leftAnchored || !rightAnchored) return null
+      // Drop the adjacent space runs; the em-dash and any edge boundaries
+      // stay. A side whose anchor is another dash keeps its spaces: deleting
+      // them would fuse the dashes into a run the multiple-dash rule then
+      // collapses on the next pass. Glyphs folded from a dash-edged source
+      // (`±`, `←`, `→`) anchor like the dash they fold from.
+      const leftAnchor = text[left - 1]
+      const rightAnchor = text[right]
+      if (left < d && !isDashChar(leftAnchor) && !FOLDED_FROM_TRAILING_DASH.has(leftAnchor)) v.replace(left, d, "")
+      if (right > d + 1 && !isDashChar(rightAnchor) && !FOLDED_FROM_LEADING_DASH.has(rightAnchor)) v.replace(d + 1, right, "")
+      return null
+    },
   )
+}
 
-  return text
+/** Preserve a space after a line-leading em-dash before an uppercase/digit. */
+function preserveLineLeadingEmDashSpace(view: ProseView): void {
+  pass(
+    view,
+    cachedRegExp(`^${EM_DASH}(?<after>[A-Z0-9])`, "gm"),
+    (match, v) => {
+      // One boundary is tolerated between the line start and the em-dash; two or
+      // more exceed that slot and block the match.
+      if (boundaryCountAt(v, match.index) > 1) return null
+      const groups = namedGroups<{ after: string }>(matchArgs(match))
+      return `${EM_DASH} ${groups.after}`
+    },
+  )
 }
 
-export function hyphenReplace(text: string, options: DashOptions = {}): string {
-  const sep = options.separator ?? DEFAULT_SEPARATOR
+// ---------------------------------------------------------------------------
+// Public dash pipeline
+// ---------------------------------------------------------------------------
+
+export function hyphenReplace(input: string, options?: DashOptions): string
+export function hyphenReplace(input: ProseView, options?: DashOptions): void
+export function hyphenReplace(input: string | ProseView, options: DashOptions = {}): string | void {
   const style = options.dashStyle ?? "american"
-  if (style === "none") return text
-  text = minusReplace(text, options)
-  text = enDashDateRange(text, options)
-  text = enDashNumberRange(text, options)
-  text = convertParentheticalDashes(text, sep, style)
-  if (style === "american") text = normalizeEmDashSpacing(text, sep)
-  return text
+  return overInput(input, (view) => {
+    if (style === "none") return
+    minusReplace(view)
+    enDashDateRangeOverView(view, style)
+    enDashNumberRange(view)
+    convertParentheticalDashes(view, style)
+    if (style === "american") normalizeEmDashSpacing(view)
+  })
 }
 
 const { WORD_JOINER } = UNICODE_SYMBOLS
@@ -292,7 +1292,16 @@ const { WORD_JOINER } = UNICODE_SYMBOLS
  * `nbspTransform` (U+00A0 breaks `Fig. 1` searches). Apply only in rendered
  * HTML; do not write into Markdown source.
  */
-export function dashWordJoiner(text: string): string {
-  const re = cachedRegExp(`(?<=[^\\s${WORD_JOINER}])[${EM_DASH}${EN_DASH}]`, "gu")
-  return text.replace(re, `${WORD_JOINER}$&`)
+export function dashWordJoiner(input: string): string
+export function dashWordJoiner(input: ProseView): void
+export function dashWordJoiner(input: string | ProseView): string | void {
+  return overInput(input, (view) => {
+    const re = cachedRegExp(`(?<=[^\\s${WORD_JOINER}])[${EM_DASH}${EN_DASH}]`, "gu")
+    replaceAllInView(view, re, (match) => `${WORD_JOINER}${match[0]}`)
+  })
+}
+
+/** Reconstructs `.replace()`-style callback args from a RegExpExecArray. */
+function matchArgs(match: RegExpExecArray): unknown[] {
+  return [...match, match.index, match.input, match.groups]
 }