diff --git a/vendor/golang.org/x/text/unicode/bidi/core.go b/vendor/golang.org/x/text/unicode/bidi/core.go
new file mode 100644
index 0000000..d4c1399
--- /dev/null
+++ b/vendor/golang.org/x/text/unicode/bidi/core.go
@@ -0,0 +1,1058 @@
+// Copyright 2015 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package bidi
+
+import "log"
+
+// This implementation is a port based on the reference implementation found at:
+// http://www.unicode.org/Public/PROGRAMS/BidiReferenceJava/
+//
+// described in Unicode Bidirectional Algorithm (UAX #9).
+//
+// Input:
+// There are two levels of input to the algorithm, since clients may prefer to
+// supply some information from out-of-band sources rather than relying on the
+// default behavior.
+//
+// - Bidi class array
+// - Bidi class array, with externally supplied base line direction
+//
+// Output:
+// Output is separated into several stages:
+//
+//  - levels array over entire paragraph
+//  - reordering array over entire paragraph
+//  - levels array over line
+//  - reordering array over line
+//
+// Note that for conformance to the Unicode Bidirectional Algorithm,
+// implementations are only required to generate correct reordering and
+// character directionality (odd or even levels) over a line. Generating
+// identical level arrays over a line is not required. Bidi explicit format
+// codes (LRE, RLE, LRO, RLO, PDF) and BN can be assigned arbitrary levels and
+// positions as long as the rest of the input is properly reordered.
+//
+// As the algorithm is defined to operate on a single paragraph at a time, this
+// implementation is written to handle single paragraphs. Thus rule P1 is
+// presumed by this implementation-- the data provided to the implementation is
+// assumed to be a single paragraph, and either contains no 'B' codes, or a
+// single 'B' code at the end of the input. 'B' is allowed as input to
+// illustrate how the algorithm assigns it a level.
+//
+// Also note that rules L3 and L4 depend on the rendering engine that uses the
+// result of the bidi algorithm. This implementation assumes that the rendering
+// engine expects combining marks in visual order (e.g. to the left of their
+// base character in RTL runs) and that it adjusts the glyphs used to render
+// mirrored characters that are in RTL runs so that they render appropriately.
+
+// level is the embedding level of a character. Even embedding levels indicate
+// left-to-right order and odd levels indicate right-to-left order. The special
+// level of -1 is reserved for undefined order.
+type level int8
+
+const implicitLevel level = -1
+
+// in returns if x is equal to any of the values in set.
+func (c Class) in(set ...Class) bool {
+	for _, s := range set {
+		if c == s {
+			return true
+		}
+	}
+	return false
+}
+
+// A paragraph contains the state of a paragraph.
+type paragraph struct {
+	initialTypes []Class
+
+	// Arrays of properties needed for paired bracket evaluation in N0
+	pairTypes  []bracketType // paired Bracket types for paragraph
+	pairValues []rune        // rune for opening bracket or pbOpen and pbClose; 0 for pbNone
+
+	embeddingLevel level // default: = implicitLevel;
+
+	// at the paragraph levels
+	resultTypes  []Class
+	resultLevels []level
+
+	// Index of matching PDI for isolate initiator characters. For other
+	// characters, the value of matchingPDI will be set to -1. For isolate
+	// initiators with no matching PDI, matchingPDI will be set to the length of
+	// the input string.
+	matchingPDI []int
+
+	// Index of matching isolate initiator for PDI characters. For other
+	// characters, and for PDIs with no matching isolate initiator, the value of
+	// matchingIsolateInitiator will be set to -1.
+	matchingIsolateInitiator []int
+}
+
+// newParagraph initializes a paragraph. The user needs to supply a few arrays
+// corresponding to the preprocessed text input. The types correspond to the
+// Unicode BiDi classes for each rune. pairTypes indicates the bracket type for
+// each rune. pairValues provides a unique bracket class identifier for each
+// rune (suggested is the rune of the open bracket for opening and matching
+// close brackets, after normalization). The embedding levels are optional, but
+// may be supplied to encode embedding levels of styled text.
+//
+// TODO: return an error.
+func newParagraph(types []Class, pairTypes []bracketType, pairValues []rune, levels level) *paragraph {
+	validateTypes(types)
+	validatePbTypes(pairTypes)
+	validatePbValues(pairValues, pairTypes)
+	validateParagraphEmbeddingLevel(levels)
+
+	p := &paragraph{
+		initialTypes:   append([]Class(nil), types...),
+		embeddingLevel: levels,
+
+		pairTypes:  pairTypes,
+		pairValues: pairValues,
+
+		resultTypes: append([]Class(nil), types...),
+	}
+	p.run()
+	return p
+}
+
+func (p *paragraph) Len() int { return len(p.initialTypes) }
+
+// The algorithm. Does not include line-based processing (Rules L1, L2).
+// These are applied later in the line-based phase of the algorithm.
+func (p *paragraph) run() {
+	p.determineMatchingIsolates()
+
+	// 1) determining the paragraph level
+	// Rule P1 is the requirement for entering this algorithm.
+	// Rules P2, P3.
+	// If no externally supplied paragraph embedding level, use default.
+	if p.embeddingLevel == implicitLevel {
+		p.embeddingLevel = p.determineParagraphEmbeddingLevel(0, p.Len())
+	}
+
+	// Initialize result levels to paragraph embedding level.
+	p.resultLevels = make([]level, p.Len())
+	setLevels(p.resultLevels, p.embeddingLevel)
+
+	// 2) Explicit levels and directions
+	// Rules X1-X8.
+	p.determineExplicitEmbeddingLevels()
+
+	// Rule X9.
+	// We do not remove the embeddings, the overrides, the PDFs, and the BNs
+	// from the string explicitly. But they are not copied into isolating run
+	// sequences when they are created, so they are removed for all
+	// practical purposes.
+
+	// Rule X10.
+	// Run remainder of algorithm one isolating run sequence at a time
+	for _, seq := range p.determineIsolatingRunSequences() {
+		// 3) resolving weak types
+		// Rules W1-W7.
+		seq.resolveWeakTypes()
+
+		// 4a) resolving paired brackets
+		// Rule N0
+		resolvePairedBrackets(seq)
+
+		// 4b) resolving neutral types
+		// Rules N1-N3.
+		seq.resolveNeutralTypes()
+
+		// 5) resolving implicit embedding levels
+		// Rules I1, I2.
+		seq.resolveImplicitLevels()
+
+		// Apply the computed levels and types
+		seq.applyLevelsAndTypes()
+	}
+
+	// Assign appropriate levels to 'hide' LREs, RLEs, LROs, RLOs, PDFs, and
+	// BNs. This is for convenience, so the resulting level array will have
+	// a value for every character.
+	p.assignLevelsToCharactersRemovedByX9()
+}
+
+// determineMatchingIsolates determines the matching PDI for each isolate
+// initiator and vice versa.
+//
+// Definition BD9.
+//
+// At the end of this function:
+//
+//  - The member variable matchingPDI is set to point to the index of the
+//    matching PDI character for each isolate initiator character. If there is
+//    no matching PDI, it is set to the length of the input text. For other
+//    characters, it is set to -1.
+//  - The member variable matchingIsolateInitiator is set to point to the
+//    index of the matching isolate initiator character for each PDI character.
+//    If there is no matching isolate initiator, or the character is not a PDI,
+//    it is set to -1.
+func (p *paragraph) determineMatchingIsolates() {
+	p.matchingPDI = make([]int, p.Len())
+	p.matchingIsolateInitiator = make([]int, p.Len())
+
+	for i := range p.matchingIsolateInitiator {
+		p.matchingIsolateInitiator[i] = -1
+	}
+
+	for i := range p.matchingPDI {
+		p.matchingPDI[i] = -1
+
+		if t := p.resultTypes[i]; t.in(LRI, RLI, FSI) {
+			depthCounter := 1
+			for j := i + 1; j < p.Len(); j++ {
+				if u := p.resultTypes[j]; u.in(LRI, RLI, FSI) {
+					depthCounter++
+				} else if u == PDI {
+					if depthCounter--; depthCounter == 0 {
+						p.matchingPDI[i] = j
+						p.matchingIsolateInitiator[j] = i
+						break
+					}
+				}
+			}
+			if p.matchingPDI[i] == -1 {
+				p.matchingPDI[i] = p.Len()
+			}
+		}
+	}
+}
+
+// determineParagraphEmbeddingLevel reports the resolved paragraph direction of
+// the substring limited by the given range [start, end).
+//
+// Determines the paragraph level based on rules P2, P3. This is also used
+// in rule X5c to find if an FSI should resolve to LRI or RLI.
+func (p *paragraph) determineParagraphEmbeddingLevel(start, end int) level {
+	var strongType Class = unknownClass
+
+	// Rule P2.
+	for i := start; i < end; i++ {
+		if t := p.resultTypes[i]; t.in(L, AL, R) {
+			strongType = t
+			break
+		} else if t.in(FSI, LRI, RLI) {
+			i = p.matchingPDI[i] // skip over to the matching PDI
+			if i > end {
+				log.Panic("assert (i <= end)")
+			}
+		}
+	}
+	// Rule P3.
+	switch strongType {
+	case unknownClass: // none found
+		// default embedding level when no strong types found is 0.
+		return 0
+	case L:
+		return 0
+	default: // AL, R
+		return 1
+	}
+}
+
+const maxDepth = 125
+
+// This stack will store the embedding levels and override and isolated
+// statuses
+type directionalStatusStack struct {
+	stackCounter        int
+	embeddingLevelStack [maxDepth + 1]level
+	overrideStatusStack [maxDepth + 1]Class
+	isolateStatusStack  [maxDepth + 1]bool
+}
+
+func (s *directionalStatusStack) empty()     { s.stackCounter = 0 }
+func (s *directionalStatusStack) pop()       { s.stackCounter-- }
+func (s *directionalStatusStack) depth() int { return s.stackCounter }
+
+func (s *directionalStatusStack) push(level level, overrideStatus Class, isolateStatus bool) {
+	s.embeddingLevelStack[s.stackCounter] = level
+	s.overrideStatusStack[s.stackCounter] = overrideStatus
+	s.isolateStatusStack[s.stackCounter] = isolateStatus
+	s.stackCounter++
+}
+
+func (s *directionalStatusStack) lastEmbeddingLevel() level {
+	return s.embeddingLevelStack[s.stackCounter-1]
+}
+
+func (s *directionalStatusStack) lastDirectionalOverrideStatus() Class {
+	return s.overrideStatusStack[s.stackCounter-1]
+}
+
+func (s *directionalStatusStack) lastDirectionalIsolateStatus() bool {
+	return s.isolateStatusStack[s.stackCounter-1]
+}
+
+// Determine explicit levels using rules X1 - X8
+func (p *paragraph) determineExplicitEmbeddingLevels() {
+	var stack directionalStatusStack
+	var overflowIsolateCount, overflowEmbeddingCount, validIsolateCount int
+
+	// Rule X1.
+	stack.push(p.embeddingLevel, ON, false)
+
+	for i, t := range p.resultTypes {
+		// Rules X2, X3, X4, X5, X5a, X5b, X5c
+		switch t {
+		case RLE, LRE, RLO, LRO, RLI, LRI, FSI:
+			isIsolate := t.in(RLI, LRI, FSI)
+			isRTL := t.in(RLE, RLO, RLI)
+
+			// override if this is an FSI that resolves to RLI
+			if t == FSI {
+				isRTL = (p.determineParagraphEmbeddingLevel(i+1, p.matchingPDI[i]) == 1)
+			}
+			if isIsolate {
+				p.resultLevels[i] = stack.lastEmbeddingLevel()
+				if stack.lastDirectionalOverrideStatus() != ON {
+					p.resultTypes[i] = stack.lastDirectionalOverrideStatus()
+				}
+			}
+
+			var newLevel level
+			if isRTL {
+				// least greater odd
+				newLevel = (stack.lastEmbeddingLevel() + 1) | 1
+			} else {
+				// least greater even
+				newLevel = (stack.lastEmbeddingLevel() + 2) &^ 1
+			}
+
+			if newLevel <= maxDepth && overflowIsolateCount == 0 && overflowEmbeddingCount == 0 {
+				if isIsolate {
+					validIsolateCount++
+				}
+				// Push new embedding level, override status, and isolated
+				// status.
+				// No check for valid stack counter, since the level check
+				// suffices.
+				switch t {
+				case LRO:
+					stack.push(newLevel, L, isIsolate)
+				case RLO:
+					stack.push(newLevel, R, isIsolate)
+				default:
+					stack.push(newLevel, ON, isIsolate)
+				}
+				// Not really part of the spec
+				if !isIsolate {
+					p.resultLevels[i] = newLevel
+				}
+			} else {
+				// This is an invalid explicit formatting character,
+				// so apply the "Otherwise" part of rules X2-X5b.
+				if isIsolate {
+					overflowIsolateCount++
+				} else { // !isIsolate
+					if overflowIsolateCount == 0 {
+						overflowEmbeddingCount++
+					}
+				}
+			}
+
+		// Rule X6a
+		case PDI:
+			if overflowIsolateCount > 0 {
+				overflowIsolateCount--
+			} else if validIsolateCount == 0 {
+				// do nothing
+			} else {
+				overflowEmbeddingCount = 0
+				for !stack.lastDirectionalIsolateStatus() {
+					stack.pop()
+				}
+				stack.pop()
+				validIsolateCount--
+			}
+			p.resultLevels[i] = stack.lastEmbeddingLevel()
+
+		// Rule X7
+		case PDF:
+			// Not really part of the spec
+			p.resultLevels[i] = stack.lastEmbeddingLevel()
+
+			if overflowIsolateCount > 0 {
+				// do nothing
+			} else if overflowEmbeddingCount > 0 {
+				overflowEmbeddingCount--
+			} else if !stack.lastDirectionalIsolateStatus() && stack.depth() >= 2 {
+				stack.pop()
+			}
+
+		case B: // paragraph separator.
+			// Rule X8.
+
+			// These values are reset for clarity, in this implementation B
+			// can only occur as the last code in the array.
+			stack.empty()
+			overflowIsolateCount = 0
+			overflowEmbeddingCount = 0
+			validIsolateCount = 0
+			p.resultLevels[i] = p.embeddingLevel
+
+		default:
+			p.resultLevels[i] = stack.lastEmbeddingLevel()
+			if stack.lastDirectionalOverrideStatus() != ON {
+				p.resultTypes[i] = stack.lastDirectionalOverrideStatus()
+			}
+		}
+	}
+}
+
+type isolatingRunSequence struct {
+	p *paragraph
+
+	indexes []int // indexes to the original string
+
+	types          []Class // type of each character using the index
+	resolvedLevels []level // resolved levels after application of rules
+	level          level
+	sos, eos       Class
+}
+
+func (i *isolatingRunSequence) Len() int { return len(i.indexes) }
+
+func maxLevel(a, b level) level {
+	if a > b {
+		return a
+	}
+	return b
+}
+
+// Rule X10, second bullet: Determine the start-of-sequence (sos) and end-of-sequence (eos) types,
+// 			 either L or R, for each isolating run sequence.
+func (p *paragraph) isolatingRunSequence(indexes []int) *isolatingRunSequence {
+	length := len(indexes)
+	types := make([]Class, length)
+	for i, x := range indexes {
+		types[i] = p.resultTypes[x]
+	}
+
+	// assign level, sos and eos
+	prevChar := indexes[0] - 1
+	for prevChar >= 0 && isRemovedByX9(p.initialTypes[prevChar]) {
+		prevChar--
+	}
+	prevLevel := p.embeddingLevel
+	if prevChar >= 0 {
+		prevLevel = p.resultLevels[prevChar]
+	}
+
+	var succLevel level
+	lastType := types[length-1]
+	if lastType.in(LRI, RLI, FSI) {
+		succLevel = p.embeddingLevel
+	} else {
+		// the first character after the end of run sequence
+		limit := indexes[length-1] + 1
+		for ; limit < p.Len() && isRemovedByX9(p.initialTypes[limit]); limit++ {
+
+		}
+		succLevel = p.embeddingLevel
+		if limit < p.Len() {
+			succLevel = p.resultLevels[limit]
+		}
+	}
+	level := p.resultLevels[indexes[0]]
+	return &isolatingRunSequence{
+		p:       p,
+		indexes: indexes,
+		types:   types,
+		level:   level,
+		sos:     typeForLevel(maxLevel(prevLevel, level)),
+		eos:     typeForLevel(maxLevel(succLevel, level)),
+	}
+}
+
+// Resolving weak types Rules W1-W7.
+//
+// Note that some weak types (EN, AN) remain after this processing is
+// complete.
+func (s *isolatingRunSequence) resolveWeakTypes() {
+
+	// on entry, only these types remain
+	s.assertOnly(L, R, AL, EN, ES, ET, AN, CS, B, S, WS, ON, NSM, LRI, RLI, FSI, PDI)
+
+	// Rule W1.
+	// Changes all NSMs.
+	preceedingCharacterType := s.sos
+	for i, t := range s.types {
+		if t == NSM {
+			s.types[i] = preceedingCharacterType
+		} else {
+			if t.in(LRI, RLI, FSI, PDI) {
+				preceedingCharacterType = ON
+			}
+			preceedingCharacterType = t
+		}
+	}
+
+	// Rule W2.
+	// EN does not change at the start of the run, because sos != AL.
+	for i, t := range s.types {
+		if t == EN {
+			for j := i - 1; j >= 0; j-- {
+				if t := s.types[j]; t.in(L, R, AL) {
+					if t == AL {
+						s.types[i] = AN
+					}
+					break
+				}
+			}
+		}
+	}
+
+	// Rule W3.
+	for i, t := range s.types {
+		if t == AL {
+			s.types[i] = R
+		}
+	}
+
+	// Rule W4.
+	// Since there must be values on both sides for this rule to have an
+	// effect, the scan skips the first and last value.
+	//
+	// Although the scan proceeds left to right, and changes the type
+	// values in a way that would appear to affect the computations
+	// later in the scan, there is actually no problem. A change in the
+	// current value can only affect the value to its immediate right,
+	// and only affect it if it is ES or CS. But the current value can
+	// only change if the value to its right is not ES or CS. Thus
+	// either the current value will not change, or its change will have
+	// no effect on the remainder of the analysis.
+
+	for i := 1; i < s.Len()-1; i++ {
+		t := s.types[i]
+		if t == ES || t == CS {
+			prevSepType := s.types[i-1]
+			succSepType := s.types[i+1]
+			if prevSepType == EN && succSepType == EN {
+				s.types[i] = EN
+			} else if s.types[i] == CS && prevSepType == AN && succSepType == AN {
+				s.types[i] = AN
+			}
+		}
+	}
+
+	// Rule W5.
+	for i, t := range s.types {
+		if t == ET {
+			// locate end of sequence
+			runStart := i
+			runEnd := s.findRunLimit(runStart, ET)
+
+			// check values at ends of sequence
+			t := s.sos
+			if runStart > 0 {
+				t = s.types[runStart-1]
+			}
+			if t != EN {
+				t = s.eos
+				if runEnd < len(s.types) {
+					t = s.types[runEnd]
+				}
+			}
+			if t == EN {
+				setTypes(s.types[runStart:runEnd], EN)
+			}
+			// continue at end of sequence
+			i = runEnd
+		}
+	}
+
+	// Rule W6.
+	for i, t := range s.types {
+		if t.in(ES, ET, CS) {
+			s.types[i] = ON
+		}
+	}
+
+	// Rule W7.
+	for i, t := range s.types {
+		if t == EN {
+			// set default if we reach start of run
+			prevStrongType := s.sos
+			for j := i - 1; j >= 0; j-- {
+				t = s.types[j]
+				if t == L || t == R { // AL's have been changed to R
+					prevStrongType = t
+					break
+				}
+			}
+			if prevStrongType == L {
+				s.types[i] = L
+			}
+		}
+	}
+}
+
+// 6) resolving neutral types Rules N1-N2.
+func (s *isolatingRunSequence) resolveNeutralTypes() {
+
+	// on entry, only these types can be in resultTypes
+	s.assertOnly(L, R, EN, AN, B, S, WS, ON, RLI, LRI, FSI, PDI)
+
+	for i, t := range s.types {
+		switch t {
+		case WS, ON, B, S, RLI, LRI, FSI, PDI:
+			// find bounds of run of neutrals
+			runStart := i
+			runEnd := s.findRunLimit(runStart, B, S, WS, ON, RLI, LRI, FSI, PDI)
+
+			// determine effective types at ends of run
+			var leadType, trailType Class
+
+			// Note that the character found can only be L, R, AN, or
+			// EN.
+			if runStart == 0 {
+				leadType = s.sos
+			} else {
+				leadType = s.types[runStart-1]
+				if leadType.in(AN, EN) {
+					leadType = R
+				}
+			}
+			if runEnd == len(s.types) {
+				trailType = s.eos
+			} else {
+				trailType = s.types[runEnd]
+				if trailType.in(AN, EN) {
+					trailType = R
+				}
+			}
+
+			var resolvedType Class
+			if leadType == trailType {
+				// Rule N1.
+				resolvedType = leadType
+			} else {
+				// Rule N2.
+				// Notice the embedding level of the run is used, not
+				// the paragraph embedding level.
+				resolvedType = typeForLevel(s.level)
+			}
+
+			setTypes(s.types[runStart:runEnd], resolvedType)
+
+			// skip over run of (former) neutrals
+			i = runEnd
+		}
+	}
+}
+
+func setLevels(levels []level, newLevel level) {
+	for i := range levels {
+		levels[i] = newLevel
+	}
+}
+
+func setTypes(types []Class, newType Class) {
+	for i := range types {
+		types[i] = newType
+	}
+}
+
+// 7) resolving implicit embedding levels Rules I1, I2.
+func (s *isolatingRunSequence) resolveImplicitLevels() {
+
+	// on entry, only these types can be in resultTypes
+	s.assertOnly(L, R, EN, AN)
+
+	s.resolvedLevels = make([]level, len(s.types))
+	setLevels(s.resolvedLevels, s.level)
+
+	if (s.level & 1) == 0 { // even level
+		for i, t := range s.types {
+			// Rule I1.
+			if t == L {
+				// no change
+			} else if t == R {
+				s.resolvedLevels[i] += 1
+			} else { // t == AN || t == EN
+				s.resolvedLevels[i] += 2
+			}
+		}
+	} else { // odd level
+		for i, t := range s.types {
+			// Rule I2.
+			if t == R {
+				// no change
+			} else { // t == L || t == AN || t == EN
+				s.resolvedLevels[i] += 1
+			}
+		}
+	}
+}
+
+// Applies the levels and types resolved in rules W1-I2 to the
+// resultLevels array.
+func (s *isolatingRunSequence) applyLevelsAndTypes() {
+	for i, x := range s.indexes {
+		s.p.resultTypes[x] = s.types[i]
+		s.p.resultLevels[x] = s.resolvedLevels[i]
+	}
+}
+
+// Return the limit of the run consisting only of the types in validSet
+// starting at index. This checks the value at index, and will return
+// index if that value is not in validSet.
+func (s *isolatingRunSequence) findRunLimit(index int, validSet ...Class) int {
+loop:
+	for ; index < len(s.types); index++ {
+		t := s.types[index]
+		for _, valid := range validSet {
+			if t == valid {
+				continue loop
+			}
+		}
+		return index // didn't find a match in validSet
+	}
+	return len(s.types)
+}
+
+// Algorithm validation. Assert that all values in types are in the
+// provided set.
+func (s *isolatingRunSequence) assertOnly(codes ...Class) {
+loop:
+	for i, t := range s.types {
+		for _, c := range codes {
+			if t == c {
+				continue loop
+			}
+		}
+		log.Panicf("invalid bidi code %v present in assertOnly at position %d", t, s.indexes[i])
+	}
+}
+
+// determineLevelRuns returns an array of level runs. Each level run is
+// described as an array of indexes into the input string.
+//
+// Determines the level runs. Rule X9 will be applied in determining the
+// runs, in the way that makes sure the characters that are supposed to be
+// removed are not included in the runs.
+func (p *paragraph) determineLevelRuns() [][]int {
+	run := []int{}
+	allRuns := [][]int{}
+	currentLevel := implicitLevel
+
+	for i := range p.initialTypes {
+		if !isRemovedByX9(p.initialTypes[i]) {
+			if p.resultLevels[i] != currentLevel {
+				// we just encountered a new run; wrap up last run
+				if currentLevel >= 0 { // only wrap it up if there was a run
+					allRuns = append(allRuns, run)
+					run = nil
+				}
+				// Start new run
+				currentLevel = p.resultLevels[i]
+			}
+			run = append(run, i)
+		}
+	}
+	// Wrap up the final run, if any
+	if len(run) > 0 {
+		allRuns = append(allRuns, run)
+	}
+	return allRuns
+}
+
+// Definition BD13. Determine isolating run sequences.
+func (p *paragraph) determineIsolatingRunSequences() []*isolatingRunSequence {
+	levelRuns := p.determineLevelRuns()
+
+	// Compute the run that each character belongs to
+	runForCharacter := make([]int, p.Len())
+	for i, run := range levelRuns {
+		for _, index := range run {
+			runForCharacter[index] = i
+		}
+	}
+
+	sequences := []*isolatingRunSequence{}
+
+	var currentRunSequence []int
+
+	for _, run := range levelRuns {
+		first := run[0]
+		if p.initialTypes[first] != PDI || p.matchingIsolateInitiator[first] == -1 {
+			currentRunSequence = nil
+			// int run = i;
+			for {
+				// Copy this level run into currentRunSequence
+				currentRunSequence = append(currentRunSequence, run...)
+
+				last := currentRunSequence[len(currentRunSequence)-1]
+				lastT := p.initialTypes[last]
+				if lastT.in(LRI, RLI, FSI) && p.matchingPDI[last] != p.Len() {
+					run = levelRuns[runForCharacter[p.matchingPDI[last]]]
+				} else {
+					break
+				}
+			}
+			sequences = append(sequences, p.isolatingRunSequence(currentRunSequence))
+		}
+	}
+	return sequences
+}
+
+// Assign level information to characters removed by rule X9. This is for
+// ease of relating the level information to the original input data. Note
+// that the levels assigned to these codes are arbitrary, they're chosen so
+// as to avoid breaking level runs.
+func (p *paragraph) assignLevelsToCharactersRemovedByX9() {
+	for i, t := range p.initialTypes {
+		if t.in(LRE, RLE, LRO, RLO, PDF, BN) {
+			p.resultTypes[i] = t
+			p.resultLevels[i] = -1
+		}
+	}
+	// now propagate forward the levels information (could have
+	// propagated backward, the main thing is not to introduce a level
+	// break where one doesn't already exist).
+
+	if p.resultLevels[0] == -1 {
+		p.resultLevels[0] = p.embeddingLevel
+	}
+	for i := 1; i < len(p.initialTypes); i++ {
+		if p.resultLevels[i] == -1 {
+			p.resultLevels[i] = p.resultLevels[i-1]
+		}
+	}
+	// Embedding information is for informational purposes only so need not be
+	// adjusted.
+}
+
+//
+// Output
+//
+
+// getLevels computes levels array breaking lines at offsets in linebreaks.
+// Rule L1.
+//
+// The linebreaks array must include at least one value. The values must be
+// in strictly increasing order (no duplicates) between 1 and the length of
+// the text, inclusive. The last value must be the length of the text.
+func (p *paragraph) getLevels(linebreaks []int) []level {
+	// Note that since the previous processing has removed all
+	// P, S, and WS values from resultTypes, the values referred to
+	// in these rules are the initial types, before any processing
+	// has been applied (including processing of overrides).
+	//
+	// This example implementation has reinserted explicit format codes
+	// and BN, in order that the levels array correspond to the
+	// initial text. Their final placement is not normative.
+	// These codes are treated like WS in this implementation,
+	// so they don't interrupt sequences of WS.
+
+	validateLineBreaks(linebreaks, p.Len())
+
+	result := append([]level(nil), p.resultLevels...)
+
+	// don't worry about linebreaks since if there is a break within
+	// a series of WS values preceding S, the linebreak itself
+	// causes the reset.
+	for i, t := range p.initialTypes {
+		if t.in(B, S) {
+			// Rule L1, clauses one and two.
+			result[i] = p.embeddingLevel
+
+			// Rule L1, clause three.
+			for j := i - 1; j >= 0; j-- {
+				if isWhitespace(p.initialTypes[j]) { // including format codes
+					result[j] = p.embeddingLevel
+				} else {
+					break
+				}
+			}
+		}
+	}
+
+	// Rule L1, clause four.
+	start := 0
+	for _, limit := range linebreaks {
+		for j := limit - 1; j >= start; j-- {
+			if isWhitespace(p.initialTypes[j]) { // including format codes
+				result[j] = p.embeddingLevel
+			} else {
+				break
+			}
+		}
+		start = limit
+	}
+
+	return result
+}
+
+// getReordering returns the reordering of lines from a visual index to a
+// logical index for line breaks at the given offsets.
+//
+// Lines are concatenated from left to right. So for example, the fifth
+// character from the left on the third line is
+//
+// 		getReordering(linebreaks)[linebreaks[1] + 4]
+//
+// (linebreaks[1] is the position after the last character of the second
+// line, which is also the index of the first character on the third line,
+// and adding four gets the fifth character from the left).
+//
+// The linebreaks array must include at least one value. The values must be
+// in strictly increasing order (no duplicates) between 1 and the length of
+// the text, inclusive. The last value must be the length of the text.
+func (p *paragraph) getReordering(linebreaks []int) []int {
+	validateLineBreaks(linebreaks, p.Len())
+
+	return computeMultilineReordering(p.getLevels(linebreaks), linebreaks)
+}
+
+// Return multiline reordering array for a given level array. Reordering
+// does not occur across a line break.
+func computeMultilineReordering(levels []level, linebreaks []int) []int {
+	result := make([]int, len(levels))
+
+	start := 0
+	for _, limit := range linebreaks {
+		tempLevels := make([]level, limit-start)
+		copy(tempLevels, levels[start:])
+
+		for j, order := range computeReordering(tempLevels) {
+			result[start+j] = order + start
+		}
+		start = limit
+	}
+	return result
+}
+
+// Return reordering array for a given level array. This reorders a single
+// line. The reordering is a visual to logical map. For example, the
+// leftmost char is string.charAt(order[0]). Rule L2.
+func computeReordering(levels []level) []int {
+	result := make([]int, len(levels))
+	// initialize order
+	for i := range result {
+		result[i] = i
+	}
+
+	// locate highest level found on line.
+	// Note the rules say text, but no reordering across line bounds is
+	// performed, so this is sufficient.
+	highestLevel := level(0)
+	lowestOddLevel := level(maxDepth + 2)
+	for _, level := range levels {
+		if level > highestLevel {
+			highestLevel = level
+		}
+		if level&1 != 0 && level < lowestOddLevel {
+			lowestOddLevel = level
+		}
+	}
+
+	for level := highestLevel; level >= lowestOddLevel; level-- {
+		for i := 0; i < len(levels); i++ {
+			if levels[i] >= level {
+				// find range of text at or above this level
+				start := i
+				limit := i + 1
+				for limit < len(levels) && levels[limit] >= level {
+					limit++
+				}
+
+				for j, k := start, limit-1; j < k; j, k = j+1, k-1 {
+					result[j], result[k] = result[k], result[j]
+				}
+				// skip to end of level run
+				i = limit
+			}
+		}
+	}
+
+	return result
+}
+
+// isWhitespace reports whether the type is considered a whitespace type for the
+// line break rules.
+func isWhitespace(c Class) bool {
+	switch c {
+	case LRE, RLE, LRO, RLO, PDF, LRI, RLI, FSI, PDI, BN, WS:
+		return true
+	}
+	return false
+}
+
+// isRemovedByX9 reports whether the type is one of the types removed in X9.
+func isRemovedByX9(c Class) bool {
+	switch c {
+	case LRE, RLE, LRO, RLO, PDF, BN:
+		return true
+	}
+	return false
+}
+
+// typeForLevel reports the strong type (L or R) corresponding to the level.
+func typeForLevel(level level) Class {
+	if (level & 0x1) == 0 {
+		return L
+	}
+	return R
+}
+
+// TODO: change validation to not panic
+
+func validateTypes(types []Class) {
+	if len(types) == 0 {
+		log.Panic("types is null")
+	}
+	for i, t := range types[:len(types)-1] {
+		if t == B {
+			log.Panicf("B type before end of paragraph at index: %d", i)
+		}
+	}
+}
+
+func validateParagraphEmbeddingLevel(embeddingLevel level) {
+	if embeddingLevel != implicitLevel &&
+		embeddingLevel != 0 &&
+		embeddingLevel != 1 {
+		log.Panicf("illegal paragraph embedding level: %d", embeddingLevel)
+	}
+}
+
+func validateLineBreaks(linebreaks []int, textLength int) {
+	prev := 0
+	for i, next := range linebreaks {
+		if next <= prev {
+			log.Panicf("bad linebreak: %d at index: %d", next, i)
+		}
+		prev = next
+	}
+	if prev != textLength {
+		log.Panicf("last linebreak was %d, want %d", prev, textLength)
+	}
+}
+
+func validatePbTypes(pairTypes []bracketType) {
+	if len(pairTypes) == 0 {
+		log.Panic("pairTypes is null")
+	}
+	for i, pt := range pairTypes {
+		switch pt {
+		case bpNone, bpOpen, bpClose:
+		default:
+			log.Panicf("illegal pairType value at %d: %v", i, pairTypes[i])
+		}
+	}
+}
+
+func validatePbValues(pairValues []rune, pairTypes []bracketType) {
+	if pairValues == nil {
+		log.Panic("pairValues is null")
+	}
+	if len(pairTypes) != len(pairValues) {
+		log.Panic("pairTypes is different length from pairValues")
+	}
+}