Pre-push review: T-1910/bugfix-m4-fixture-work-matching

diff --git a/Packages/AsterismCore/Sources/AsterismCore/URLIdentityPlanner.swift b/Packages/AsterismCore/Sources/AsterismCore/URLIdentityPlanner.swiftindex a6c9498..c435ae4 100644--- a/Packages/AsterismCore/Sources/AsterismCore/URLIdentityPlanner.swift+++ b/Packages/AsterismCore/Sources/AsterismCore/URLIdentityPlanner.swift@@ -337,6 +337,66 @@ public struct IdentityFirstWorkCandidate: Equatable, Sendable {   } } +/// One Site's Work candidates, validated once and bucketed for repeated+/// matching.+///+/// `match` and `identityReuse` take a plain array, and each call re-validates+/// the whole array before linear-scanning it. That is right for a single+/// capture, and wrong for a teaching projection: the projection calls them once+/// per Entry over one unchanging candidate list, so at 5,000 Entries × 1,000+/// Works it re-validates 5M candidates and re-scans 5M more, per preview, on+/// every keystroke (T-1910). Building this once hoists both out of the loop.+///+/// The buckets are sorted by the same `candidateOrder` the scanning arms apply,+/// so an outcome taken from a bucket is the outcome the scan would have+/// produced — including the member order of an `.ambiguous`.+public struct IdentityFirstWorkCandidateIndex: Sendable {+  /// The candidates as supplied, for the counts the planner logs.+  public let candidates: [IdentityFirstWorkCandidate]+  /// Every candidate whose exact matching title is the key.+  fileprivate let byMatchingTitle: [ExactScalarString: [IdentityFirstWorkCandidate]]+  /// Every candidate retaining a *rule*-derived identity equal to the key. The+  /// evidence half of `isIdentityMatch` is still checked per candidate; only+  /// the tuple half is what the key can settle.+  fileprivate let byRetainedRuleIdentity: [ExactScalarString: [IdentityFirstWorkCandidate]]++  /// Validates the candidates exactly as `match` does, then indexes them.+  /// Throws the same `IdentityFirstWorkPlanningError` values, so a caller that+  /// builds the index in place of its first `match` call reports the same+  /// refusal it reported before.+  public init(validating candidates: [IdentityFirstWorkCandidate]) throws {+    try IdentityFirstWorkMatchingPlanner.validate(candidates)+    self.candidates = candidates++    var byTitle: [ExactScalarString: [IdentityFirstWorkCandidate]] = [:]+    var byIdentity: [ExactScalarString: [IdentityFirstWorkCandidate]] = [:]+    for candidate in candidates {+      byTitle[candidate.matchingTitle, default: []].append(candidate)+      if candidate.previousIdentity.state == .rule, let value = candidate.previousIdentity.value {+        byIdentity[value, default: []].append(candidate)+      }+    }+    // Sorting once per bucket is the same total work as sorting the matches of+    // one call, and it is paid once rather than per Entry.+    self.byMatchingTitle = byTitle.mapValues {+      $0.count == 1 ? $0 : $0.sorted(by: IdentityFirstWorkMatchingPlanner.candidateOrder)+    }+    self.byRetainedRuleIdentity = byIdentity.mapValues {+      $0.count == 1 ? $0 : $0.sorted(by: IdentityFirstWorkMatchingPlanner.candidateOrder)+    }+  }++  fileprivate func titleMatches(_ title: ExactScalarString) -> [IdentityFirstWorkCandidate] {+    byMatchingTitle[title] ?? []+  }++  fileprivate func retainedRuleIdentityMatches(+    _ identity: ExactScalarString+  ) -> [IdentityFirstWorkCandidate] {+    byRetainedRuleIdentity[identity] ?? []+  }+}+ public enum ProspectiveWorkKey: Equatable, Hashable, Sendable {   case urlIdentity(ExactScalarString)   case title(ExactScalarString)@@ -372,36 +432,43 @@ public enum IdentityFirstWorkMatchingPlanner {     parsedTitle: ExactScalarString,     candidates: [IdentityFirstWorkCandidate]   ) throws -> IdentityFirstWorkMatchOutcome {-    guard !parsedTitle.isBlank else { throw IdentityFirstWorkPlanningError.blankParsedTitle }-    if let extractedIdentity, extractedIdentity.isBlank {-      throw IdentityFirstWorkPlanningError.blankExtractedIdentity-    }-    try validate(candidates)+    // The input guards run before the candidates are validated, as they always+    // have: a blank title or identity is reported as such even when the+    // candidate array is also invalid.+    try requireNonBlank(extractedIdentity: extractedIdentity, parsedTitle: parsedTitle)+    return try match(+      extractedIdentity: extractedIdentity, parsedTitle: parsedTitle,+      index: IdentityFirstWorkCandidateIndex(validating: candidates))+  }++  /// `match` against candidates already validated and indexed. The answer is+  /// identical; what it saves is the per-call revalidation and full scan, which+  /// a caller matching many Entries against one candidate list pays once+  /// instead of once per Entry (T-1910).+  public static func match(+    extractedIdentity: ExactScalarString?,+    parsedTitle: ExactScalarString,+    index: IdentityFirstWorkCandidateIndex+  ) throws -> IdentityFirstWorkMatchOutcome {+    try requireNonBlank(extractedIdentity: extractedIdentity, parsedTitle: parsedTitle)      guard let extractedIdentity else {-      let titleMatches = candidates-        .filter { $0.matchingTitle == parsedTitle }-        .sorted(by: candidateOrder)       let outcome = existingTitleOutcome(-        matches: titleMatches,+        matches: index.titleMatches(parsedTitle),         createKey: .title(parsedTitle)       )-      logger.debug("Used exact-title fallback across \(candidates.count) Work candidates")+      logger.debug("Used exact-title fallback across \(index.candidates.count) Work candidates")       return outcome     } -    switch identityMatchOutcome(extractedIdentity, among: candidates) {+    switch identityMatchOutcome(extractedIdentity, among: index) {     case .reuse(let workID): return .reuse(workID: workID)     case .ambiguous(let workIDs): return .ambiguous(workIDs: workIDs)     case nil: break     } -    let claimMatches = candidates-      .filter {-        $0.matchingTitle == parsedTitle-          && isClaimEligible($0, identity: extractedIdentity)-      }-      .sorted(by: candidateOrder)+    let claimMatches = index.titleMatches(parsedTitle)+      .filter { isClaimEligible($0, identity: extractedIdentity) }     if claimMatches.count == 1, let match = claimMatches.first {       logger.debug("Claimed one nil-identity Work by exact title and complete evidence")       return .claim(workID: match.id)@@ -424,18 +491,41 @@ public enum IdentityFirstWorkMatchingPlanner {     extractedIdentity: ExactScalarString,     candidates: [IdentityFirstWorkCandidate]   ) throws -> IdentityReuseOutcome? {+    // Input guard before candidate validation, as in `match`.     guard !extractedIdentity.isBlank else { throw IdentityFirstWorkPlanningError.blankExtractedIdentity }-    try validate(candidates)-    return identityMatchOutcome(extractedIdentity, among: candidates)+    return try identityReuse(+      extractedIdentity: extractedIdentity,+      index: IdentityFirstWorkCandidateIndex(validating: candidates))+  }++  /// `identityReuse` against candidates already validated and indexed. See the+  /// index-taking `match`.+  public static func identityReuse(+    extractedIdentity: ExactScalarString,+    index: IdentityFirstWorkCandidateIndex+  ) throws -> IdentityReuseOutcome? {+    guard !extractedIdentity.isBlank else { throw IdentityFirstWorkPlanningError.blankExtractedIdentity }+    return identityMatchOutcome(extractedIdentity, among: index)+  }++  private static func requireNonBlank(+    extractedIdentity: ExactScalarString?, parsedTitle: ExactScalarString+  ) throws {+    guard !parsedTitle.isBlank else { throw IdentityFirstWorkPlanningError.blankParsedTitle }+    if let extractedIdentity, extractedIdentity.isBlank {+      throw IdentityFirstWorkPlanningError.blankExtractedIdentity+    }   }    private static func identityMatchOutcome(     _ extractedIdentity: ExactScalarString,-    among candidates: [IdentityFirstWorkCandidate]+    among index: IdentityFirstWorkCandidateIndex   ) -> IdentityReuseOutcome? {-    let identityMatches = candidates+    // The bucket already holds every candidate whose *tuple* carries this+    // rule-derived identity, in `candidateOrder`; what is left is the evidence+    // half of the same predicate, and filtering preserves the order.+    let identityMatches = index.retainedRuleIdentityMatches(extractedIdentity)       .filter { isIdentityMatch($0, identity: extractedIdentity) }-      .sorted(by: candidateOrder)     if identityMatches.count == 1, let match = identityMatches.first {       logger.debug("Reused one Work by complete URL identity evidence")       return .reuse(workID: match.id)@@ -483,7 +573,7 @@ public enum IdentityFirstWorkMatchingPlanner {     }   } -  private static func validate(_ candidates: [IdentityFirstWorkCandidate]) throws {+  fileprivate static func validate(_ candidates: [IdentityFirstWorkCandidate]) throws {     var candidateIDs: Set<UUID> = []     for candidate in candidates {       guard candidateIDs.insert(candidate.id).inserted else {@@ -587,7 +677,7 @@ public enum IdentityFirstWorkMatchingPlanner {     }   } -  private static func candidateOrder(+  fileprivate static func candidateOrder(     _ lhs: IdentityFirstWorkCandidate,     _ rhs: IdentityFirstWorkCandidate   ) -> Bool {

diff --git a/Packages/AsterismCore/Sources/AsterismCore/TeachingTypes.swift b/Packages/AsterismCore/Sources/AsterismCore/TeachingTypes.swiftindex e239486..0fbd535 100644--- a/Packages/AsterismCore/Sources/AsterismCore/TeachingTypes.swift+++ b/Packages/AsterismCore/Sources/AsterismCore/TeachingTypes.swift@@ -166,36 +166,85 @@ public struct WorkMatchCandidate: Equatable, Sendable, Hashable {     } } +/// One Site's Work candidates bucketed by their exact matching title.+///+/// `WorkMatcher.match` scans every candidate per call. A teaching projection+/// calls it once per Entry over one unchanging candidate list, so at 5,000+/// Entries × 1,000 Works that is 5M scalar-string comparisons per preview, on+/// every keystroke (T-1910). Building this once hoists the scan out of the loop.+///+/// The key is `ExactScalarString`, whose `==` and `hash(into:)` both work on+/// Unicode scalars — the same equality the scan applied. Keying on `String`+/// would silently switch the matcher to canonical equivalence.+public struct WorkMatchIndex: Sendable {+    fileprivate let byMatchingTitle: [ExactScalarString: [WorkMatchCandidate]]++    public init(candidates: [WorkMatchCandidate]) {+        var buckets: [ExactScalarString: [WorkMatchCandidate]] = [:]+        for candidate in candidates {+            buckets[ExactScalarString(WorkMatcher.matchingTitle(of: candidate)), default: []]+                .append(candidate)+        }+        // Pre-sorted into the order an `.ambiguous` reports, so reading a bucket+        // gives the answer the scan gave.+        byMatchingTitle = buckets.mapValues {+            $0.count == 1 ? $0 : $0.sorted { $0.id.uuidString < $1.id.uuidString }+        }+    }++    fileprivate func matches(_ title: String) -> [WorkMatchCandidate] {+        byMatchingTitle[ExactScalarString(title)] ?? []+    }+}+ /// Performs exact scalar Work matching per Requirements 3.4-3.7. public enum WorkMatcher {      /// Match a parsed Work title against candidates using exact scalar equality.     /// Uses `lastParsedTitle` when present and nonblank, otherwise `displayTitle`.+    ///+    /// Deliberately still a scan rather than a call through `WorkMatchIndex`: a+    /// one-shot match pays only a comparison per candidate, and building an+    /// index for it would cost a hash and an allocation per candidate instead.+    /// The index is for callers that match *many* titles against one list.     public static func match(         parsedWorkTitle: String,         candidates: [WorkMatchCandidate]     ) -> WorkMatchOutcome {         var matchingCandidates: [WorkMatchCandidate] = []-        for candidate in candidates {-            let matchTitle: String-            if let lpt = candidate.lastParsedTitle, !M2Unicode.isBlank(lpt) {-                matchTitle = lpt-            } else {-                matchTitle = candidate.displayTitle-            }-            if scalarEqual(matchTitle, parsedWorkTitle) {-                matchingCandidates.append(candidate)-            }+        for candidate in candidates where scalarEqual(matchingTitle(of: candidate), parsedWorkTitle) {+            matchingCandidates.append(candidate)         }-         switch matchingCandidates.count {         case 0: return .create(parsedWorkTitle: parsedWorkTitle)         case 1: return .reuse(workID: matchingCandidates[0].id)         default:             // Sort by UUID string for deterministic order.-            let sorted = matchingCandidates.sorted { $0.id.uuidString < $1.id.uuidString }-            return .ambiguous(candidates: sorted)+            return .ambiguous(candidates: matchingCandidates.sorted { $0.id.uuidString < $1.id.uuidString })+        }+    }++    /// `match` against candidates already indexed. Identical answers; what it+    /// saves is the per-call scan (T-1910).+    public static func match(+        parsedWorkTitle: String,+        index: WorkMatchIndex+    ) -> WorkMatchOutcome {+        let matchingCandidates = index.matches(parsedWorkTitle)+        switch matchingCandidates.count {+        case 0: return .create(parsedWorkTitle: parsedWorkTitle)+        case 1: return .reuse(workID: matchingCandidates[0].id)+        default: return .ambiguous(candidates: matchingCandidates)+        }+    }++    /// The title a candidate matches on: `lastParsedTitle` when present and+    /// nonblank, otherwise `displayTitle`.+    static func matchingTitle(of candidate: WorkMatchCandidate) -> String {+        if let lastParsed = candidate.lastParsedTitle, !M2Unicode.isBlank(lastParsed) {+            return lastParsed         }+        return candidate.displayTitle     }      /// Exact Unicode scalar equality without normalization.

diff --git a/Packages/AsterismCore/Sources/AsterismCore/ComposedTeachingProjection.swift b/Packages/AsterismCore/Sources/AsterismCore/ComposedTeachingProjection.swiftindex ed826f6..174243b 100644--- a/Packages/AsterismCore/Sources/AsterismCore/ComposedTeachingProjection.swift+++ b/Packages/AsterismCore/Sources/AsterismCore/ComposedTeachingProjection.swift@@ -456,12 +456,19 @@ public enum ComposedTeachingProjectionPlanner {          // Per-Entry assignment (identity-first with title fallback) and the         // eligible create consumers for prospective-Work batching.-        let identityCandidates = identityBearing-            ? try identityFirstCandidates(basis: basis, evidenceByWorkID: evidenceByWorkID)-            : []-        let titleCandidates = basis.works.map {+        //+        // Both candidate sets are **indexed once, here** (T-1910). The matchers+        // also take plain arrays, and doing that would re-validate and re-scan+        // the Site's whole Work list once per Entry: 5,000 Entries × 1,000 Works+        // measured 5.81 s against Req 8.5's 1 s preview budget, for work whose+        // answer cannot change across the loop.+        let identityIndex = try IdentityFirstWorkCandidateIndex(+            validating: identityBearing+                ? identityFirstCandidates(basis: basis, evidenceByWorkID: evidenceByWorkID)+                : [])+        let titleIndex = WorkMatchIndex(candidates: basis.works.map {             WorkMatchCandidate(id: $0.id, lastParsedTitle: $0.lastParsedTitle, displayTitle: $0.displayTitle)-        }+        })          var entryProjections: [ComposedEntryProjection] = []         var prospectiveEntries: [ProspectiveWorkEntry] = []@@ -469,7 +476,7 @@ public enum ComposedTeachingProjectionPlanner {             guard let derivation = derivations[entry.id] else { continue }             let assignment = try assign(                 entry: entry, derivation: derivation, identityBearing: identityBearing,-                identityCandidates: identityCandidates, titleCandidates: titleCandidates)+                identityIndex: identityIndex, titleIndex: titleIndex)             let projection = makeProjection(entry: entry, derivation: derivation, assignment: assignment)             entryProjections.append(projection) @@ -627,7 +634,7 @@ public enum ComposedTeachingProjectionPlanner {      private static func assign(         entry: ComposedEntryBasis, derivation: ComposedDerivation, identityBearing: Bool,-        identityCandidates: [IdentityFirstWorkCandidate], titleCandidates: [WorkMatchCandidate]+        identityIndex: IdentityFirstWorkCandidateIndex, titleIndex: WorkMatchIndex     ) throws -> ComposedAssignmentProjection {         if entry.intentionallyUnattached || entry.workAssignmentProvenance == .manual {             return .protected@@ -638,7 +645,7 @@ public enum ComposedTeachingProjectionPlanner {         // of optional-chapter-sequence), so the guard sits after this.         if identityBearing, let identity = derivation.workIdentity,            let reuse = try IdentityFirstWorkMatchingPlanner.identityReuse(-               extractedIdentity: identity, candidates: identityCandidates) {+               extractedIdentity: identity, index: identityIndex) {             switch reuse {             case .reuse(let workID): return .reuse(workID: workID)             case .ambiguous(let workIDs): return .ambiguous(workIDs: workIDs)@@ -650,7 +657,7 @@ public enum ComposedTeachingProjectionPlanner {         if identityBearing {             let outcome = try IdentityFirstWorkMatchingPlanner.match(                 extractedIdentity: derivation.workIdentity,-                parsedTitle: ExactScalarString(name), candidates: identityCandidates)+                parsedTitle: ExactScalarString(name), index: identityIndex)             switch outcome {             case .reuse(let workID): return .reuse(workID: workID)             case .claim(let workID): return .claim(workID: workID)@@ -658,7 +665,7 @@ public enum ComposedTeachingProjectionPlanner {             case .ambiguous(let workIDs): return .ambiguous(workIDs: workIDs)             }         }-        switch WorkMatcher.match(parsedWorkTitle: name, candidates: titleCandidates) {+        switch WorkMatcher.match(parsedWorkTitle: name, index: titleIndex) {         case .reuse(let workID): return .reuse(workID: workID)         case .create(let title): return .create(key: .title(ExactScalarString(title)))         case .ambiguous(let candidates): return .ambiguous(workIDs: candidates.map(\.id))

diff --git a/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixture.swift b/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixture.swiftindex 279b6d2..fac4aef 100644--- a/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixture.swift+++ b/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixture.swift@@ -15,6 +15,24 @@ import Foundation ///   "collapsed to title-only" preview requests (Req 8.5). /// - `captureBasis` / `captureRequests` → a taught `CaptureBasis` and a batch of ///   new-chapter captures for the capture rule-application budget (Req 6.5, Q9).+///+/// **Work-bearing bases (T-1910).** `composedBasis(entryCount:)` carries no+/// Works, because an untaught Site has none — it is the *initial* teaching+/// sweep. Every Work-matching arm of `ComposedTeachingProjectionPlanner` is+/// therefore unreachable from it, and those arms are the expensive ones. The+/// re-teaching shape, where the Site already holds its 1,000 Works and every+/// Entry re-matches against all of them, is built by:+/// - `titleMatchBasis(entryCount:)` → taught, sequence-only URL rule, 1,000+///   title candidates: the `WorkMatcher` arm.+/// - `identityMatchBasis(entryCount:)` + `identityRequest` → taught, an+///   identity-bearing URL rule, 1,000 candidates each already carrying the+///   rule-derived identity its Entries extract: the+///   `IdentityFirstWorkMatchingPlanner` arms.+///+/// They are deliberately separate members rather than a change to+/// `composedBasis(entryCount:)`: the existing preview arms keep measuring the+/// initial-teaching shape they were banded against, so no other suite over this+/// fixture shifts. struct M4ScaleFixture {     let hostname = LibraryRepository.m4FixtureHostname     let entryCount = LibraryRepository.m4FixtureEntryCount@@ -54,6 +72,124 @@ struct M4ScaleFixture {             workAssignmentProvenance: .none, intentionallyUnattached: false)     } +    // MARK: - Work-bearing (re-teaching) preview bases — T-1910++    /// The URL rule id the taught bases cite. Works carrying a rule-derived+    /// identity must reference a rule row, and `deriveURLIdentity` takes the+    /// basis's own current rule id, so both sides have to name the same UUID.+    static let urlRuleID = LibraryRepository.m4FixtureUUID(namespace: 2, index: 0)++    /// The per-Work slug the identity-bearing URL rule extracts.+    static func workSlug(workIndex: Int) -> String { "story-\(workIndex)" }++    /// A raw URL carrying the Work slug *and* the chapter sequence, so an+    /// identity-bearing rule has something to extract. `uniqueBy` keeps every+    /// Entry's immutable raw URL distinct without affecting either field.+    static func identityRawURL(+        hostname: String, workIndex: Int, chapter: Int, uniqueBy: Int+    ) -> String {+        "https://\(hostname)/read?story=\(workSlug(workIndex: workIndex))&chapter=\(chapter)&e=\(uniqueBy)"+    }++    /// `.workAndSequence`, the simplest identity-bearing form (`suppliesIdentity`+    /// is true for every M3 form and false only for the sequence-only rule the+    /// rest of this fixture teaches).+    static func identityURLDefinition() -> URLRuleDefinition {+        .workAndSequence(+            work: URLFieldSelector(locator: .query(name: ExactScalarString("story"))),+            sequence: URLFieldSelector(locator: .query(name: ExactScalarString("chapter"))))+    }++    /// The composed request whose URL rule supplies a Work identity, so the+    /// projection takes the identity-first assignment arms.+    var identityRequest: ComposedTeachingRequest {+        ComposedTeachingRequest(+            titleDefinition: LibraryRepository.m4FixtureTitleDefinition(),+            trimPrefix: LibraryRepository.m4FixtureTrimPrefix,+            trimSuffix: LibraryRepository.m4FixtureTrimSuffix,+            urlDefinition: Self.identityURLDefinition())+    }++    /// A taught basis over the sequence-only URL rule: `entryCount` Entries and+    /// the Site's full 1,000 Works, none of them carrying a URL identity. The+    /// projection's title-matching arm (`WorkMatcher.match`) then runs once per+    /// Entry against all 1,000 candidates.+    ///+    /// The Work list is always the whole Site, whatever `entryCount` is: a+    /// single-Entry edit acknowledgement re-matches against every Work the Site+    /// holds, which is exactly the keystroke-path cost this measures.+    func titleMatchBasis(entryCount: Int) -> ComposedTeachingBasis {+        ComposedTeachingBasis(+            siteMode: .taught, hostname: hostname,+            entries: (0..<entryCount).map { taughtEntryBasis(entryIndex: $0, identityBearing: false) },+            works: (0..<workCount).map { workBasis(workIndex: $0, identityBearing: false) },+            currentTitleRule: titleRuleBasis,+            currentURLRule: urlRuleBasis(LibraryRepository.m4FixtureURLDefinition()))+    }++    /// A taught basis over an identity-bearing URL rule: `entryCount` Entries+    /// whose URLs carry their Work's slug, and the Site's full 1,000 Works, each+    /// already retaining the identity its own Entries extract. Every Entry then+    /// reaches `IdentityFirstWorkMatchingPlanner` and reuses its Work by+    /// identity — the steady state of a taught, identity-bearing Site.+    func identityMatchBasis(entryCount: Int) -> ComposedTeachingBasis {+        ComposedTeachingBasis(+            siteMode: .taught, hostname: hostname,+            entries: (0..<entryCount).map { taughtEntryBasis(entryIndex: $0, identityBearing: true) },+            works: (0..<workCount).map { workBasis(workIndex: $0, identityBearing: true) },+            currentTitleRule: titleRuleBasis,+            currentURLRule: urlRuleBasis(Self.identityURLDefinition()))+    }++    private var titleRuleBasis: ComposedTitleRuleBasis {+        ComposedTitleRuleBasis(+            id: LibraryRepository.m4FixtureUUID(namespace: 1, index: 0), version: 1,+            definition: LibraryRepository.m4FixtureTitleDefinition(),+            trimPrefix: LibraryRepository.m4FixtureTrimPrefix,+            trimSuffix: LibraryRepository.m4FixtureTrimSuffix)+    }++    private func urlRuleBasis(_ definition: URLRuleDefinition) -> ComposedURLRuleBasis {+        ComposedURLRuleBasis(+            id: Self.urlRuleID, version: 1, origin: .readerTaught, definition: definition)+    }++    /// An Entry already assigned to its Work with pattern provenance — the shape+    /// a taught Site's rows carry. Pattern provenance is deliberate: `.manual`+    /// would make `assign` return `.protected` and skip every matching arm this+    /// basis exists to reach.+    private func taughtEntryBasis(entryIndex: Int, identityBearing: Bool) -> ComposedEntryBasis {+        let workIndex = entryIndex / LibraryRepository.m4FixtureEntriesPerWork+        let chapter = entryIndex % LibraryRepository.m4FixtureEntriesPerWork + 1+        let rawURL = identityBearing+            ? Self.identityRawURL(+                hostname: hostname, workIndex: workIndex, chapter: chapter, uniqueBy: entryIndex)+            : LibraryRepository.m4FixtureRawURL(+                hostname: hostname, chapter: chapter, uniqueBy: entryIndex)+        return ComposedEntryBasis(+            id: LibraryRepository.m4FixtureUUID(namespace: 11, index: entryIndex),+            captureTitle: LibraryRepository.m4FixtureCaptureTitle(workIndex: workIndex, chapter: chapter),+            rawURLString: rawURL,+            hostname: hostname,+            firstCapturedAt: Date(timeIntervalSince1970: TimeInterval(entryIndex)),+            chapterTitle: nil, chapterTitleProvenance: .none,+            workID: LibraryRepository.m4FixtureUUID(namespace: 10, index: workIndex),+            workAssignmentProvenance: .pattern, intentionallyUnattached: false)+    }++    private func workBasis(workIndex: Int, identityBearing: Bool) -> ComposedWorkBasis {+        let name = LibraryRepository.m4FixtureWorkTitle(workIndex: workIndex)+        let identity: WorkIdentitySnapshot = identityBearing+            ? WorkIdentitySnapshot(+                value: ExactScalarString(Self.workSlug(workIndex: workIndex)),+                state: .rule, ruleReference: URLRuleReference(id: Self.urlRuleID))+            : .none+        return ComposedWorkBasis(+            id: LibraryRepository.m4FixtureUUID(namespace: 10, index: workIndex),+            displayTitle: name, lastParsedTitle: name,+            titleProvenance: .parsed, identity: identity)+    }+     // MARK: - Capture rule-application basis      /// The number of Works the taught capture basis carries (5,000 / 5).

diff --git a/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixtureTests.swift b/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixtureTests.swiftindex a7af7db..d0a81be 100644--- a/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixtureTests.swift+++ b/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixtureTests.swift@@ -55,6 +55,71 @@ struct M4ScaleFixtureTests {         #expect(outcome.entries.first?.projectedKeyVersion == 3)     } +    // MARK: - Work-bearing (re-teaching) preview bases — T-1910++    /// The bug this pins: the budgeted basis has to *reach* the Work-matching+    /// arms. `composedBasis` carries no Works at all, so nothing in+    /// `WorkMatcher` or `IdentityFirstWorkMatchingPlanner` runs over it — the+    /// arms the 1-second preview budget exists to bound were unmeasured.+    @Test("The initial-teaching basis reaches no Work-matching arm at all")+    func initialTeachingBasisHasNoWorkCandidates() throws {+        let fixture = M4ScaleFixture()+        #expect(fixture.composedBasis(entryCount: fixture.entryCount).works.isEmpty)+    }++    @Test("The title-matching basis reuses all 1,000 existing Works by exact title")+    func titleMatchBasisReusesEveryWork() throws {+        let fixture = M4ScaleFixture()+        let basis = fixture.titleMatchBasis(entryCount: fixture.entryCount)+        #expect(basis.works.count == fixture.workCount)+        let outcome = try ComposedTeachingProjectionPlanner.project(+            basis: basis, request: fixture.composedRequest)++        #expect(outcome.entries.count == fixture.entryCount)+        // Every Entry matched an existing Work by title, so nothing is+        // prospective and every projected Work ID is one the basis supplied.+        #expect(outcome.prospectiveWorks.isEmpty)+        let workIDs = Set(basis.works.map(\.id))+        #expect(outcome.entries.allSatisfy { $0.projectedWorkID.map(workIDs.contains) == true })+        #expect(outcome.entries.allSatisfy { $0.projectedKeyVersion == 3 })+    }++    @Test("The identity-matching basis reuses all 1,000 existing Works by URL identity")+    func identityMatchBasisReusesEveryWork() throws {+        let fixture = M4ScaleFixture()+        let basis = fixture.identityMatchBasis(entryCount: fixture.entryCount)+        let request = fixture.identityRequest+        #expect(request.urlDefinition?.suppliesIdentity == true)++        let outcome = try ComposedTeachingProjectionPlanner.project(basis: basis, request: request)+        #expect(outcome.entries.count == fixture.entryCount)+        #expect(outcome.prospectiveWorks.isEmpty)+        #expect(outcome.issues.isEmpty)+        // Identity-bearing rules derive per-Work evidence, so the projection+        // reports one Work row per candidate — the shape the identity-first+        // matching arms consume.+        #expect(outcome.works.count == fixture.workCount)+        let workIDs = Set(basis.works.map(\.id))+        #expect(outcome.entries.allSatisfy { $0.projectedWorkID.map(workIDs.contains) == true })+    }++    @Test("A single-Entry Work-bearing preview still matches against all 1,000 Works")+    func editAckBasesCarryEveryWork() throws {+        let fixture = M4ScaleFixture()+        for basis in [+            fixture.titleMatchBasis(entryCount: 1), fixture.identityMatchBasis(entryCount: 1)+        ] {+            #expect(basis.entries.count == 1)+            #expect(basis.works.count == fixture.workCount)+        }+        let title = try ComposedTeachingProjectionPlanner.project(+            basis: fixture.titleMatchBasis(entryCount: 1), request: fixture.composedRequest)+        #expect(title.entries.first?.projectedWorkID != nil)+        let identity = try ComposedTeachingProjectionPlanner.project(+            basis: fixture.identityMatchBasis(entryCount: 1), request: fixture.identityRequest)+        #expect(identity.entries.first?.projectedWorkID != nil)+    }+     // MARK: - Capture rule application      @Test("Capture rule application derives a v3 key and reuses the title-matched Work")

diff --git a/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScalePerformanceTests.swift b/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScalePerformanceTests.swiftindex 9134f65..e6e62be 100644--- a/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScalePerformanceTests.swift+++ b/Packages/AsterismCore/Tests/AsterismCoreTests/M4ScalePerformanceTests.swift@@ -98,6 +98,63 @@ struct M4ScalePerformanceTests {         expectWithinBudget("complete-preview-collapsed", measured, completePreviewBudget)     } +    // MARK: - Preview budgets over a Site that already holds its Works (T-1910)++    /// The four arms above all project a basis with **no Works** — the initial+    /// teaching sweep of an untaught Site, which by construction has none. That+    /// left every Work-matching arm of `ComposedTeachingProjectionPlanner`+    /// outside the budget the arms exist to guard, and those are the expensive+    /// ones: matching runs once per Entry against every Work the Site holds.+    ///+    /// The four arms below project the re-teaching shape instead — the same+    /// 5,000 Entries over the Site's own 1,000 Works — so the same Req 8.5+    /// budgets bound the title-matching arm (`WorkMatcher`) and the+    /// identity-first arms (`IdentityFirstWorkMatchingPlanner`).++    @Test("Edit acknowledgement ≤ 100 ms — title matching against 1,000 Works")+    func editAckTitleMatching() throws {+        let fixture = M4ScaleFixture()+        let basis = fixture.titleMatchBasis(entryCount: 1)+        let request = fixture.composedRequest+        let measured = try measureDistribution(iterations: iterations) {+            _ = try ComposedTeachingProjectionPlanner.project(basis: basis, request: request)+        }+        expectWithinBudget("edit-ack-title-matching", measured, editAckBudget)+    }++    @Test("Edit acknowledgement ≤ 100 ms — identity-first matching against 1,000 Works")+    func editAckIdentityMatching() throws {+        let fixture = M4ScaleFixture()+        let basis = fixture.identityMatchBasis(entryCount: 1)+        let request = fixture.identityRequest+        let measured = try measureDistribution(iterations: iterations) {+            _ = try ComposedTeachingProjectionPlanner.project(basis: basis, request: request)+        }+        expectWithinBudget("edit-ack-identity-matching", measured, editAckBudget)+    }++    @Test("Complete 5,000-Entry preview ≤ 1 s — title matching against 1,000 Works")+    func completePreviewTitleMatching() throws {+        let fixture = M4ScaleFixture()+        let basis = fixture.titleMatchBasis(entryCount: fixture.entryCount)+        let request = fixture.composedRequest+        let measured = try measureDistribution(iterations: iterations) {+            _ = try ComposedTeachingProjectionPlanner.project(basis: basis, request: request)+        }+        expectWithinBudget("complete-preview-title-matching", measured, completePreviewBudget)+    }++    @Test("Complete 5,000-Entry preview ≤ 1 s — identity-first matching against 1,000 Works")+    func completePreviewIdentityMatching() throws {+        let fixture = M4ScaleFixture()+        let basis = fixture.identityMatchBasis(entryCount: fixture.entryCount)+        let request = fixture.identityRequest+        let measured = try measureDistribution(iterations: iterations) {+            _ = try ComposedTeachingProjectionPlanner.project(basis: basis, request: request)+        }+        expectWithinBudget("complete-preview-identity-matching", measured, completePreviewBudget)+    }+     // MARK: - Capture rule-application budget (Req 6.5, Q9)      @Test("Capture rule-application ≤ 100 ms against the composed fixture")

diff --git a/Packages/AsterismCore/Tests/AsterismCoreTests/URLIdentityPlannerTests.swift b/Packages/AsterismCore/Tests/AsterismCoreTests/URLIdentityPlannerTests.swiftindex 2ecaf46..0182dbf 100644--- a/Packages/AsterismCore/Tests/AsterismCoreTests/URLIdentityPlannerTests.swift+++ b/Packages/AsterismCore/Tests/AsterismCoreTests/URLIdentityPlannerTests.swift@@ -364,6 +364,44 @@ struct IdentityFirstWorkMatchingTests {     #expect(intents.contains { $0.key == .title(ExactScalarString("Latest winner")) })   } +  @Test("Blank inputs are reported before the candidate array is validated")+  func blankInputsBeforeCandidateValidation() throws {+    // The array-taking forms build an index, which validates the candidates.+    // The input guards still run first, so a blank title or identity is what+    // a caller hears about even when its candidates are also invalid (T-1910).+    let duplicated = candidate(+      1,+      title: "Series 42",+      identity: .none,+      evidence: .noEntries(previousIdentity: .none)+    )+    let invalidCandidates = [duplicated, duplicated]+    #expect(throws: IdentityFirstWorkPlanningError.duplicateCandidateID(duplicated.id)) {+      try IdentityFirstWorkCandidateIndex(validating: invalidCandidates)+    }++    #expect(throws: IdentityFirstWorkPlanningError.blankParsedTitle) {+      try IdentityFirstWorkMatchingPlanner.match(+        extractedIdentity: ExactScalarString("series-42"),+        parsedTitle: ExactScalarString("   "),+        candidates: invalidCandidates+      )+    }+    #expect(throws: IdentityFirstWorkPlanningError.blankExtractedIdentity) {+      try IdentityFirstWorkMatchingPlanner.match(+        extractedIdentity: ExactScalarString(""),+        parsedTitle: ExactScalarString("Series 42"),+        candidates: invalidCandidates+      )+    }+    #expect(throws: IdentityFirstWorkPlanningError.blankExtractedIdentity) {+      try IdentityFirstWorkMatchingPlanner.identityReuse(+        extractedIdentity: ExactScalarString(" "),+        candidates: invalidCandidates+      )+    }+  }+   private func candidate(     _ suffix: Int,     title: String,

diff --git a/Packages/AsterismCore/Tests/AsterismCoreTests/TitleParsingTests.swift b/Packages/AsterismCore/Tests/AsterismCoreTests/TitleParsingTests.swiftindex 4ab92ea..ea7bf1b 100644--- a/Packages/AsterismCore/Tests/AsterismCoreTests/TitleParsingTests.swift+++ b/Packages/AsterismCore/Tests/AsterismCoreTests/TitleParsingTests.swift@@ -556,6 +556,34 @@ struct WorkMatchingTests {         let outcome = WorkMatcher.match(parsedWorkTitle: "Cafe\u{0301}", candidates: candidates)         #expect(outcome == .create(parsedWorkTitle: "Cafe\u{0301}"))     }++    @Test("Indexed matching answers as the scan does, ambiguous member order included")+    func indexMatchesScan() {+        // Two Works share a title, a third does not; supplied in the reverse+        // of the UUID order an `.ambiguous` reports (T-1910).+        let low = UUID(uuidString: "00000000-0000-0000-0000-000000000001")!+        let high = UUID(uuidString: "00000000-0000-0000-0000-000000000002")!+        let other = UUID(uuidString: "00000000-0000-0000-0000-000000000003")!+        let candidates = [+            WorkMatchCandidate(id: high, lastParsedTitle: "Fiction Name", displayTitle: "Fiction Name (B)"),+            WorkMatchCandidate(id: other, lastParsedTitle: nil, displayTitle: "Other Name"),+            WorkMatchCandidate(id: low, lastParsedTitle: "   ", displayTitle: "Fiction Name"),+        ]+        let index = WorkMatchIndex(candidates: candidates)++        for title in ["Fiction Name", "Other Name", "Fiction Name (B)", "Cafe\u{0301}"] {+            #expect(+                WorkMatcher.match(parsedWorkTitle: title, index: index)+                    == WorkMatcher.match(parsedWorkTitle: title, candidates: candidates))+        }+        #expect(+            WorkMatcher.match(parsedWorkTitle: "Fiction Name", index: index)+                == .ambiguous(candidates: [candidates[2], candidates[0]]))+        #expect(WorkMatcher.match(parsedWorkTitle: "Other Name", index: index) == .reuse(workID: other))+        #expect(+            WorkMatcher.match(parsedWorkTitle: "Fiction Name (B)", index: index)+                == .create(parsedWorkTitle: "Fiction Name (B)"))+    } }  // MARK: - Actionability tests

diff --git a/Packages/AsterismCore/Tests/AsterismCoreTests/ComposedTeachingProjectionTests.swift b/Packages/AsterismCore/Tests/AsterismCoreTests/ComposedTeachingProjectionTests.swiftindex fd22cf9..564dba4 100644--- a/Packages/AsterismCore/Tests/AsterismCoreTests/ComposedTeachingProjectionTests.swift+++ b/Packages/AsterismCore/Tests/AsterismCoreTests/ComposedTeachingProjectionTests.swift@@ -247,6 +247,33 @@ struct ComposedTeachingProjectionTests {         #expect(p2.assignment == .protected)     } +    @Test("An invalid Work list is refused even when every Entry is protected")+    func invalidWorksRefusedWithProtectedEntries() throws {+        // Candidate validation is eager — once, before the per-Entry loop — so+        // a Work no matcher would ever be asked about is still refused. Before+        // T-1910 this library projected, because validation ran lazily inside+        // the first match call and protected Entries never make one.+        let blankWork = UUID(), e1 = UUID()+        let basis = ComposedTeachingBasis(+            siteMode: .untaught, hostname: host,+            entries: [+                entry(e1, title: "Chapter 7 - Real Work", url: "https://ex.com/read?id=42&chapter=7",+                      at: 1, workID: blankWork, assignmentProv: .manual),+            ],+            works: [ComposedWorkBasis(+                id: blankWork, displayTitle: "   ", lastParsedTitle: nil,+                titleProvenance: .manual, identity: .none)],+            currentTitleRule: nil, currentURLRule: nil)+        let request = ComposedTeachingRequest(+            titleDefinition: try wcSegment(), urlDefinition: identitySequenceURL())++        #expect(throws: IdentityFirstWorkPlanningError.invalidCandidate(+            id: blankWork, reason: "matching title is blank")+        ) {+            try ComposedTeachingProjectionPlanner.project(basis: basis, request: request)+        }+    }+     // MARK: - Failure template: S-Site title parse failure (Q24)      @Test("Sequence-only Site with a title parse failure: conservative key, sequence evidence, actionable name")

diff --git a/CLAUDE.md b/CLAUDE.mdindex 32fee26..921d519 100644--- a/CLAUDE.md+++ b/CLAUDE.md@@ -58,7 +58,7 @@ invocations where a target exists. - `make test-quick` — unit-test bundle only (simulator), preceded by `build-mac`: a macOS compile failure fails it (Req 9.1). The Mac build is never installed or launched. `SKIP_MAC=1` drops that dependency loudly and owes a clean `make build-mac` before the push. - `make test` / `make test-ui` — full suites (simulator, iPhone); they skip the iPad-only suites by name - `make test-ui-ipad` — the wide-layout and wide-layout-accessibility suites on `IPAD_SIMULATOR` (simulator, safe)-- `make test-performance-m4` — M4 Core budgets, host only, no device, safe to run. **~21 minutes** (1,093 s of test time measured 2026-08-28, and 1,120 s over 28 tests on 2026-08-30 after `character-ranking` added its own, plus a ~190 s release build): most of it is the worst-case single-hostname consolidation in `M4ScalePerformanceTests` (5 samples, each paying its own ~40 s divert before a ~40 s measurement) and the Req 10.1 settling pass (10 samples, each re-seeding 1,350 duplicate rows plus an untimed observation pass). The V4→V5 migration measurement is **gone** — `retire-migration-chain` deleted the pass it timed along with the suite. **The target exits 0**, with the accepted breaches reported as `withKnownIssue` known issues rather than failures — **eight** since `drop-superseded-columns` (four before `multi-site-works`, nine after it). Four are long-standing: Req 10.1's settling pass and Req 5.5's three diagnosis re-derivations. Three are Req 5.4's capture-projection arms (0.093–0.102 s pre-V8 → 0.160–0.170 s at V9 → **0.169–0.176 s at V10**, the one on a path the reader waits on; the three new `Work` columns and the wider `orderComponents` cost them 3–6%, still well inside a 250 ms ceiling). The eighth is the **full**-tier no-op reconcile, and V9 recovered most of it: 1.07 s → **0.0296–0.0302 s** once `V8PopulationPass` was deleted with the columns and `MembershipReconciler.heal` was gated on the diagnosis, which is still 3.0× a 10 ms ceiling drawn before the library had a membership table; V10 left it there (0.0301 s). Req 10.1's *observation* pass **retired** with that fall (2.69 s → 1.01 s, back inside its 2 s budget, and 1.02 s at V10). Every one has a regression ceiling asserted *outside* its known-issue block, so a run that drifts further still fails; `RUNS=3` completes all three runs. See `specs/work-and-reading-status/verification-run.md` §4 for the current numbers, `specs/drop-superseded-columns/verification-run.md` and `specs/multi-site-works/verification-run.md` §4 and §7 for the previous ones, and `docs/agent-notes/testing.md` for recording a band.+- `make test-performance-m4` — M4 Core budgets, host only, no device, safe to run. **~21 minutes** (1,093 s of test time measured 2026-08-28, and 1,120 s over 28 tests on 2026-08-30 after `character-ranking` added its own, and 1,380 s over 32 tests on 2026-09-05 after T-1910 added four Work-bearing preview arms, plus a ~190 s release build): most of it is the worst-case single-hostname consolidation in `M4ScalePerformanceTests` (5 samples, each paying its own ~40 s divert before a ~40 s measurement) and the Req 10.1 settling pass (10 samples, each re-seeding 1,350 duplicate rows plus an untimed observation pass). The V4→V5 migration measurement is **gone** — `retire-migration-chain` deleted the pass it timed along with the suite. **The target exits 0**, with the accepted breaches reported as `withKnownIssue` known issues rather than failures — **eight** since `drop-superseded-columns` (four before `multi-site-works`, nine after it). Four are long-standing: Req 10.1's settling pass and Req 5.5's three diagnosis re-derivations. Three are Req 5.4's capture-projection arms (0.093–0.102 s pre-V8 → 0.160–0.170 s at V9 → **0.169–0.176 s at V10**, the one on a path the reader waits on; the three new `Work` columns and the wider `orderComponents` cost them 3–6%, still well inside a 250 ms ceiling). The eighth is the **full**-tier no-op reconcile, and V9 recovered most of it: 1.07 s → **0.0296–0.0302 s** once `V8PopulationPass` was deleted with the columns and `MembershipReconciler.heal` was gated on the diagnosis, which is still 3.0× a 10 ms ceiling drawn before the library had a membership table; V10 left it there (0.0301 s). Req 10.1's *observation* pass **retired** with that fall (2.69 s → 1.01 s, back inside its 2 s budget, and 1.02 s at V10). Every one has a regression ceiling asserted *outside* its known-issue block, so a run that drifts further still fails; `RUNS=3` completes all three runs. See `specs/work-and-reading-status/verification-run.md` §4 for the current numbers, `specs/drop-superseded-columns/verification-run.md` and `specs/multi-site-works/verification-run.md` §4 and §7 for the previous ones, and `docs/agent-notes/testing.md` for recording a band. - `make test-performance-chunks` — host-only calibration sweep of the shared bulk chunk constant (import commits and the reconciler re-pin). No device, safe to run, but gated on `ASTERISM_RUN_CHUNK_SWEEP=1` and **~20 minutes per run**, so it is deliberately *not* part of `make test-performance-m4`. It asserts nothing — a calibration is reported, not budgeted. Re-run it when the bulk write paths change (Q53 and the task 25 section of `specs/cloudkit-mirroring/implementation.md`). - `make test-performance-m4-recent` — **physical device, see above** 

diff --git a/docs/agent-notes/testing.md b/docs/agent-notes/testing.mdindex be5cd09..a6897af 100644--- a/docs/agent-notes/testing.md+++ b/docs/agent-notes/testing.md@@ -220,7 +220,8 @@ Consequences: added the membership suite (1,494 s measured 2026-08-26), and **~21 minutes** since `drop-superseded-columns` (1,093 s of test time measured 2026-08-28, and 1,120 s over 28 tests measured 2026-08-30, after `character-ranking` added its-own).+own, and 1,380 s over **32** tests on 2026-09-05, after T-1910 added four+Work-bearing preview arms). It reported **four or five known issues** — Req 10.1's settling pass (`duplicate-reconciliation` Decision 27), Req 5.5's three diagnosis re-derivations (`library-integrity-tolerance` Decision 11), and,@@ -265,6 +266,16 @@ Zero would mean the filter or the opt-in gate stopped the suites from running at all, which is the failure mode the Makefile's no-xcbeautify comment exists for. +**A scale fixture can be silently missing the shape a budget guards.** T-1910:+`M4ScaleFixture.composedBasis` built its 5,000-Entry preview basis with+`works: []`, so every Work-matching arm of `ComposedTeachingProjectionPlanner`+sat outside the Req 8.5 budget that exists to bound it. Once the Work-bearing+re-teaching shape was given a basis, the identity-bearing preview measured+**5.82 s against a 1 s budget**. The lesson generalises: a green budget over a+fixture nobody re-read after the code gained a second path proves nothing about+that path. `M4ScaleFixtureTests` now pins that the budgeted bases actually reach+the arms — assert reachability, not only timings.+ **A budget that leaves its band is not always a code regression — check the fixture first.** Task 22 of `multi-site-works` found `reconcile-worst-case-consolidation` failing at 5,000 of 6,000 records re-pinned

diff --git a/specs/bugfixes/m4-fixture-work-matching/report.md b/specs/bugfixes/m4-fixture-work-matching/report.mdnew file mode 100644index 0000000..740f842--- /dev/null+++ b/specs/bugfixes/m4-fixture-work-matching/report.md@@ -0,0 +1,245 @@+# Bugfix Report: M4 performance fixture doesn't exercise Work matching++**Date:** 2026-09-05+**Status:** Fixed+**Ticket:** T-1910++## Description of the Issue++`M4ScaleFixture.composedBasis(entryCount:)` builds its 5,000-Entry composed+teaching basis with `works: []` and a sequence-only URL rule. Every+Work-matching path in `ComposedTeachingProjectionPlanner` was therefore+unreachable from the four preview arms of `M4ScalePerformanceTests` — the arms+that exist to guard Req 8.5 (edit acknowledgement p95 ≤ 100 ms, complete preview+p95 ≤ 1 s).++The budget was green because it measured a shape that skips the expensive work,+not because the expensive work fits inside it.++**Reproduction steps:**++1. Read `M4ScaleFixture.composedBasis(entryCount:)`: `works: []`.+2. Read `ComposedTeachingProjectionPlanner.project`: with no Works both the+   identity-first candidate list and the title candidate list are empty, so+   `assign` returns `.create` for every Entry without ever scanning a candidate.+3. Run `make test-performance-m4`: `complete-preview-expanded` reports 0.083 s+   against a 1 s budget, and no measurement anywhere covers the case where the+   Site already holds its 1,000 Works.++**Impact:** a silent gap in a stated requirement's guard. The unmeasured paths+are the ones a reader actually pays for — re-teaching a taught Site runs Work+matching once per Entry over every Work the Site holds, on every+keystroke-triggered preview. Measured for the first time here, the+identity-bearing shape took **5.82 s median against a 1 s budget**.++## Investigation Summary++- **Symptoms examined:** the fixture's basis shape versus the code paths the+  budget claims to bound.+- **Code inspected:**+  - `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixture.swift`+  - `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScalePerformanceTests.swift`+  - `Packages/AsterismCore/Sources/AsterismCore/ComposedTeachingProjection.swift`+    (`project`, `identityFirstCandidates`, `assign`)+  - `Packages/AsterismCore/Sources/AsterismCore/URLIdentityPlanner.swift`+    (`IdentityFirstWorkMatchingPlanner.match` / `identityReuse` / `validate`)+  - `Packages/AsterismCore/Sources/AsterismCore/TeachingTypes.swift`+    (`WorkMatcher.match`)+- **Hypotheses tested:**+  - *Could the existing arms reach the matching code at all?* No — `basis.works`+    is empty, so both candidate arrays are empty and every arm short-circuits.+  - *Does the capture budget already cover it?* No — capture matches **one**+    Entry against the Work list. That is linear and comfortably inside its+    100 ms budget (0.0002 s measured); the projection's cost is the *per-Entry*+    repetition, which capture never pays.+  - *Is the cost the derivation or the matching?* The matching. The same 5,000+    Entries with no Works project in 0.083 s; with 1,000 Works and a+    sequence-only rule, 0.506 s; with 1,000 Works and an identity-bearing rule,+    5.82 s.++## Discovered Root Cause++Two defects, the first hiding the second.++**1. The fixture (the reported bug).** `composedBasis` models the *initial*+teaching sweep of an untaught Site, which by construction owns no Works. That is+a legitimate shape, but it was the only shape the preview budget measured, and+it is the cheap one. The re-teaching shape — a taught Site with its 1,000 Works,+every Entry re-matched against all of them — had no fixture and therefore no+budget.++**2. The code the fixture was hiding.** Once measured, the identity-bearing+preview costs 5.82 s against a 1 s budget, because+`IdentityFirstWorkMatchingPlanner` is called once per Entry and each call:++- re-validates the **entire** candidate array (`validate` builds a fresh+  `Set<UUID>` over all candidates and, per candidate, `validateEvidence`+  allocates another `Set<UUID>`), and+- linear-scans and sorts every candidate for the identity arm, and again for the+  title arm.++At 5,000 Entries × 1,000 Works that is ~5M candidate validations, ~5M `Set`+allocations and ~10M exact-scalar comparisons per preview — work whose result+cannot change across the loop, because the candidate list does not change.++`WorkMatcher.match` has the same shape on the title arm: a full scan per Entry+over an unchanging candidate list. That one is cheaper per candidate (a+comparison, no allocation), which is why it costs 0.42 s rather than 5.7 s.++**Defect type:** missing test coverage (1) concealing an algorithmic defect —+loop-invariant work performed once per iteration (2).++**Why it occurred:** the fixture was written for the initial-teaching journey+the feature was being built for, and the budget was attached to it. Identity-first+matching landed on the same planner a milestone later and nothing re-read the+fixture.++## Resolution for the Issue++**Changes made:**++- `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixture.swift` — added+  `titleMatchBasis(entryCount:)` and `identityMatchBasis(entryCount:)` plus+  `identityRequest` and the identity-bearing URL shape they need+  (`.workAndSequence` over a `story` slug and the existing `chapter` sequence).+  Both carry the Site's full 1,000 Works whatever the Entry count, because a+  single-Entry edit acknowledgement also re-matches against every Work. Entries+  carry their Work id at `.pattern` provenance — `.manual` would make `assign`+  return `.protected` and skip the arms outright.+  `composedBasis(entryCount:)` is **unchanged**, so every existing measurement+  over this fixture keeps the shape it was banded against.+- `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScalePerformanceTests.swift`+  — four new arms holding the same Req 8.5 budgets over the Work-bearing bases.+- `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixtureTests.swift` —+  correctness tests (they run in the default `make test-core`) pinning that the+  new bases really do reach the matching arms and reuse every existing Work, and+  that the old basis reaches none of them.+- `Packages/AsterismCore/Sources/AsterismCore/URLIdentityPlanner.swift` — added+  `IdentityFirstWorkCandidateIndex`: candidates validated once and bucketed by+  exact matching title and by retained rule identity. `match` and+  `identityReuse` gained index-taking overloads; the array-taking entry points+  are thin wrappers that build a single-use index, so every existing caller+  keeps its behaviour and its error cases.+- `Packages/AsterismCore/Sources/AsterismCore/TeachingTypes.swift` — added+  `WorkMatchIndex` and `WorkMatcher.match(parsedWorkTitle:index:)`. The+  array-taking `match` deliberately keeps its scan rather than routing through+  the index: a one-shot match pays only a comparison per candidate, and building+  an index for it would cost a hash and an allocation per candidate instead.+- `Packages/AsterismCore/Sources/AsterismCore/ComposedTeachingProjection.swift`+  — `project` builds both indices once, before the per-Entry loop, and `assign`+  takes them.++**Approach rationale:** the buckets are pre-sorted with the same comparator the+scanning arms applied (`candidateOrder` for identity-first, UUID string order for+`WorkMatcher`), so an outcome read out of a bucket is identical to the one the+scan produced — including the member order of an `.ambiguous`.+`ExactScalarString` hashes and compares by Unicode scalar, so a dictionary keyed+on it is exactly the equality the matchers already used; keying on `String` would+have silently switched them to canonical equivalence.++**One deliberate behaviour change.** Candidate validation is now eager: `project`+validates the whole candidate list once at the top rather than lazily inside the+first `match` call that happens to run. A library whose Work list is invalid *and*+whose every Entry is protected therefore now throws where it previously projected.+That is the stricter and more deterministic answer — the refusal no longer depends+on which Entries happen to reach the arm — and every shape it refuses is one the+library validator already rejects.++**Alternatives considered:**++- **Fixture only, no optimisation** (the ticket's stated minimum). Rejected: it+  would leave `make test-performance-m4` red, and that target's contract is that+  it exits 0 with accepted breaches recorded as known issues. A 5.8× breach on a+  path a reader hits per keystroke is not an accepted breach.+- **Change `composedBasis` to carry Works.** Rejected: several suites measure+  over it and their recorded bands are in CLAUDE.md and the verification-run+  files. Adding Works there would have shifted every one of them at once, which+  is exactly what the ticket warned against.+- **Memoise the validation inside the planner** (a static cache keyed on the+  candidate array). Rejected: shared mutable state in a pure planner, and the key+  would have to hash the whole array anyway.++## Regression Test++**Test files:**++- `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixtureTests.swift` —+  `initialTeachingBasisHasNoWorkCandidates`, `titleMatchBasisReusesEveryWork`,+  `identityMatchBasisReusesEveryWork`, `editAckBasesCarryEveryWork`.+- `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScalePerformanceTests.swift`+  — `editAckTitleMatching`, `editAckIdentityMatching`,+  `completePreviewTitleMatching`, `completePreviewIdentityMatching`.++**What they verify:** that the budgeted basis reaches the Work-matching arms at+all (the correctness tests, which fail loudly if a future change empties the+candidate lists again), and that those arms fit inside the Req 8.5 budgets (the+timed arms). `completePreviewIdentityMatching` is the failing test the fix is+anchored to: 5.82 s before, well inside 1 s after.++**Run commands:**++- `make test-core` — the correctness tests.+- `make test-performance-m4` — the timed arms (host only, ~21 minutes, safe).++## Affected Files++| File | Change |+|------|--------|+| `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixture.swift` | Work-bearing re-teaching bases |+| `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScaleFixtureTests.swift` | Correctness pins for the new bases |+| `Packages/AsterismCore/Tests/AsterismCoreTests/M4ScalePerformanceTests.swift` | Four new Req 8.5 budget arms |+| `Packages/AsterismCore/Sources/AsterismCore/URLIdentityPlanner.swift` | `IdentityFirstWorkCandidateIndex`; index-taking `match` / `identityReuse` |+| `Packages/AsterismCore/Sources/AsterismCore/TeachingTypes.swift` | `WorkMatchIndex`; index-taking `WorkMatcher.match` |+| `Packages/AsterismCore/Sources/AsterismCore/ComposedTeachingProjection.swift` | Build both indices once per projection |++## Verification++**Automated:**++- [x] The regression test fails before the fix and passes after it:+      `complete-preview-identity-matching` 5.82 s → 0.19 s against a 1 s budget.+- [x] `make test-core` exits 0 (a later re-run tripped on the two live Apple+      Intelligence calls, which pass on their own and are not in AsterismCore —+      see `measurements.md` §5).+- [x] `make test-performance-m4` reports the same **8 known issues** CLAUDE.md+      documents, over 32 tests rather than 28. One run failed a regression+      ceiling on `capture-projection-duplicateSiteRows`, an arm this change does+      not touch; a control run with the change stashed and a third run on the+      branch both pass, and `measurements.md` §4 sets out the evidence that it+      was host contention.+- [x] No new compiler warnings from the changed files (`swift build` and+      `swift build --build-tests`, unwrapped).+- [x] No linter: the repo has none. A clean `make test-core` is the bar.++**Owed:** one `make test-performance-m4` on a quiet machine, end to end.++The measured numbers and the verbatim run summaries are in `measurements.md`+beside this file.++## Prevention++- A performance fixture is part of the assertion, not scaffolding for it. When a+  budget guards a code path, something has to pin that the fixture *reaches* it —+  which is why the new correctness tests assert candidate counts and reuse+  outcomes rather than only timings.+- When a planner gains a second matching arm, re-read the fixtures that budget+  the planner.+- A matcher that validates its whole input on every call is fine for a capture+  and quadratic for a sweep. Where a caller matches many things against one list,+  the list belongs in an index built outside the loop.++## Not Done++- `TitleProjectionPlanner` (`TitleProjectionPlanner.swift`, the M2 title-only+  projection) calls `WorkMatcher.match` once per Entry over a loop-invariant+  candidate list and would benefit from the same `WorkMatchIndex` hoist. It is+  out of scope here: nothing in this ticket measures it, and there is no failing+  test to anchor the change. Worth its own ticket.++## Related++- Ticket T-1910, raised as a skipped finding in the pre-push review of+  `feature/unified-teaching-composition`.+- `CLAUDE.md` § "Build and test tooling" — the `make test-performance-m4`+  known-issue bands.+- `docs/agent-notes/testing.md` § "`make test-performance-m4` exits 0".

diff --git a/specs/bugfixes/m4-fixture-work-matching/measurements.md b/specs/bugfixes/m4-fixture-work-matching/measurements.mdnew file mode 100644index 0000000..ab52e5e--- /dev/null+++ b/specs/bugfixes/m4-fixture-work-matching/measurements.md@@ -0,0 +1,148 @@+# Measurements: T-1910++All numbers are `swift test -c release -Xswiftc -DASTERISM_PERFORMANCE_TESTING`+on the host (`make test-performance-m4`'s own configuration), 2026-09-05.+`median` is the regression statistic; `p95` is reported but asserted only on a+run declared `CONTROLLED=1` (Decision 10 of `library-integrity-tolerance`).++**Read the host caveat (§4).** Three sibling agent worktrees were running their+own `swift test` loops throughout, and the host load average sat between 13 and+29. The arms this change touches moved far enough that noise cannot explain+them. One run of the full target failed a regression ceiling on an arm this+change does not touch; §4 shows, by a control run with the change stashed and a+third run on the branch, that it was the host.++## 1. The bug, measured++The four preview arms that existed all project a basis with **no Works**. The+four added here project the same 5,000 Entries over the Site's own 1,000 Works.++| arm | before | after | budget |+|---|---|---|---|+| `edit-ack-title-matching` | 0.000290 s | 0.000702 s | 100 ms |+| `edit-ack-identity-matching` | 0.002756 s | 0.004372 s | 100 ms |+| `complete-preview-title-matching` | 0.505781 s | **0.083442 s** | 1 s |+| `complete-preview-identity-matching` | **5.822792 s** | **0.189502 s** | 1 s |++`complete-preview-identity-matching` is the failing test the fix is anchored to.+Before the fix it exceeded its 1-second budget by **5.8×**; it was reproduced+twice at 5.813 s and 5.823 s, so it is not a noise artefact. After the fix it is+**31× faster** and sits at 19% of the budget.++`complete-preview-title-matching` fell 6.1× and now costs the same as the+Work-free arm — the `WorkMatcher` scan is gone from the per-Entry loop entirely.++The two edit-acknowledgement arms got *slower* in absolute terms and this is+expected: with one Entry, building an index over 1,000 candidates costs more+than one scan over them. They are 140× and 23× inside their 100 ms budget, and+the shape they measure — one keystroke — is not the shape the bug was about.+Trading a 0.4 ms regression on a single Entry for 5.6 s on the whole preview is+the right trade, and the budgets say so.++## 2. The arms this change does not touch, before and after++Unchanged within run-to-run noise:++| arm | before | after |+|---|---|---|+| `edit-ack-expanded` | 0.000033 s | 0.000016 s |+| `edit-ack-collapsed` | 0.000006 s | 0.000006 s |+| `complete-preview-expanded` | 0.082584 s | 0.096977 s |+| `complete-preview-collapsed` | 0.029928 s | 0.041257 s |++The two `complete-preview-*` arms rose 15–35% between the two runs. They project+a basis with no Works, so nothing in this change runs over them; the rise is the+host (see §4), and both remain an order of magnitude inside their 1 s budget.++## 3. `make test-performance-m4` — the full target++Run with the fix, `PERFORMANCE_LOG` collected:++```+✘ Test run with 32 tests in 5 suites failed after 1380.227 seconds with 9 issues (including 8 known issues).+```++- **32 tests**, up from the 28 CLAUDE.md records — the four new preview arms.+- **8 known issues**, exactly the steady state CLAUDE.md documents (Req 10.1's+  settling pass, Req 5.4's three capture-projection arms, Req 5.5's three+  diagnosis re-derivations, and the full-tier no-op reconcile). No known issue+  was added or retired.+- **1 real failure**, and it is not on a path this change touches:+  `capture-projection-duplicateSiteRows` measured 0.3079 s against the 0.25 s+  regression ceiling asserted outside its known-issue block. See §4.++## 4. The failing ceiling is the host, not the change++`capture-projection-*` measures `LibraryRepository.projectCapture`, which is a+store read plus `computeCaptureOutcome` → `captureWorkMatch`. None of those is in+this change's diff — the diff touches+`ComposedTeachingProjectionPlanner.project`, `IdentityFirstWorkMatchingPlanner`+and `WorkMatcher`, and capture reaches none of them.++Four runs of `M4ToleratedScalePerformanceTests` were taken to settle it — three+on this branch and one **control run with the change stashed**, on the same+host:++| arm | ceiling | branch 1 | branch 2 | branch 3 | control (no change) |+|---|---|---|---|---|---|+| `capture-projection-duplicateSiteRows` | 0.25 s | **0.3079** ✘ | 0.2185 | 0.1789 | 0.1902 |+| `capture-projection-siteMissing` | 0.25 s | 0.2488 | 0.1881 | 0.1724 | 0.1863 |+| `capture-projection-duplicateIdentity` | 0.25 s | 0.2029 | 0.2149 | 0.2073 | — |+| `diagnosis-refresh-foreground` | 0.4 s | 0.3280 | 0.3451 | 0.3358 | 0.2989 |+| `diagnosis-refresh-after-write` | 0.4 s | 0.3346 | **0.4465** ✘ | 0.3383 | 0.3046 |+| `diagnosis-refresh-duplicateSiteRows` | 0.4 s | 0.3260 | 0.3371 | 0.3229 | 0.3224 |++A **different** arm breached on each of the first two branch runs and neither+breached on the third. That is the signature of a noisy host, not a regression: a+regression moves the same arm every time. Sample spreads across this suite ran+1.2–2.8×, and the host load average was 13–29 because three sibling worktrees+were running their own `swift test` loops.++Both the control run and the branch's third run reported the same verdict:++```+━ Test run with 6 tests in 1 suite passed after 214.460 seconds with 6 known issues.   (control, no change)+━ Test run with 6 tests in 1 suite passed after 217.116 seconds with 6 known issues.   (branch, run 3)+```++Six known issues, no failure, on both. On the branch's third run every+`capture-projection` arm is at or below the control's number.++**Owed:** one `make test-performance-m4` on a quiet machine to confirm the whole+target exits 0 in one pass. Each suite has now been shown to pass individually,+but the 32-test run that failed has not been repeated end to end. CLAUDE.md's own+guidance applies — "Do not treat a single run as a baseline or a single failure+as a regression."++## 5. `make test-core`++```+[exited with code 0]+```++A later re-run exited 2 on the two live Apple Intelligence calls+(`FoundationCharacterExtractionModelClient`, `FoundationRuleSuggestionModelClient`)+with a `GenerationError`. Both passed on an immediate re-run of just those+suites, and neither is in AsterismCore:++```+✔ Test "A live model call returns a decodable ExtractionResult for one note" (33.499 seconds)+Suite "FoundationCharacterExtractionModelClient" passed after 33.518 seconds+✔ Test "A live model call returns a decodable RuleProposal for a two-example corpus" (19.559 seconds)+Suite "FoundationRuleSuggestionModelClient" passed after 19.561 seconds+```++The on-device model was being contended for by the sibling worktrees at the time.++The new correctness tests all pass:++```+✔ Test "The initial-teaching basis reaches no Work-matching arm at all" passed after 0.023 seconds.+✔ Test "The title-matching basis reuses all 1,000 existing Works by exact title" passed after 5.030 seconds.+✔ Test "The identity-matching basis reuses all 1,000 existing Works by URL identity" passed after 15.417 seconds.+✔ Test "A single-Entry Work-bearing preview still matches against all 1,000 Works" passed after 0.025 seconds.+```++(Those are debug timings, and they are their own small confirmation of the bug:+the Work-free basis projects in 0.367 s where the identity-bearing one takes+15.4 s.)