perf(quartz): one pass for list search's other tags, ASCII whitespace fast path

- The `t` pass and the natural-language pass are one walk over the tags, so
  `t` values now come in tag order among the others (golden re-recorded; the
  set of fields per event is unchanged on all 1.5M corpus events).
- The extra-field walk is private; the two text builders share one join.
- Whitespace checks take an ASCII fast path instead of the JVM's two Unicode
  table lookups per character (identical results on every corpus value).

Corpus benchmark: classifier ~400 -> ~300 ms per 3.7M values, walk ~490 ->
~365 ms per 300k events.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016XJSf4ekFVuhSj5m8eeNsP

.claude/skills/searchable-events/references/searchable-kinds.md

@@ -65,7 +65,7 @@ Separator legend: **NL** = `joinToString("\n")`, **SP** = `joinToString(" ")`.
 | 9736 | Bolt12ZapEvent | nipB1Bolt12Zaps/zap | `content` |
 | 9737 | Bolt12ZapIntentEvent | nipB1Bolt12Zaps/intent | `content` |
 | 9802 | HighlightEvent | nip84Highlights | `listOfNotNull(comment(), context(), content)` NL |
-| 9998 | ListHeaderEvent | experimental/decentralizedLists/header | `tags.searchableListContent()` NL — `names` (singular, plural), `titles` (singular, plural), `name`, `title`, `description`, `comments`, then every `t` value that is not `isMachineValue`, then every value of every other tag (except `alt`, `client`, `imeta`) that passes `isNaturalLanguageValue` (has whitespace or non-ASCII, or is one capitalized letters-only word; never JSON, numbers, URIs of any scheme, addresses, hex ids, UUIDs, bech32) |
+| 9998 | ListHeaderEvent | experimental/decentralizedLists/header | `tags.searchableListContent()` NL — `names` (singular, plural), `titles` (singular, plural), `name`, `title`, `description`, `comments`, then in tag order every `t` value that is not `isMachineValue` and every value of every other tag (except `alt`, `client`, `imeta`) that passes `isNaturalLanguageValue` (has whitespace or non-ASCII, or is one capitalized letters-only word; never JSON, numbers, URIs of any scheme, addresses, hex ids, UUIDs, bech32) |
 | 9999 | ListItemEvent | experimental/decentralizedLists/item | same as 9998 |
 | 10003 | BookmarkListEvent | nip51Lists/bookmarkList | `listOfNotNull(title())` NL |
 | 10100 | AgentProfileEvent | buzz/agentProfiles | `profileOrNull()?.let { listOfNotNull(it.name, it.displayName).joinToString("\n") } ?: ""` |

quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/experimental/decentralizedLists/README.md

@@ -85,8 +85,8 @@ Filter(
 
 All four kinds are `SearchableEvent`s sharing one walk
 (`forEachSearchableListField`): `names` and `titles` (singular then plural),
-`name`, `title`, `description`, `comments`, then every `t` item value, then
-every value of every other tag that reads as natural language.
+`name`, `title`, `description`, `comments`, then, in tag order, every `t`
+item value and every value of every other tag that reads as natural language.
 
 The tag set is open — deployments add `author`, `subject`, `artist`,
 `relationshipType` and tags nobody has named yet — so that last step decides

quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/experimental/decentralizedLists/SearchExt.kt

@@ -43,11 +43,12 @@ import com.vitorpamplona.quartz.nip89AppHandlers.clientTag.ClientTag
 import com.vitorpamplona.quartz.nip92IMeta.IMetaTag
 
 /**
- * The human-authored text of any kind in the family, in a fixed order: the header's `names`
- * and `titles` (singular, then plural), the item's `name` and `title`, `description`,
- * `comments`, then each `t` value that is not a machine value, then
- * [forEachSearchableListExtraField]. Headers and items share one walk because the spec's
- * nonstandard method lets an item carry header tags; each kind simply has fewer of them set.
+ * The human-authored text of any kind in the family: first, in a fixed order, the header's `names`
+ * and `titles` (singular, then plural), the item's `name` and `title`, `description` and
+ * `comments`; then, in tag order, each `t` value that is not a machine value and every value of
+ * every other tag that reads as natural language ([forEachOtherField]). Headers and items share
+ * one walk because the spec's nonstandard method lets an item carry header tags; each kind simply
+ * has fewer of them set.
  *
  * `content` is not part of the spec and ids/pubkeys/coordinates are served by tag filters,
  * so neither is indexed.
@@ -57,7 +58,8 @@ import com.vitorpamplona.quartz.nip92IMeta.IMetaTag
 fun TagArray.forEachSearchableListField(visitor: IndexableFieldVisitor): Boolean {
     // The read path runs once per event per keystroke, so this is two allocation-free passes
     // instead of one full scan (and one parsed object) per field. The first pass only remembers
-    // the first well-formed tag of each field; the fixed visiting order is applied afterwards.
+    // the first well-formed tag of each named field, so their fixed order can be applied before
+    // the second pass visits everything else.
     var names: Array<String>? = null
     var titles: Array<String>? = null
     var name: String? = null
@@ -89,53 +91,55 @@ fun TagArray.forEachSearchableListField(visitor: IndexableFieldVisitor): Boolean
     title?.let { if (!visitor.visit(it)) return false }
     description?.let { if (!visitor.visit(it)) return false }
     comments?.let { if (!visitor.visit(it)) return false }
-    fastForEach { tag ->
-        HashtagTag.parse(tag)?.let { if (!isMachineValue(it) && !visitor.visit(it)) return false }
-    }
-    return forEachSearchableListExtraField(visitor)
+    return forEachOtherField(visitor, withHashtags = true)
 }
 
 /**
- * Every value of every other tag that reads as natural language ([isNaturalLanguageValue]), in
- * tag order. The family's tag set is open — deployments add `author`, `subject`, `artist`,
- * `relationshipType` and tags nobody has named yet — so this walk decides by what a value looks
- * like rather than by the tag it sits in.
+ * Every tag the first pass of [forEachSearchableListField] does not own, in tag order: each `t`
+ * value that is not [isMachineValue] (only [withHashtags]), and every value of every other tag
+ * that [isNaturalLanguageValue]. The family's tag set is open — deployments add `author`,
+ * `subject`, `artist`, `relationshipType` and tags nobody has named yet — so this decides by what
+ * a value looks like rather than by the tag it sits in.
  *
- * Skips the tags [forEachSearchableListField] already visits, plus the NIP-defined metadata tags
- * that are not the event's own text: `alt` (NIP-31 fallback text, which here restates `title`
- * and `artist` behind a fixed "Song: … by …" prefix), `client` (NIP-89 app name) and `imeta`
- * (NIP-92 `key value` pairs, whose space would otherwise pass every URL and hash as text).
+ * Skips the NIP-defined metadata tags that are not the event's own text: `alt` (NIP-31 fallback
+ * text, which here restates `title` and `artist` behind a fixed "Song: … by …" prefix), `client`
+ * (NIP-89 app name) and `imeta` (NIP-92 `key value` pairs, whose space would otherwise pass every
+ * URL and hash as text).
  *
  * @return false when the visitor stopped the walk.
  */
-fun TagArray.forEachSearchableListExtraField(visitor: IndexableFieldVisitor): Boolean {
+private fun TagArray.forEachOtherField(
+    visitor: IndexableFieldVisitor,
+    withHashtags: Boolean,
+): Boolean {
     fastForEach { tag ->
-        if (tag.size < 2 || !isExtraFieldTagName(tag[0])) return@fastForEach
-        for (i in 1 until tag.size) {
-            val value = tag[i]
-            if (isNaturalLanguageValue(value) && !visitor.visit(value)) return false
+        if (tag.size < 2) return@fastForEach
+        when (tag[0]) {
+            HashtagTag.TAG_NAME -> {
+                if (withHashtags && !isMachineValue(tag[1]) && !visitor.visit(tag[1])) return false
+            }
+
+            NamesTag.TAG_NAME,
+            TitlesTag.TAG_NAME,
+            NameTag.TAG_NAME,
+            TitleTag.TAG_NAME,
+            DescriptionTag.TAG_NAME,
+            CommentsTag.TAG_NAME,
+            AltTag.TAG_NAME,
+            ClientTag.TAG_NAME,
+            IMetaTag.TAG_NAME,
+            -> {}
+
+            else -> {
+                for (i in 1 until tag.size) {
+                    if (isNaturalLanguageValue(tag[i]) && !visitor.visit(tag[i])) return false
+                }
+            }
         }
     }
     return true
 }
 
-private fun isExtraFieldTagName(name: String) =
-    when (name) {
-        NamesTag.TAG_NAME,
-        TitlesTag.TAG_NAME,
-        NameTag.TAG_NAME,
-        TitleTag.TAG_NAME,
-        DescriptionTag.TAG_NAME,
-        CommentsTag.TAG_NAME,
-        HashtagTag.TAG_NAME,
-        AltTag.TAG_NAME,
-        ClientTag.TAG_NAME,
-        IMetaTag.TAG_NAME,
-        -> false
-
-        else -> true
-    }
-
 /**
  * The TITLE role of [forEachSearchableListField], for search engines that weight fields: what the
  * list or item is called — the header's `names` and `titles` (singular, then plural) and the
@@ -151,25 +155,18 @@ fun TagArray.searchableListTitles(): List<String?> {
 fun TagArray.searchableListDescriptions(): List<String?> = listOf(description(), comments())
 
 /**
- * The TEXT role of [forEachSearchableListField]: the [forEachSearchableListExtraField] values, one
- * per line, or null when the event has none. The `t` values are not here; the search funnel
- * carries them as hashtags.
+ * The TEXT role of [forEachSearchableListField]: the natural-language values of the tags the
+ * fixed-order fields do not own, one per line, or null when the event has none. The `t` values
+ * are not here; the search funnel carries them as hashtags.
  */
-fun TagArray.searchableListExtraText(): String? =
-    buildString {
-        forEachSearchableListExtraField { field ->
-            if (field != null) {
-                if (isNotEmpty()) append('\n')
-                append(field)
-            }
-            true
-        }
-    }.ifEmpty { null }
+fun TagArray.searchableListExtraText(): String? = joinFields { forEachOtherField(it, withHashtags = false) }.ifEmpty { null }
 
 /** The write-path join of [forEachSearchableListField]: one field per line. */
-fun TagArray.searchableListContent() =
+fun TagArray.searchableListContent() = joinFields { forEachSearchableListField(it) }
+
+private inline fun joinFields(walk: (IndexableFieldVisitor) -> Boolean) =
     buildString {
-        forEachSearchableListField { field ->
+        walk { field ->
             if (field != null) {
                 if (isNotEmpty()) append('\n')
                 append(field)

quartz/src/commonMain/kotlin/com/vitorpamplona/quartz/nip50Search/NaturalLanguageValue.kt

@@ -36,15 +36,15 @@ package com.vitorpamplona.quartz.nip50Search
 fun isNaturalLanguageValue(value: String): Boolean {
     var start = 0
     var end = value.length
-    while (start < end && value[start].isWhitespace()) start++
-    while (end > start && value[end - 1].isWhitespace()) end--
+    while (start < end && value[start].isSpace()) start++
+    while (end > start && value[end - 1].isSpace()) end--
     if (start == end || isStructuredValue(value, start, end)) return false
 
     var nonAscii = false
     for (i in start until end) {
         val c = value[i]
         // The token shapes in isMachineToken cannot contain whitespace.
-        if (c.isWhitespace()) return true
+        if (c.isSpace()) return true
         if (c.code > 127) nonAscii = true
     }
 
@@ -68,10 +68,10 @@ fun isNaturalLanguageValue(value: String): Boolean {
 fun isMachineValue(value: String): Boolean {
     var start = 0
     var end = value.length
-    while (start < end && value[start].isWhitespace()) start++
-    while (end > start && value[end - 1].isWhitespace()) end--
+    while (start < end && value[start].isSpace()) start++
+    while (end > start && value[end - 1].isSpace()) end--
     if (start == end || isStructuredValue(value, start, end)) return true
-    for (i in start until end) if (value[i].isWhitespace()) return false
+    for (i in start until end) if (value[i].isSpace()) return false
     return isMachineToken(value, start, end)
 }
 
@@ -99,6 +99,12 @@ private fun isMachineToken(
         isUuid(v, start, end) ||
         isBech32(v, start)
 
+/**
+ * [isWhitespace] with an ASCII fast path: on the JVM, [isWhitespace] is two Unicode table lookups
+ * per character, and nearly every character this file scans is ASCII.
+ */
+private fun Char.isSpace() = if (code < 0x80) this == ' ' || this in '\t'..'\r' || this in '\u001C'..'\u001F' else isWhitespace()
+
 private fun Char.isAsciiDigit() = this in '0'..'9'
 
 private fun Char.isAsciiLetter() = this in 'a'..'z' || this in 'A'..'Z'

quartz/src/jvmTest/resources/indexable-content.golden

@@ -58,8 +58,8 @@
 9736	The content body.
 9737	The content body.
 9802	The Comment\nThe Context\nThe content body.
-9998	The Name\nThe Title\nThe Description\nhashtag1\nhashtag2\nThe Subject\nThe Summary\nThe About\nThe Comment\nThe Context\nThe Rules\nThe Text\nThe Instructions
-9999	The Name\nThe Title\nThe Description\nhashtag1\nhashtag2\nThe Subject\nThe Summary\nThe About\nThe Comment\nThe Context\nThe Rules\nThe Text\nThe Instructions
+9998	The Name\nThe Title\nThe Description\nThe Subject\nThe Summary\nThe About\nThe Comment\nThe Context\nThe Rules\nThe Text\nThe Instructions\nhashtag1\nhashtag2
+9999	The Name\nThe Title\nThe Description\nThe Subject\nThe Summary\nThe About\nThe Comment\nThe Context\nThe Rules\nThe Text\nThe Instructions\nhashtag1\nhashtag2
 10003	The Title
 10100	
 10154	The Title\nThe Description
@@ -141,8 +141,8 @@
 39092	The Title\nThe Description
 39307	The content body.
 39701	The Title\nThe content body.
-39998	The Name\nThe Title\nThe Description\nhashtag1\nhashtag2\nThe Subject\nThe Summary\nThe About\nThe Comment\nThe Context\nThe Rules\nThe Text\nThe Instructions
-39999	The Name\nThe Title\nThe Description\nhashtag1\nhashtag2\nThe Subject\nThe Summary\nThe About\nThe Comment\nThe Context\nThe Rules\nThe Text\nThe Instructions
+39998	The Name\nThe Title\nThe Description\nThe Subject\nThe Summary\nThe About\nThe Comment\nThe Context\nThe Rules\nThe Text\nThe Instructions\nhashtag1\nhashtag2
+39999	The Name\nThe Title\nThe Description\nThe Subject\nThe Summary\nThe About\nThe Comment\nThe Context\nThe Rules\nThe Text\nThe Instructions\nhashtag1\nhashtag2
 40002	The content body.
 40100	The content body.
 45001	The content body.