lanework/Kanban/LiveStore/BannerCenter.swift

import Foundation
import Observation

// MARK: - Tone

/// The three kinds a banner row can be (02-architecture.md § The banner surface, "Tones, not
/// components"): one layout, one accessibility path, three colorings. The card window's
/// remote-change signpost (07-sync-collab.md) is this same component in `.info` — visually calm,
/// no error color — which is precisely why the distinction is a *tone* on a shared row rather
/// than three view types that would drift apart.
public enum BannerTone: Sendable, Equatable {
    case error
    case warning
    case info
}

// MARK: - Row payloads

/// A one-shot write failure: it happened once, it is over, and it waits to be read
/// (02-architecture.md § The banner surface, "One-shots dismiss, conditions heal").
///
/// **No timeout, ever** — "an error never evaporates unread". The row leaves the strip only when
/// the user dismisses it, which is why it carries an id of its own rather than being identified by
/// its content: two identical failures a minute apart are two rows, and dismissing one must not
/// take the other with it.
public struct OneShotBanner: Identifiable, Sendable, Equatable {
    public let id: UUID
    public let error: BoardWriteError
    /// When the failure happened — the sort key for "newest first within a class".
    public let occurredAt: Date

    public init(id: UUID = UUID(), error: BoardWriteError, occurredAt: Date = Date()) {
        self.id = id
        self.error = error
        self.occurredAt = occurredAt
    }
}

/// **The one-shot failure class's second shape** (02-architecture.md § The banner surface, settled
/// 2026-07-31): "the `BoardWriteError`-shaped write failure, and a message-carrying **git-operation
/// failure** — the operation named in the user's words plus the underlying error, phrasing still
/// BannerCenter's — because failures rank by what they are, not by which error vocabulary threw
/// them".
///
/// It is a row of its own rather than a `BoardWriteError` with a git-shaped case because "git
/// operations stay off the closed `WriteOperation` vocabulary — only the banner tier learns the
/// second shape": a `WriteOperation` is something the Writer does to a file the model knows, and a
/// checkout is not one. What the two shapes *do* share is everything the strip cares about — the
/// error tone, the failure rank, and the one-shot's dismissable, untimed lifecycle — which is why
/// they sit in one precedence class rather than in two (`BannerCenter.rows(...)`).
///
/// It replaces the shipped compromise, and the compromise is worth naming because its reasoning is
/// the mirror of this one's: a failed restore used to post a `LossBanner` — the nearest class with a
/// free-form message — which read as warning-tone content-didn't-arrive when what actually happened
/// was an action that didn't happen at all. "A failed undo restore, branch switch, or (pro-m2)
/// pull/push … presents in the error tone at the failure rank, never as a warning-tone loss row."
public struct GitFailureBanner: Identifiable, Sendable, Equatable {
    public let id: UUID
    /// Which operation didn't happen — a closed vocabulary, so the sentence is composed here rather
    /// than carried in from the git layer (`GitOperation`).
    public let operation: GitOperation
    /// The underlying error, verbatim — libgit2's own message, the free-form tail the phrasing rules
    /// allow ("could not write to '…': Permission denied"). Never the banner's verb.
    public let reason: String
    /// When it failed — the sort key for "newest first within a class", shared with the write
    /// failures it ranks beside.
    public let occurredAt: Date

    public init(id: UUID = UUID(), operation: GitOperation, reason: String, occurredAt: Date = Date()) {
        self.id = id
        self.operation = operation
        self.reason = reason
        self.occurredAt = occurredAt
    }
}

/// **The git operations the banner can name** — the closed vocabulary behind `GitFailureBanner`,
/// and the git half of the rule the `WriteOperation` enum already carries for writes: "the banner
/// owns all user-facing phrasing and localization from that vocabulary", and an operation added
/// without a sentence to say about it is a compile-time hole rather than a silent default
/// (`BannerCenter.actionPhrase(for:)` switches over it exhaustively).
///
/// Deliberately small. 02 names the members: "a failed undo restore, branch switch, or (pro-m2)
/// pull/push" — the first three are here because they have producers today
/// (`GitHistoryProvider`, `GitBranchSwitcher`), and pull/push join by adding a case and a sentence
/// when the remote half ships.
///
/// No associated values: a branch switch's target is in the in-progress row the failure replaces
/// ("Switching to 'main'…"), and a failure that named it again would be the only member of this
/// vocabulary carrying a payload — a case can grow one the day a sentence genuinely needs it.
public enum GitOperation: Sendable, Equatable, CaseIterable {
    case undo
    case redo
    case branchSwitch

    /// **Add-git, when the form that asked has been dismissed** (06-history-undo.md ▸ Interaction with
    /// external writers, ruled 2026-07-31): "if the sheet has been dismissed before the answer
    /// arrives, the failure falls back to the one-shot banner above — inline is the primary surface,
    /// never a silence trap." Inline is `HistoryStore.lastFailure`; this is the fallback.
    case addGit

    /// The pair ⌘Z and ⇧⌘Z map to — so a caller that knows a `HistoryDirection` (the provider's
    /// crossing) never has to spell the mapping itself, and never gets it backwards.
    public static func restore(_ direction: HistoryDirection) -> GitOperation {
        switch direction {
        case .undo: .undo
        case .redo: .redo
        }
    }
}

/// A loss row: content that didn't arrive though nothing failed (02-architecture.md § The banner
/// surface, "Loss rows are the warning-tone class for non-failure losses", settled 2026-07-28) —
/// folders skipped from a Finder drop, the app's own relocation and repair notices, their future kin.
///
/// **It takes the one-shot's lifecycle**, deliberately: "a loss the user didn't notice is the harm,
/// so it never auto-expires" is `OneShotBanner`'s "an error never evaporates unread", read for a row
/// that reports something incomplete rather than something failed. Carrying its own `id` is the same
/// consequence: two identical losses a minute apart are two rows, and dismissing one must not take
/// the other with it.
///
/// **It ranks below the true failures and above the ambient notices** — "an action that didn't
/// happen outranks one that partially did" — which is why it is its own `BannerRow` case rather than
/// an `OneShotBanner` with a `nil` error or a `InfoSignpost` with a heavier tone: neither of those
/// vocabularies has a slot at loss's precedence, and bending one to fit would blur the reading that
/// the class exists to make precise.
public struct LossBanner: Identifiable, Sendable, Equatable {
    public let id: UUID
    public let message: String
    /// When the loss happened — the sort key for "newest first within a class".
    public let occurredAt: Date

    public init(id: UUID = UUID(), message: String, occurredAt: Date = Date()) {
        self.id = id
        self.message = message
        self.occurredAt = occurredAt
    }
}

/// The standing condition "changes aren't being recorded to history" (02-architecture.md §
/// Write-failure surfacing, "Auto-commit failures beyond `index.lock` contention").
///
/// `reason` is **diagnostic English, not the banner's verb** — the free-form tail the phrasing
/// rules allow ("disk full", "the repository is corrupt"). The user-facing sentence is
/// `BannerCenter.headline(for:)`'s, like every other row's.
///
/// `since` exists because this is a condition, not an event: it started at some point and has been
/// true ever since. Nothing renders it today; it is what a later "suspended for 4 minutes" reading
/// would be built from, and recording it costs a `Date`.
public struct HistorySuspension: Sendable, Equatable {
    public let reason: String
    public let since: Date

    public init(reason: String, since: Date = Date()) {
        self.reason = reason
        self.since = since
    }
}

/// A calm notice that something happened elsewhere — the info tone's *passive* half.
///
/// Its one producer-to-be is the card window's **remote-change signpost** (07-sync-collab.md): a
/// pull landed commits touching the very card being edited, the dirty buffer stays put and wins,
/// and the window says so without a modal or a merge UI. 02-architecture.md calls that "this same
/// component in the info tone", which is exactly what this is.
///
/// **Ranks last and may collapse** (settled, § The banner surface): "calm by design, nothing gated
/// on seeing them instantly" — the opposite end of the strip from an in-progress row, whose spinner
/// may never hide.
///
/// Dismissable, like a one-shot and unlike a condition: it reports something that already happened,
/// so there is nothing for it to heal into.
public struct InfoSignpost: Identifiable, Sendable, Equatable {
    public let id: UUID
    public let message: String
    public let occurredAt: Date

    public init(id: UUID = UUID(), message: String, occurredAt: Date = Date()) {
        self.id = id
        self.message = message
        self.occurredAt = occurredAt
    }
}

/// Work in flight, shown as an info row with a spinner (02-architecture.md § The banner surface,
/// "In-progress operations are info rows"): bracketed git operations ("Pulling…", "Switching to
/// 'main'…") and long non-git work (big-board Duplicate, template instantiation, large attachment
/// imports).
///
/// `label` is the caller's, deliberately: unlike a failure — whose vocabulary is the closed
/// `WriteOperation` enum precisely so the banner can own every word — an in-progress row names an
/// operation the banner layer has no enum for, and inventing one would mean a new git verb could
/// not describe itself without touching this file. The tradeoff is stated rather than hidden.
///
/// **Cancel appears on safe copies only** (settled): copy-shaped work — attachment imports,
/// Duplicate, template instantiation — carries `cancel`, meaning "remove the partial copy, nothing
/// lost". Git brackets pass `nil`: they are seconds long, and aborting a rebase mid-flight is a
/// repair job, not a cancel.
public struct InProgressOperation: Identifiable, Sendable {
    public let id: UUID
    public let label: String
    public var isCancelable: Bool { cancel != nil }
    /// `@MainActor @Sendable` rather than the plainer `@MainActor () -> Void` the sketch carried:
    /// this value is stored in a `Sendable` struct, and a non-`Sendable` function type would make
    /// that conformance a lie under strict concurrency. The isolation is unchanged — cancelling
    /// runs on the main actor, where the operation's own state lives.
    public let cancel: (@MainActor @Sendable () -> Void)?

    public init(id: UUID = UUID(), label: String, cancel: (@MainActor @Sendable () -> Void)? = nil) {
        self.id = id
        self.label = label
        self.cancel = cancel
    }
}

// MARK: - Rows

/// One row in a window's banner strip.
///
/// The eight cases are the whole vocabulary of 02-architecture.md § The banner surface, and they
/// divide into three lifecycles that the view renders differently and that the ordering rule
/// treats as classes:
///
/// - **Conditions heal**: `readOnlyLock`, `reloadBreakage`, `historySuspended`. They describe
///   ongoing state and carry no dismiss control — "an error never evaporates unread" has a twin,
///   "a condition is never dismissed while it is still true". Each leaves when the thing it
///   describes stops being true.
/// - **One-shots dismiss**: `oneShot`, `gitFailure`, `loss`, and `signpost`. Each reports something
///   that already happened, so only the user can clear it. `gitFailure` is the same failure class as
///   `oneShot` in a second shape (settled 2026-07-31) — same tone, same rank, same lifecycle, a
///   message where the other carries a `BoardWriteError`. `loss` shares the lifecycle deliberately
///   (settled 2026-07-28) even though it reports no failure: "a loss the user didn't notice is the
///   harm, so it never auto-expires" is the same reasoning that keeps a one-shot from evaporating
///   unread, aimed at a row that isn't an error at all.
/// - **In-progress rows complete or fail**: `inProgress`. Completion clears the row; failure swaps
///   it for a one-shot (`BannerCenter.endOperation(_:)` + `post(_:)`).
///
/// **`Identifiable` by a synthetic string id**, not by content: the condition rows are singletons
/// per window (there is one lock, one breakage, one history suspension), so their ids are constant
/// and a changing *reason* updates the row rather than replacing it — no view churn, no lost
/// animation, and no diffing surprise when a lock's cause changes underneath a standing row.
public enum BannerRow: Identifiable, Sendable {
    /// The board refuses writes. Condition, error tone. Producers: the failed bracketed reload, the
    /// vanished root, and the writability probe — at open and, symmetrically, on every reconciling
    /// reload thereafter.
    case readOnlyLock(ReadOnlyLockReason)
    /// A reload failed and the last good snapshot is still on screen. Condition, error tone.
    case reloadBreakage(BoardLoadError)
    /// A write that did not happen. Dismissable, error tone.
    case oneShot(OneShotBanner)
    /// A git operation that did not happen — an undo restore, a branch switch, and (pro-m2) a pull
    /// or a push. Dismissable, error tone, and it ranks *with* `oneShot`: the one-shot failure
    /// class's second shape (settled 2026-07-31, see `GitFailureBanner`).
    case gitFailure(GitFailureBanner)
    /// Content that didn't arrive though nothing failed — folders skipped from a Finder drop, the
    /// app's own relocation and repair notices. Dismissable, warning tone: below the true failures above it,
    /// above the ambient notices below it (settled 2026-07-28, see `LossBanner`).
    case loss(LossBanner)
    /// History has stopped advancing. Condition, warning tone — the files are safe, only the undo
    /// trail is degraded, which is a warning rather than an error. (m7's committer drives it.)
    case historySuspended(HistorySuspension)
    /// Work in flight. Info tone, spinner, **pinned above everything** and exempt from the collapse.
    /// (m5's copy-shaped work and m7's git brackets drive it.)
    case inProgress(InProgressOperation)
    /// A calm notice — the passive half of the info tone, ranking last of all. (m6's card window
    /// drives it, as 07-sync-collab.md's remote-change signpost.)
    case signpost(InfoSignpost)

    public var id: String {
        switch self {
        case .readOnlyLock: "read-only-lock"
        case .reloadBreakage: "reload-breakage"
        case let .oneShot(banner): "one-shot:\(banner.id.uuidString)"
        case let .gitFailure(banner): "git-failure:\(banner.id.uuidString)"
        case let .loss(loss): "loss:\(loss.id.uuidString)"
        case .historySuspended: "history-suspension"
        case let .inProgress(operation): "operation:\(operation.id.uuidString)"
        case let .signpost(signpost): "signpost:\(signpost.id.uuidString)"
        }
    }

    public var tone: BannerTone {
        switch self {
        // A git operation that failed is an action that didn't happen, so it takes the failure
        // tone with the write failures it ranks beside — "never as a warning-tone loss row"
        // (02-architecture.md § The banner surface, settled 2026-07-31).
        case .readOnlyLock, .reloadBreakage, .oneShot, .gitFailure: .error
        case .historySuspended, .loss: .warning
        case .inProgress, .signpost: .info
        }
    }

    /// Whether this row is pinned above the strip's collapse — true for in-progress rows and
    /// nothing else (settled, 02-architecture.md § The banner surface: "a spinner may never hide
    /// behind '+N more'"). The view reads this rather than re-deriving the case.
    public var isPinned: Bool {
        if case .inProgress = self { true } else { false }
    }

    /// The user-facing line. Every word of it comes from `BannerCenter`'s phrasing statics except
    /// an in-progress row's, whose label is its caller's (see `InProgressOperation`).
    public var headline: String {
        switch self {
        case let .readOnlyLock(reason): BannerCenter.headline(for: reason)
        case let .reloadBreakage(error): BannerCenter.headline(for: error)
        case let .oneShot(banner): BannerCenter.headline(for: banner.error)
        case let .gitFailure(banner): BannerCenter.headline(for: banner)
        case let .loss(loss): loss.message
        case let .historySuspended(suspension): BannerCenter.headline(for: suspension)
        case let .inProgress(operation): operation.label
        case let .signpost(signpost): signpost.message
        }
    }

    /// Only the rows reporting something that already happened carry a dismiss control —
    /// one-shot failures, loss rows, and signposts. Conditions stand until they heal; in-progress
    /// rows complete, fail, or are cancelled — neither is something a user can wave away.
    public var dismissID: UUID? {
        switch self {
        case let .oneShot(banner): banner.id
        case let .gitFailure(banner): banner.id
        case let .loss(loss): loss.id
        case let .signpost(signpost): signpost.id
        case .readOnlyLock, .reloadBreakage, .historySuspended, .inProgress: nil
        }
    }

    /// **This row's buttons, in the order Tab visits them** — Cancel, then Dismiss.
    ///
    /// It exists because 10-accessibility.md ▸ Full Keyboard Access rules the banner's buttons in by
    /// name (2026-07-29): "'Every control' is literal and includes banner-row buttons — a Dismiss or
    /// Cancel on a banner must be a Tab stop … Cancel on an in-progress operation is exactly the
    /// control that cannot require a pointer". A claim about *which* controls a row has is then a
    /// fact about the row's data rather than about a view's `if` ladder, so it can be pinned
    /// headlessly and the strip can render straight from it (`BannerStripView`) — which is the same
    /// posture the rest of this type already takes ("the per-kind affordances hang off the row's
    /// data, not off separate views").
    ///
    /// No row has both today: the two conditions are disjoint by construction (only an in-progress
    /// row cancels, and an in-progress row is never dismissable). The order is stated anyway, since
    /// it is the Tab order the moment one does.
    public var controls: [BannerRowControl] {
        var controls: [BannerRowControl] = []
        if case let .inProgress(operation) = self, let cancel = operation.cancel {
            controls.append(.cancel(cancel))
        }
        if let dismissID {
            controls.append(.dismiss(dismissID))
        }
        return controls
    }
}

// MARK: - A row's buttons

/// One button on a banner row — the strip's whole vocabulary of per-row controls, as data.
///
/// **Identified by its label**, which is legitimate rather than lazy here: a row carries at most one
/// of each kind, the label is the user-facing name of exactly that kind, and it is what both
/// surfaces the ruling cares about need — the button's title (Cancel) or its accessibility label
/// (Dismiss, whose face is an ✕ glyph).
public enum BannerRowControl: Identifiable, Sendable {

    /// Stop the operation this row is reporting and remove its partial work — carried by cancelable
    /// in-progress rows only (`InProgressOperation`: "safe copies only").
    case cancel(@MainActor @Sendable () -> Void)

    /// Clear this row, by the id `BannerCenter.dismiss(_:)` takes.
    case dismiss(UUID)

    public var label: String {
        switch self {
        case .cancel: "Cancel"
        case .dismiss: "Dismiss"
        }
    }

    public var id: String { label }
}

// MARK: - BannerCenter

/// The model behind one window's banner strip: the rows a window owns outright, plus the ordering
/// and phrasing rules the whole app renders through (02-architecture.md § The banner surface).
///
/// ### What lives here and what does not
///
/// A center holds the state nothing else does: dismissable one-shot failures in both shapes (write
/// and git-operation), loss rows, the history suspension, in-progress operations, and passive
/// signposts. It deliberately does **not** hold the read-only lock or the reload breakage — those
/// are `BoardStore`'s truths, and copying them here would create a second place for them to be
/// stale. `BoardStore.bannerRows` composes both halves through
/// `rows(lock:breakage:oneShots:losses:suspension:operations:signposts:gitFailures:)`, which is a
/// *pure function* precisely so the precedence rule can be tested without a store, a window, or a
/// filesystem.
///
/// ### Phrasing lives here too
///
/// "The banner owns all user-facing phrasing and localization from that vocabulary" (02 §
/// Write-failure surfacing). `BoardWriteError` carries a closed `WriteOperation` enum and a
/// diagnostic `reason`; `headline(for:)` switches over that enum **exhaustively, with no
/// `default`**, so a Writer operation added without a sentence to say about it is a compile-time
/// hole rather than a silent fallback. The same rule covers locks, breakage, and the history
/// suspension: their user-facing lines are here, not on the error types. **And the git operations
/// too** (settled 2026-07-31): `GitOperation` is that vocabulary's git-side twin — "the operation
/// named in the user's words plus the underlying error, phrasing still BannerCenter's" — which is
/// why `GitOperationFailure.operation`, a developer-facing string, never reaches a row.
///
/// ### One center per window, not per board
///
/// The strip is "hosted by the window of origin": a card window's own save and attachment failures
/// belong to that window's center, and re-home to the board window's when it closes (m6's job).
/// `BoardStore` owns *the board window's* center, which is why it is a stored `let` there rather
/// than something injected — a board window's strip has exactly one lifetime, the store's.
@MainActor
@Observable
public final class BannerCenter {

    // MARK: State

    /// Newest first, which is the order the class renders in. Kept sorted on insertion rather than
    /// at render time so that two failures sharing a timestamp — plausible inside one run loop
    /// turn — still order deterministically.
    public private(set) var oneShots: [OneShotBanner] = []

    /// The failure class's git-shaped half, newest first like `oneShots` — a failed undo restore,
    /// a failed branch switch, and (pro-m2) a failed pull or push (settled 2026-07-31, see
    /// `GitFailureBanner`).
    ///
    /// **A second array rather than a second case inside `OneShotBanner`**: the two shapes have no
    /// payload in common — one carries the closed `WriteOperation` vocabulary and a path, the other
    /// a git operation and libgit2's message — and every reader of a write failure would have to
    /// start asking whether it was one. They meet where the design says they meet, in the strip's
    /// order: `rows(...)` merges them into one precedence class, newest first across both.
    public private(set) var gitFailures: [GitFailureBanner] = []

    /// Newest first, like `oneShots` — content that didn't arrive though nothing failed,
    /// dismissable and untimed for the same reason a one-shot failure is (settled 2026-07-28, see
    /// `LossBanner`).
    public private(set) var losses: [LossBanner] = []

    /// The standing "history isn't advancing" condition, or `nil` when commits are landing.
    public private(set) var historySuspension: HistorySuspension?

    /// Work in flight, newest first for the same reason `oneShots` is.
    public private(set) var operations: [InProgressOperation] = []

    /// Passive notices, newest first. Empty until m6's card window starts posting the
    /// remote-change signpost.
    public private(set) var signposts: [InfoSignpost] = []

    public init() {}

    // MARK: One-shots

    /// Records a write failure. **Every failed write lands here before it reaches its caller** —
    /// `BoardStore.performWrite` posts before it rethrows — because the banner is how the one-way
    /// flow stays honest: the action visibly did not happen, and the strip is the only place that
    /// says why (02 § Write-failure surfacing).
    public func post(_ error: BoardWriteError) {
        oneShots.insert(OneShotBanner(error: error), at: 0)
    }

    /// **Records a git operation that didn't happen** — the failure class's second shape (settled
    /// 2026-07-31, see `GitFailureBanner`): an undo or redo restore that failed cleanly
    /// (06-history-undo.md ▸ Interaction with external writers), a branch switch that could not run
    /// (06 ▸ Branch switching), and the remote pair when pro-m2 wires them.
    ///
    /// `reason` is the underlying error and nothing else — libgit2's own message, or the app's own
    /// diagnosis of a repository it could not read. **Every user-facing word is composed here**
    /// (`headline(for:)`), which is what keeps a caller from inventing a verb: the git layer's
    /// `GitOperationFailure.operation` is a developer-facing string and never reaches the strip.
    public func postGitFailure(_ operation: GitOperation, reason: String) {
        gitFailures.insert(GitFailureBanner(operation: operation, reason: reason), at: 0)
    }

    /// Posts a loss row — content that didn't arrive though nothing failed (settled 2026-07-28, see
    /// `LossBanner`). Newest first, like the one-shots it shares a lifecycle with.
    public func postLoss(_ message: String) {
        losses.insert(LossBanner(message: message), at: 0)
    }

    /// Posts a passive notice — m6's remote-change signpost and whatever joins it. Newest first,
    /// like the one-shots it shares a lifecycle with.
    public func postSignpost(_ message: String) {
        signposts.insert(InfoSignpost(message: message), at: 0)
    }

    /// **The skipped undo step** (13-native-undo.md ▸ Rules ▸ staleness validation): an inverse found
    /// its target holding somebody else's newer value, so it was popped rather than applied and ⌘Z
    /// fell through to the next step. This is the row that says so.
    ///
    /// **A signpost, and no new class** — the vocabulary's answer rather than a compromise. 13 asks
    /// for an "info-tone banner", and 02-architecture.md § The banner surface gives the info tone
    /// exactly two halves: the pinned in-progress row with its spinner, and the passive signpost.
    /// Nothing is in flight here, so the passive half is the whole of the choice. It also reads
    /// right: unlike a `loss` row (warning tone, "content that didn't arrive though nothing failed"),
    /// **nothing was lost and nothing failed** — the file holds exactly what its most recent writer
    /// meant it to, the stack moved on to a step that did apply, and the user's ⌘Z did something. A
    /// row that ranks last and may collapse behind "+N more" is the honest weight for that: calm by
    /// design, nothing gated on seeing it instantly. And it is emphatically not a `oneShot`, which
    /// carries a `BoardWriteError` — a *failed* inverse posts one of those instead, and the two rows
    /// must stay distinguishable (`HistoryStepOutcome`).
    public func postSkippedStep(_ direction: HistoryDirection, subject: String) {
        postSignpost(Self.skippedStepMessage(direction, subject: subject))
    }

    /// **The refused paste** (04-interactions.md ▸ Clipboard, re-ruled 2026-07-29): the staged
    /// snapshot was missing or unreadable, so the paste produced **nothing**, and this is the row that
    /// says so — "the paste produces nothing, and a one-shot failure banner names it from the
    /// manifest's metadata".
    ///
    /// **A `oneShot`, not a loss row** — which is the pivot, and it is the vocabulary reading the
    /// event correctly rather than a reclassification for its own sake. The degraded paste *was* a loss
    /// row because the items landed and only their attachments did not: content that didn't arrive
    /// though nothing failed. Under refuse-don't-degrade nothing lands at all, which is exactly
    /// 02-architecture.md's definition of a one-shot — a write that did not happen — so the row
    /// carries a `BoardWriteError` like every other failure, ranks with the true failures, and says
    /// "Couldn't paste" rather than "Pasted … without".
    ///
    /// The retired member is `postDegradedPaste(_:)` and its `AttachmentLoss` payload: with the
    /// degraded materialization gone there is no partial arrival to account for, and the
    /// loss-accounting problem it existed to report — what didn't arrive, and whether the totals were
    /// honest — dissolves rather than being solved. **The loss class itself is untouched**: folder-drop
    /// skips still post one (`postSkippedFolders`), and the relocation, migration, displacement and
    /// remint notices are all still its.
    ///
    /// `title` is the offending entry's, from the manifest's own metadata, `nil` for an untitled item —
    /// "Untitled" is a rendering, never a value (03-board-ui.md § Card face), so `actionPhrase(for:)`
    /// says "the item" instead. The staging path is what the error names as its `path`: it is the file
    /// that was not there, and naming it is what makes a bug report about this actionable.
    public func postRefusedPaste(title: String?, stagedAt path: String) {
        post(BoardWriteError(
            operation: .paste(title: title),
            path: path,
            reason: .clipboardContentGone
        ))
    }

    /// One card whose loose files were relocated into `attachments/` — what
    /// `relocatedLooseFilesMessage(for:)` names.
    ///
    /// `fileNames` are the names the files had **beside `index.md`**, not the Finder-renamed ones
    /// they may have landed under: those are the names the user or their agent wrote, and the one
    /// they would recognize in a sentence (`WriteOperation.importAttachment`'s own rule, read for
    /// the relocation). `title` is the card's as written, `nil` for an untitled one — "Untitled" is
    /// a rendering, never a value (03-board-ui.md § Card face).
    public struct Relocation: Sendable, Equatable {
        public let title: String?
        public let fileNames: [String]

        public init(title: String?, fileNames: [String]) {
            self.title = title
            self.fileNames = fileNames
        }
    }

    /// **The loose-file relocation** (01-storage-format.md § Fractal layout ▸ Rules, settled
    /// 2026-07-28): a file was sitting beside a card's `index.md`, the app moved it into that card's
    /// `attachments/`, and this is the row that says so — "surfacing a graceful warning-tone notice
    /// naming the card and files".
    ///
    /// **A loss row, though nothing was lost.** The class is the vocabulary's warning-tone,
    /// user-dismissed, never-expiring one — `LossBanner`'s "their future kin" — and this is exactly
    /// that shape read once more: the app did something to the user's files that they did not ask
    /// for, so it must be said out loud, it must not evaporate unread, and it must not rank as an
    /// error, because no action failed. `signpost` would be too quiet (it ranks last and may
    /// collapse behind "+N more"); `oneShot` would be a lie (it carries a `BoardWriteError`, and
    /// the write succeeded). The name of the class is about its *lifecycle and tone*, not about
    /// loss being the only thing it can report.
    ///
    /// A relocation that moved nothing posts nothing.
    public func postRelocatedLooseFiles(_ relocations: [Relocation]) {
        guard let message = Self.relocatedLooseFilesMessage(for: relocations) else { return }
        postLoss(message)
    }

    /// **The legacy tombstone migration** (01-storage-format.md § Deletion, resettled 2026-07-28:
    /// "Legacy `deleted:` keys migrate on load-and-write, never destroy … a graceful warning-tone
    /// notice"): a board written by an older version carried `deleted:` keys, the app moved the
    /// cards those keys named into `.trash/` and returned the lanes live, and this is the row that
    /// says so.
    ///
    /// **A loss row for `postRelocatedLooseFiles`' exact reason**, and it is the same shape of event:
    /// the app moved the user's folders on its own initiative, on a board it opened rather than on a
    /// gesture they made. That must be said out loud, must not evaporate unread, and must not rank
    /// as an error, because no action failed.
    ///
    /// **Cards only** (01-storage-format.md § Deletion, lane clause re-ruled 2026-07-29): the lane
    /// half of this notice retired with the lane migration itself — a lane's `deleted:` is inert now,
    /// nothing is written for it, and a row announcing an act the app did not perform would be worse
    /// than silence.
    ///
    /// `cards` holds the migrated cards' titles, in the order they were written, `nil` for an
    /// untitled one — "Untitled" is a rendering, never a value (03-board-ui.md § Card face), so the
    /// phrasing layer decides what to call it. A migration that migrated nothing posts nothing.
    public func postMigratedTombstones(cards: [String?]) {
        guard let message = Self.migratedTombstonesMessage(cards: cards) else { return }
        postLoss(message)
    }

    /// One claimed board-root name whose squatter was moved aside — what
    /// `displacedClaimedNamesMessage(for:)` names.
    ///
    /// Both names are carried because the notice owes **old and new** (01-storage-format.md
    /// § Fractal layout ▸ Rules, ruled 2026-07-29: "with the relocation-style warning-tone notice
    /// naming old and new"): the user needs to know which of their files moved *and* where to find
    /// it, and a sentence naming only one of the two would be half an answer.
    public struct Displacement: Sendable, Equatable {
        /// The claimed name the app took back — `.trash`, `CLAUDE.md`.
        public let name: String
        /// The Finder-ladder name the displaced node now has — `.trash 2`.
        public let movedTo: String

        public init(name: String, movedTo: String) {
            self.name = name
            self.movedTo = movedTo
        }
    }

    /// **The claimed-name displacement** (01-storage-format.md § Fractal layout ▸ Rules, ruled
    /// 2026-07-29): a folder, file or symlink was sitting on a name the app owns, the app moved it
    /// aside — preserved verbatim, never destroyed — and this is the row that says so.
    ///
    /// **A loss row, on `postRelocatedLooseFiles`' exact reasoning**, which is also what the ruling
    /// asks for by name ("the relocation-style warning-tone notice"): the app moved something of the
    /// user's that they did not ask it to move, so it must be said out loud, must not evaporate
    /// unread, and must not rank as an error, because nothing failed.
    ///
    /// A displacement that displaced nothing posts nothing.
    public func postDisplacedClaimedNames(_ displacements: [Displacement]) {
        guard let message = Self.displacedClaimedNamesMessage(for: displacements) else { return }
        postLoss(message)
    }

    /// **The duplicate-id remint** (01-storage-format.md § Fractal layout ▸ Rules, re-ruled
    /// 2026-07-29): two folders were carrying one id, the app gave the later one the fresh identity a
    /// copy should have had, and this is the row that says so — "Announced, not invisible: the
    /// relocation-style warning-tone notice names the repair … identity changed, so a line says so,
    /// but nothing waits on consent".
    ///
    /// **A loss row, on `postRelocatedLooseFiles`' exact reasoning**, and the ruling asks for that
    /// class by name: the app renamed a folder of the user's that they did not ask it to rename, so it
    /// must be said out loud, must not evaporate unread, and must not rank as an error, because
    /// nothing failed. It is emphatically **not** a condition banner with a button — the user-gated
    /// Repair retired on 2026-07-29 and this row is what replaced it.
    ///
    /// `titles` are the reminted items' as the load found them, `nil` for an untitled one — "Untitled"
    /// is a rendering, never a value (03-board-ui.md § Card face). A remint that reminted nothing
    /// posts nothing.
    public func postRemintedDuplicateIDs(_ titles: [String?]) {
        guard let message = Self.remintedDuplicateIDsMessage(for: titles) else { return }
        postLoss(message)
    }

    /// Posts the skipped-folders loss row for a Finder drop that imported its files but refused its
    /// folders (04-interactions.md ▸ Selection, drag & drop, "Folders are refused at hover"): "a
    /// mixed drag proposes for its files only, and the drop imports the files while a one-shot
    /// banner names the skipped folders" — now a loss row (settled 2026-07-28): folders that never
    /// arrived are a non-failure loss, not a write failure. With the degraded paste retired
    /// (`postRefusedPaste`) this is the loss class's clearest remaining instance: the drop *did* land,
    /// and only the payload the attachment model cannot hold stayed behind.
    ///
    /// A drop with no skipped folders posts nothing — nothing was lost, so there is nothing to say.
    public func postSkippedFolders(count: Int) {
        guard count > 0 else { return }
        postLoss(Self.skippedFoldersMessage(count: count))
    }

    /// **The mixed-kind drag that tried to leave the trash** (04-interactions.md ▸ The trash, ruled
    /// 2026-07-31 with kind-blind trash selection): "pickup is allowed — the selection is legal — but
    /// every out-of-trash drop target refuses the mixed payload, and the release surfaces a notice
    /// explaining the rule … the refused drag ends like any refusal, rows staying put".
    ///
    /// **A loss row**, with the relocation family and `postSkippedFolders`: nothing failed — no write
    /// was attempted — and nothing is wrong with the board, but the gesture the user made did not
    /// happen, which is exactly the warning-tone "didn't arrive" register. The wording is 04's own,
    /// verbatim, and lives here because BannerCenter owns the phrasing (▸ Clipboard).
    ///
    /// It is the *teaching* half of the guard: ⌘C/⌘X grey out silently (menu validation says no
    /// before the gesture starts), while the drag has no such surface to say it in advance — so the
    /// drop-time explanation "is where the rule teaches itself".
    public func postMixedTrashDrag() {
        postLoss(Self.mixedTrashDragMessage)
    }

    /// 04-interactions.md ▸ The trash's own sentence.
    nonisolated static let mixedTrashDragMessage =
        "Cards and lanes leave the trash separately \u{2014} restore one kind at a time"

    /// Removes a dismissable row: a one-shot failure in either shape, a loss row, or a signpost.
    /// **An id that names an in-progress operation is ignored** rather than ending it, because
    /// "dismiss" and "cancel" are different promises and a row that offers one must never quietly do
    /// the other.
    public func dismiss(_ id: UUID) {
        oneShots.removeAll { $0.id == id }
        gitFailures.removeAll { $0.id == id }
        losses.removeAll { $0.id == id }
        signposts.removeAll { $0.id == id }
    }

    /// Removes every dismissable row — both failure shapes, losses, and signposts alike. The strip's
    /// own "clear all" affordance later; today it is what a window uses when it re-homes its rows
    /// elsewhere (m6).
    public func dismissAllDismissableRows() {
        oneShots.removeAll()
        gitFailures.removeAll()
        losses.removeAll()
        signposts.removeAll()
    }

    // MARK: History suspension

    /// Raises (or refreshes) the "changes aren't being recorded to history" condition — m7's
    /// committer calls this when a commit fails past `index.lock` contention (06-history-undo.md
    /// covers the lock itself, which is deliberately *not* a banner).
    ///
    /// A second call while already suspended keeps the original `since` and takes the newer
    /// `reason`: the condition never stopped being true, so restarting its clock would misreport
    /// how long history has been stalled, while the newest diagnosis is the useful one.
    public func suspendHistory(reason: String) {
        historySuspension = HistorySuspension(reason: reason, since: historySuspension?.since ?? Date())
    }

    /// Clears it, on the first successful commit. Idempotent — clearing a condition that is not
    /// standing is not an error, it is the ordinary shape of "commit succeeded".
    public func clearHistorySuspension() {
        historySuspension = nil
    }

    // MARK: In-progress operations

    /// Starts an info row with a spinner and hands back its id.
    ///
    /// - Parameter cancel: non-`nil` only for copy-shaped work, where cancelling means "remove the
    ///   partial copy, nothing lost" (02, settled). Git brackets pass `nil`.
    @discardableResult
    public func beginOperation(label: String, cancel: (@MainActor @Sendable () -> Void)? = nil) -> UUID {
        let operation = InProgressOperation(label: label, cancel: cancel)
        operations.insert(operation, at: 0)
        return operation.id
    }

    /// **Relabels a running operation** — the same row, still spinning, now saying something else.
    ///
    /// It exists for one sentence in 06-history-undo.md ▸ Interaction with external writers:
    /// "contention outlasting the brief retry surfaces as a *waiting* state in the operation's
    /// in-progress banner row ('waiting for another writer's git lock'), retrying on its cadence".
    /// The waiting state is explicitly *the operation's own row*, not a second row and not a
    /// replacement — the operation has not restarted, it is explaining itself — so the id is stable
    /// and the view neither churns nor re-animates.
    ///
    /// An unknown id is a no-op: a completion racing a relabel is ordinary, not a bug.
    public func updateOperation(_ id: UUID, label: String) {
        guard let index = operations.firstIndex(where: { $0.id == id }) else { return }
        let existing = operations[index]
        operations[index] = InProgressOperation(id: existing.id, label: label, cancel: existing.cancel)
    }

    /// Ends one — the row leaves the strip.
    ///
    /// **Completion and failure both come through here.** There is no `failOperation`: a failure is
    /// `endOperation(id)` followed by `post(error)`, which is exactly 02's "failure swaps it for
    /// the error row" expressed as two calls that already exist. Ending an unknown id is a no-op;
    /// a close racing a completion is ordinary, not a bug.
    public func endOperation(_ id: UUID) {
        operations.removeAll { $0.id == id }
    }

    // MARK: - Ordering

    /// The strip's rows, precedence-ordered — **the whole of the ordering rule, in one pure
    /// function** (02 § The banner surface, "Concurrent conditions stack").
    ///
    /// The precedence, settled: **in-progress rows (pinned)** > read-only lock > reload breakage >
    /// one-shot write failures > **loss rows** > commit and attachment failures > **passive info
    /// rows**. Four readings of it are worth stating because the code depends on them:
    ///
    /// - **The two info classes sit at opposite ends of the strip.** An in-progress row is pinned
    ///   above everything: it is the only explanation the strip offers for a bracket's write lock
    ///   and for a close or quit deferring teardown, and a copy row's Cancel has to stay reachable
    ///   — "a spinner may never hide behind '+N more'". A passive signpost ranks last and may
    ///   collapse: calm by design, nothing gated on seeing it instantly. Same tone, opposite
    ///   urgency.
    /// - **Loss rows slot between the true failures and the ambient notices** (settled 2026-07-28):
    ///   "an action that didn't happen outranks one that partially did" is why a loss ranks below
    ///   every one-shot write failure, while "content that didn't arrive though nothing failed" is
    ///   still more consequential than a condition or a signpost that only reports ambient state —
    ///   so a loss also ranks above `historySuspended` and the attachment one-shots that follow it.
    ///   Concretely: non-attachment one-shots, then loss rows, then the commit-and-attachment class,
    ///   then signposts.
    /// - **An attachment failure is a one-shot**, not a separate kind — its `WriteOperation` is
    ///   `.importAttachment`. It ranks in the commit-failure class rather than with its fellow
    ///   one-shots, so a fresh failed import sits below an older failed move. That is the design's
    ///   ordering read literally, and it is defensible: a failed import is the least destructive of
    ///   the failures (the drop was accepted, the partial copy was removed), so it yields its place
    ///   to failures that stopped the user's actual work.
    /// - **Within the commit-and-attachment class the standing suspension leads**, then the
    ///   attachment one-shots newest first. A deliberate reading of "newest first within a class":
    ///   a condition and a one-shot are not comparable by recency in any way a user would read as
    ///   order — the condition's `since` is when it *started* being true, not when it happened —
    ///   so they are ordered by kind, and recency orders only the one-shots among themselves.
    /// - **The failure rank holds both shapes, interleaved by recency** (settled 2026-07-31):
    ///   "failures rank by what they are, not by which error vocabulary threw them", so a failed
    ///   undo posted a second ago sits above a failed move from a minute ago and below one from a
    ///   second before it. Ties — two rows sharing a `Date` to the microsecond — put the write
    ///   failure first; the two shapes are posted from different call sites, so a tie is an
    ///   accident of the clock rather than an order anyone can read.
    ///
    /// `signposts` and `gitFailures` carry defaults: the first because its producer is m6's card
    /// window, the second because a center that hosts no git surface (a card window's own) can never
    /// hold one. Every other class is spelled out at every call site — `losses` included, since a
    /// Finder drop that skipped folders already posts one (`postSkippedFolders`).
    public nonisolated static func rows(
        lock: ReadOnlyLockReason?,
        breakage: BoardLoadError?,
        oneShots: [OneShotBanner],
        losses: [LossBanner],
        suspension: HistorySuspension?,
        operations: [InProgressOperation],
        signposts: [InfoSignpost] = [],
        gitFailures: [GitFailureBanner] = []
    ) -> [BannerRow] {
        var rows: [BannerRow] = []

        rows.append(contentsOf: operations.map(BannerRow.inProgress))

        if let lock {
            rows.append(.readOnlyLock(lock))
        }
        if let breakage {
            rows.append(.reloadBreakage(breakage))
        }

        let ordered = newestFirst(oneShots, by: \.occurredAt)
        rows.append(contentsOf: failureRank(
            writes: ordered.filter { !$0.isAttachmentImport },
            git: newestFirst(gitFailures, by: \.occurredAt)
        ))

        rows.append(contentsOf: newestFirst(losses, by: \.occurredAt).map(BannerRow.loss))

        if let suspension {
            rows.append(.historySuspended(suspension))
        }
        rows.append(contentsOf: ordered.lazy.filter(\.isAttachmentImport).map(BannerRow.oneShot))

        rows.append(contentsOf: signposts.map(BannerRow.signpost))
        return rows
    }

    /// **The failure rank, both shapes** (settled 2026-07-31): the non-attachment write failures and
    /// the git-operation failures, merged into the one precedence class they share — "failures rank
    /// by what they are, not by which error vocabulary threw them".
    ///
    /// A merge rather than a re-sort, because both inputs arrive newest-first already; ties keep the
    /// write failure first, which is arbitrary and says so (`rows(...)`).
    private nonisolated static func failureRank(
        writes: [OneShotBanner],
        git: [GitFailureBanner]
    ) -> [BannerRow] {
        var merged: [BannerRow] = []
        var writes = writes[...]
        var git = git[...]
        while let write = writes.first, let failure = git.first {
            if failure.occurredAt > write.occurredAt {
                merged.append(.gitFailure(failure))
                git = git.dropFirst()
            } else {
                merged.append(.oneShot(write))
                writes = writes.dropFirst()
            }
        }
        merged.append(contentsOf: writes.map(BannerRow.oneShot))
        merged.append(contentsOf: git.map(BannerRow.gitFailure))
        return merged
    }

    /// Newest first, and **stable**: `sorted(by:)` is not, and two rows posted in the same run loop
    /// turn can share a `Date` to the microsecond. Ties fall back to the input order, which every
    /// `post…` maintains newest-first on insertion — so a tie renders in the order it was posted
    /// rather than in whatever order the sort happened to leave.
    ///
    /// One function over a date key rather than one per class: the three dismissable classes order
    /// by exactly the same rule, and three copies of it were three places for it to drift.
    private nonisolated static func newestFirst<Row>(_ rows: [Row], by occurredAt: KeyPath<Row, Date>) -> [Row] {
        rows
            .enumerated()
            .sorted { lhs, rhs in
                lhs.element[keyPath: occurredAt] == rhs.element[keyPath: occurredAt]
                    ? lhs.offset < rhs.offset
                    : lhs.element[keyPath: occurredAt] > rhs.element[keyPath: occurredAt]
            }
            .map(\.element)
    }

    // MARK: - Phrasing

    /// The user-facing line for a failed write: what the app could not do, then why.
    ///
    /// "Couldn't move 'Fix login' — disk full" is the design's own example and the shape every
    /// case takes: an action clause the banner owns, an em dash, and the diagnostic cause from the
    /// error's `reason` — the one place free-form English is allowed to survive.
    public nonisolated static func headline(for error: BoardWriteError) -> String {
        let action = actionPhrase(for: error.operation)
        let cause = causePhrase(for: error.reason)
        return cause.isEmpty ? action : "\(action) — \(cause)"
    }

    /// The user-facing line for a git operation that didn't happen: what the app could not do, then
    /// why — the write failure's shape exactly (settled 2026-07-31), because the two are one class.
    ///
    /// The cause is the underlying error verbatim, on `causePhrase(for:)`'s reasoning read one layer
    /// down: libgit2's messages are specific in a way no re-phrasing of ours would be, and the
    /// alternative to showing one is a shrug. An empty or whitespace-only reason leaves the action
    /// clause alone rather than trailing a dash into nothing.
    public nonisolated static func headline(for failure: GitFailureBanner) -> String {
        let action = actionPhrase(for: failure.operation)
        let cause = trimmed(failure.reason)
        return cause.isEmpty ? action : "\(action) — \(cause)"
    }

    /// **Exhaustive by construction — no `default`**, the `WriteOperation` rule applied to the git
    /// vocabulary: an operation added to `GitOperation` without a sentence here fails to compile.
    ///
    /// The undo pair is named by **the command the user pressed** — the skipped-step row's rule
    /// (`skippedStepMessage`), and for its reason: ⌘Z is the gesture, and "Couldn't restore an
    /// earlier state" would describe machinery the user has no model of. The branch switch is named
    /// by the control they used (the popover's branch picker) and stays plural — "branches" — rather
    /// than naming the target, which the in-progress row this failure replaces already said
    /// ("Switching to 'main'…") and which `GitOperation` deliberately doesn't carry.
    private nonisolated static func actionPhrase(for operation: GitOperation) -> String {
        switch operation {
        case .undo: "Undo failed"
        case .redo: "Redo failed"
        case .branchSwitch: "Couldn't switch branches"
        // Named by the control the user pressed, like the pair above — "Add Git" is the button, and
        // the row is only ever seen by someone who pressed it and looked away.
        case .addGit: "Couldn't add git to this board"
        }
    }

    /// **Exhaustive by construction — no `default`.** A `WriteOperation` case added without a
    /// sentence here fails to compile, which is the settled contract ("a new Writer operation
    /// without a banner rendering is a compile-time hole, not a silent default").
    ///
    /// Titles are quoted where the operation carries one and the phrasing stays graceful where it
    /// does not: the enum knows an item's title, never its *kind*, so an untitled failure says
    /// "the item" rather than guessing "card" and being wrong about a lane.
    ///
    /// The trash verbs match the commands the user pressed — **Delete**, Empty Trash — which is
    /// 03-board-ui.md § Trash's naming constraint, settled with the trash UI copy: "Finder's 'Move to
    /// Trash' phrasing is reserved for the system Trash; board deletion says 'Delete'". A banner
    /// saying a card could not be *moved to the trash* would name the wrong one of the app's two
    /// trashes (the card window's attachment Remove is the other).
    private nonisolated static func actionPhrase(for operation: WriteOperation) -> String {
        switch operation {
        case .createBoard:
            "Couldn't create the board"
        case .createLane:
            "Couldn't create a lane"
        case .createCard:
            "Couldn't create a card"
        case let .move(title):
            if let title { "Couldn't move '\(title)'" } else { "Couldn't move the item" }
        case let .reorder(title):
            if let title { "Couldn't reorder '\(title)'" } else { "Couldn't reorder the item" }
        case let .copy(title):
            if let title { "Couldn't copy '\(title)'" } else { "Couldn't copy the item" }
        case let .paste(title):
            // **The command's own verb** (04-interactions.md ▸ Clipboard's own example sentence,
            // "Couldn't paste 'Fix login' — the copied content is gone"). The user pressed ⌘V, and
            // "Couldn't copy" — the operation the Writer would have run — would name a gesture they
            // never made. The item is named from the manifest's metadata, which is what the embedded
            // `index.md` is kept for now that it is never a materialization source.
            if let title { "Couldn't paste '\(title)'" } else { "Couldn't paste the item" }
        case let .delete(title):
            if let title { "Couldn't delete '\(title)'" } else { "Couldn't delete the item" }
        case let .purge(title):
            if let title { "Couldn't permanently delete '\(title)'" } else { "Couldn't permanently delete the item" }
        case let .migrateTombstone(title):
            // One sentence covering both shapes of the migration, because the user's mental model
            // of either is the same non-event: a board written by an older version being brought
            // up to date. It deliberately names neither "delete" nor "trash" — a card's migration
            // moves it into the trash and a lane's brings it *back*, so any verb specific enough
            // to describe one would be actively wrong about the other.
            if let title { "Couldn't update '\(title)' to the current format" } else { "Couldn't update an item to the current format" }
        case let .style(title):
            if let title { "Couldn't restyle '\(title)'" } else { "Couldn't restyle the item" }
        case let .resize(title):
            if let title { "Couldn't resize '\(title)'" } else { "Couldn't resize the item" }
        case let .rename(title):
            // The title here is the item's name *before* the edit — the one the user is still
            // looking at — which is what makes "Couldn't rename 'Fix login'" (02's own example
            // sentence) identify the right row rather than a name that never landed.
            if let title { "Couldn't rename '\(title)'" } else { "Couldn't rename the item" }
        case let .duplicateBoard(title):
            if let title { "Couldn't duplicate '\(title)'" } else { "Couldn't duplicate the board" }
        case let .saveAsTemplate(title):
            // The command's own words (File ▸ Save as Template), because that is the button they
            // pressed and the board is still exactly as it was: nothing was saved *over*, and the
            // failure is about the copy in the templates folder, not about this board's own files.
            if let title { "Couldn't save '\(title)' as a template" } else { "Couldn't save the board as a template" }
        case let .importAttachment(filename):
            "Couldn't import '\(filename)'"
        case .listAttachments:
            "Couldn't read this card's attachments"
        case let .removeAttachment(filename):
            // **Finder's phrasing, deliberately** — and the one place in this app that is allowed
            // it. The naming constraint above reserves "move to the Trash" for the *system* Trash,
            // and this is the operation that uses it: the file really did go (or fail to go) where
            // Finder's own ⌘⌫ sends things, so saying anything else — "remove", "delete" — would
            // describe the board's own trash instead and promise the wrong recovery.
            "Couldn't move '\(filename)' to the Trash"
        case .renumberChildren:
            "Couldn't renumber cards"
        case let .relocateLooseFile(filename):
            // The verb matches the successful notice's ("Moved 'notes.txt' into attachments"), so
            // the failure reads as the same sentence negated rather than as a different event.
            // It stays in the ordinary one-shot precedence class rather than joining the attachment
            // imports at the bottom: the relocation is work the *app* started on its own, and a
            // failure the user did not provoke is exactly the one they have no other way to learn
            // about.
            "Couldn't move '\(filename)' into attachments"
        case .agentGuide:
            // **"the agent guide", not "CLAUDE.md"**: the file is the app's, written for agents, and
            // most users will never have opened it — a filename here would name something they have
            // no relationship with. It says nothing about the board's own files because none were
            // touched, and nothing is lost: the board works exactly as well without the guide, which
            // is why every *refusal* to write it is a log line and only a real I/O failure gets here.
            "Couldn't update the agent guide"
        case .seedGitignore:
            // **The agent guide's sentence, one file over**, and for its reasons: `.gitignore` is a
            // courtesy the app writes once, most users will never open it, and nothing about the
            // board's own files is at stake — a board without one simply excludes nothing, which is
            // how every board behaved until 2026-07-31. It names the file rather than a role because
            // this one *does* have a name users know from git, and "the ignore list" would be the
            // app inventing a word for something already called something.
            "Couldn't write this board's .gitignore"
        case let .displaceClaimedName(name):
            // **The name, quoted, and what the app wanted with it** — the failure's mirror of the
            // success row ("Renamed '.trash' to '.trash 2' — Lanework needs that name"). It names
            // the *consequence* the user can act on rather than the mechanics: while the name is
            // held, the feature that needs it does not work, and the fix is theirs (move or rename
            // the thing sitting there) because the app has just demonstrated it cannot.
            "Couldn't move '\(name)' aside — Lanework needs that name"
        case let .repairDuplicateID(title):
            // **The failure's mirror of the success row** ("Repaired duplicate id — 'Fix login'"), in
            // the same words, so the two read as one sentence and its negation. It names the *defect*
            // rather than the mechanics ("couldn't rename a folder" would describe an act the user has
            // no model of) and stays graceful when the item is untitled, because the enum knows a
            // title and never a kind.
            if let title { "Couldn't repair the duplicate id of '\(title)'" } else { "Couldn't repair a duplicate id" }
        case let .toggleTask(title):
            // The user's word for it, not the file's: they ticked a box. The card is named where
            // the read that preceded the flip learned its title, so a body write that refused says
            // *which* card refused it — a card window is not always the frontmost thing on screen.
            if let title { "Couldn't tick the checkbox in '\(title)'" } else { "Couldn't tick the checkbox" }
        case let .editBody(title):
            // **Save**, because that is the word for what just failed: the Edit→Preview flip is the
            // effective Save button (05-card-window.md ▸ Edit), and the debounced tick is the same
            // act happening on its own. The keystrokes are still in the buffer — the banner says the
            // app could not put them on disk, not that they are gone.
            if let title { "Couldn't save '\(title)'" } else { "Couldn't save the card" }
        case let .rawSource(title):
            // **Apply**, because that is the button they pressed (05-card-window.md ▸ Raw source
            // outlet), and "source changes" because what failed to land is the whole file as they
            // typed it — not a save of the card's body, which is what "Couldn't save" would claim.
            // The buffer is still on screen: the banner says the app could not put those bytes on
            // disk, not that they are gone.
            if let title { "Couldn't apply source changes to '\(title)'" } else { "Couldn't apply the source changes" }

        // **The comment family names the card, because that is the title these operations carry** —
        // a comment has none of its own (01-storage-format.md § Enhanced schema: "No `title`, no
        // `order`"), and the card is the thing on the window in front of them. Each sentence is the
        // gesture's own word, the vocabulary's standing rule: they pressed ⌘↩, or Delete, or nothing
        // at all in the draft's case.
        case let .saveCommentDraft(title):
            // **"draft", never "save the comment"**: nothing has been posted, and a banner claiming a
            // comment could not be saved would name one that does not exist yet. The typed text is
            // still in the composer — this says the app could not put it on disk.
            if let title { "Couldn't save your draft comment on '\(title)'" } else { "Couldn't save your draft comment" }
        case let .postComment(title):
            // The Comment button's own word (05-card-window.md ▸ The comments column, "⌘↩ posts").
            // The draft is untouched on disk, which is what makes this a retry rather than a loss.
            if let title { "Couldn't post your comment on '\(title)'" } else { "Couldn't post your comment" }
        case let .editComment(title):
            if let title { "Couldn't save your edit to a comment on '\(title)'" } else { "Couldn't save your edit to the comment" }
        case let .deleteComment(title):
            if let title { "Couldn't delete a comment on '\(title)'" } else { "Couldn't delete the comment" }
        case .purgeCommentTrash:
            // No title, and no mention of a trash the user has never seen: `comments/.trash/` is
            // "never a UI surface", so the sentence is about the *card's* files being tidied — the
            // agent guide's posture, one level down. Nothing is lost either way; the next window
            // open sweeps again.
            "Couldn't tidy up deleted comments"
        }
    }

    /// The cause tail. `reason`'s own English is diagnostic and already reads as a phrase ("disk
    /// full", "no such file or directory"); the uneditable-frontmatter case is the one that needs
    /// translating, because its `description` is a fragment written for a developer.
    private nonisolated static func causePhrase(for reason: BoardWriteError.Reason) -> String {
        let text = switch reason {
        case let .unreadable(message):
            message
        case let .uneditableFrontmatter(shape):
            "this file's frontmatter can't be edited in place (\(shape.description))"
        case let .io(message):
            message
        case let .staleTarget(message):
            message
        case let .invalidSource(error):
            // The loader's own reason, without its path: the path is always the card's own
            // `index.md`, and the banner has already named the card. In practice the raw-source
            // outlet raises this in its alert and never here — the store validates before it opens a
            // write bracket — so this line exists for a caller that reached the Writer directly.
            error.reason.description
        case .clipboardContentGone:
            // **04's own words** ("the copied content is gone"), and the whole of what can honestly be
            // said: the snapshot the pasteboard promised is not on disk, so there is no cause to
            // diagnose past that. It deliberately says nothing about *why* — a sweep that ran early, a
            // container the system reclaimed, an unmounted volume — because the user's recovery is the
            // same in every case, and 04 names it: ⌘C again.
            "the copied content is gone"
        }
        return trimmed(text)
    }

    /// The read-only lock's line. Each cause says the same two things — what is wrong, and that
    /// what is on screen is still the last good view — because the lock's whole promise is that
    /// nothing was lost: reading, selecting, searching and copying out all stay live (02 § "The
    /// lock's scope").
    ///
    /// **The unwritable location gets two lines, not one shared one** (02 § Write-failure
    /// surfacing, settled): "which specific cause, not a shared line … the fixes being different
    /// acts". Ejecting a DMG or copying the board off it is not the same repair as a `chmod` or a
    /// Get Info panel, and a line that covered both would name neither.
    public nonisolated static func headline(for lock: ReadOnlyLockReason) -> String {
        switch lock {
        case .bracketedReloadFailed:
            "This board couldn't be re-read after the last operation — showing the last good view, read-only"
        case .vanishedRoot:
            "This board's folder is gone — showing the last good view, read-only"
        case .unwritableLocation(.readOnlyVolume):
            "This board's volume is read-only — showing the last good view, read-only"
        case .unwritableLocation(.permissionDenied):
            "You don't have permission to change this folder — showing the last good view, read-only"
        }
    }

    /// Reload breakage: fail-fast's specifics (the offending path and what is wrong with it), plus
    /// the reassurance that the board is still the board.
    ///
    /// The path is root-relative as `BoardLoadError` reports it, and `"."` — the root's own
    /// `index.md` — is spelled as "This board" rather than shown as a lone dot.
    public nonisolated static func headline(for breakage: BoardLoadError) -> String {
        let reason = trimmed(breakage.reason.description)
        let subject = breakage.path == "." || breakage.path.isEmpty
            ? "This board isn't loading"
            : "'\(breakage.path)' isn't loading"
        return "\(subject): \(reason) — showing the last good view"
    }

    /// The skipped-folders line — 04-interactions.md's own example, "Folders can't be attached — 2
    /// skipped", generalized over the count. `postSkippedFolders` never calls this at `count == 0`,
    /// so every real call already has something to report.
    ///
    /// **Singular stays "Folders can't be attached — 1 skipped"** rather than recasting the leading
    /// clause to "A folder can't be attached": the claim is always about the drag as a whole — *its*
    /// folders didn't make it in — and only the trailing count varies, which is one sentence shape
    /// for every count instead of two that would have to be kept in agreement with each other.
    public nonisolated static func skippedFoldersMessage(count: Int) -> String {
        "Folders can't be attached — \(count) skipped"
    }

    /// The loose-file relocation's line — 01-storage-format.md's own example sentence, "Moved
    /// 'notes.txt' into attachments — 'Fix login'", generalized over the two axes it varies on.
    ///
    /// **Plurals fold twice**, which is the design's word for it ("plurals fold"; "multiple files
    /// one card → 'Moved 3 files into attachments — Fix login'"):
    ///
    /// - **One card, one file** names the file *and* the card, which is the sentence the design
    ///   wrote: both facts fit, so both are said.
    /// - **One card, several files** drops the filenames for their count. A banner is one line, and
    ///   a list of names would be the first thing to truncate; the card is still named, which is
    ///   what makes the notice actionable — the user knows exactly which `attachments/` to look in.
    /// - **Several cards** folds again, to two counts: "Moved 5 files into attachments — 3 cards"
    ///   (settled here, the judgment 01 leaves to the implementation): the true total plus the true
    ///   item count is the honest summary where a truncated list of titles would not
    ///   be. This case is the whole-board sweep (a board opened after an agent scattered files
    ///   across it), where naming three cards of eleven would read as a bug.
    ///
    /// The multi-card branch never has to spell a singular: two cards carry at least two files.
    ///
    /// `nil` when nothing moved — a relocation that relocated nothing is not news. Entries with no
    /// filenames are dropped first, so a caller need not filter its own list.
    public nonisolated static func relocatedLooseFilesMessage(for relocations: [Relocation]) -> String? {
        let cards = relocations.filter { !$0.fileNames.isEmpty }
        guard let only = cards.first else { return nil }

        let total = cards.reduce(0) { $0 + $1.fileNames.count }
        guard cards.count == 1 else {
            return "Moved \(total) files into attachments — \(cards.count) cards"
        }

        let subject = only.title.map { "'\($0)'" } ?? "an untitled card"
        guard total == 1, let name = only.fileNames.first else {
            return "Moved \(total) files into attachments — \(subject)"
        }
        return "Moved '\(name)' into attachments — \(subject)"
    }

    /// The legacy tombstone migration's line, written in `relocatedLooseFilesMessage`'s voice
    /// because it is the same kind of notice: the act first, the subject after an em dash, plurals
    /// folded, a sole item named.
    ///
    /// **One clause, since the lane half retired** (01-storage-format.md § Deletion, re-ruled
    /// 2026-07-29 — lanes ignore the key rather than migrating). The shapes:
    ///
    /// - **One card**: "Moved 'Fix login' to the trash — it carried an old deleted marker".
    /// - **Several**: "Moved 3 cards to the trash — they carried old deleted markers".
    ///
    /// **The tail names the cause**, and it is the whole explanation the row owes: the user did not
    /// delete anything just now, and without the clause the sentence would read as an action they
    /// had somehow just taken.
    ///
    /// `nil` when nothing migrated — a migration that migrated nothing is not news.
    public nonisolated static func migratedTombstonesMessage(cards: [String?]) -> String? {
        guard !cards.isEmpty else { return nil }
        let subject = cards.count == 1 ? sole(cards[0]) : "\(cards.count) cards"
        let tail = cards.count == 1
            ? "it carried an old deleted marker"
            : "they carried old deleted markers"
        return "Moved \(subject) to the trash — \(tail)"
    }

    /// The claimed-name displacement's line — the relocation's own voice (the act first, the subject
    /// after an em dash), naming **old and new** as the ruling requires.
    ///
    /// - **One name**: "Renamed '.trash' to '.trash 2' — Lanework needs that name". The tail is the
    ///   whole explanation the row owes: the user did not rename anything, and without it the
    ///   sentence would read as an act they had somehow just taken. It says *needs the name* rather
    ///   than anything about what was there, because what was there is the user's business and
    ///   still exists, under the name the row just gave them.
    /// - **Several**: folded to a count in the relocation's idiom — "Renamed 2 items — Lanework
    ///   needs those names". Two claimed names can be squatted at once (a board somebody unpacked
    ///   over an old one), and a two-clause sentence would be longer than the row.
    ///
    /// `nil` when nothing moved — a displacement that displaced nothing is not news.
    public nonisolated static func displacedClaimedNamesMessage(for displacements: [Displacement]) -> String? {
        guard let only = displacements.first else { return nil }
        guard displacements.count == 1 else {
            return "Renamed \(displacements.count) items — Lanework needs those names"
        }
        return "Renamed '\(only.name)' to '\(only.movedTo)' — Lanework needs that name"
    }

    /// The duplicate-id remint's line — **the design's own sentence**, verbatim
    /// (01-storage-format.md § Fractal layout ▸ Rules, re-ruled 2026-07-29: "posts 'Repaired duplicate
    /// id — 'Fix login''").
    ///
    /// - **One**: "Repaired duplicate id — 'Fix login'". The relocation's idiom exactly — the act
    ///   first, the subject after an em dash — and it names the *id* rather than the folder, because
    ///   the folder name is a UUID no user has a relationship with. There is no explanatory tail: the
    ///   act is its own explanation, and nothing about the user's content changed.
    /// - **Several**: folded to a count, the idiom's plural — "Repaired 3 duplicate ids". Several
    ///   arrive together routinely (a hand-copied lane's cards all collide at once), and a sentence
    ///   naming each would be longer than the row.
    ///
    /// `nil` when nothing was reminted — a heal that healed nothing is not news, which is also the
    /// vanished-duplicate race's whole outward appearance: silence.
    public nonisolated static func remintedDuplicateIDsMessage(for titles: [String?]) -> String? {
        guard let only = titles.first else { return nil }
        guard titles.count == 1 else {
            return "Repaired \(titles.count) duplicate ids"
        }
        return "Repaired duplicate id — \(sole(only))"
    }

    /// A sole migrated item's name: its title in quotes, or the untitled rendering the relocation
    /// line already uses ("an untitled card" / "an untitled lane" are one phrase here, because the
    /// clause it sits in already says which level it is).
    private nonisolated static func sole(_ title: String?) -> String {
        guard let title else { return "an untitled item" }
        return "'\(title)'"
    }

    /// The skipped-step line — 13-native-undo.md ▸ Rules' own example sentence, "Undo skipped — 'Fix
    /// login' changed outside Lanework", with ⇧⌘Z's mirror ("Redo skipped — …").
    ///
    /// **The verb is the command the user pressed**, not the half of the step that declined: a step
    /// already undone sits on the redo stack reversed, so the closure ⇧⌘Z crosses is the one
    /// registered as `redo`, and a sentence naming the half would tell the user they pressed the
    /// other key (`HistoryDirection`).
    ///
    /// **`subject` is quoted whatever it names** — the item's title for a step with one target
    /// ("'Fix login'"), the step's own 06 phrase for a batch or an untitled item ("'Move 3 Cards'").
    /// One sentence shape for both readings, chosen where the step is registered
    /// (`BoardStore.registerStep`) because that is the only place that knows how many items it named.
    ///
    /// **"changed outside Lanework"** is the design's own wording and stays literal: it is the whole
    /// explanation the row owes — the app did not decline out of caution, somebody else wrote to that
    /// item, and the reason the step is gone is that applying it would have thrown their edit away.
    public nonisolated static func skippedStepMessage(_ direction: HistoryDirection, subject: String) -> String {
        let verb = switch direction {
        case .undo: "Undo"
        case .redo: "Redo"
        }
        return "\(verb) skipped — '\(subject)' changed outside Lanework"
    }

    /// The suspended-history line. It names the *consequence* the user cares about — undo and the
    /// flush-before-overwrite guarantee are degraded — rather than the git mechanics, and carries
    /// the diagnosis as its tail.
    public nonisolated static func headline(for suspension: HistorySuspension) -> String {
        let reason = trimmed(suspension.reason)
        let line = "Changes aren't being recorded to history"
        return reason.isEmpty ? line : "\(line) — \(reason)"
    }

    /// Trims whitespace and a trailing period: these fragments are tails inside a longer line, and
    /// a stray full stop mid-sentence is the tell of machine-assembled text.
    private nonisolated static func trimmed(_ text: String) -> String {
        var value = text.trimmingCharacters(in: .whitespacesAndNewlines)
        while value.hasSuffix(".") {
            value.removeLast()
        }
        return value
    }
}

// MARK: - Helpers

private extension OneShotBanner {
    /// Whether this failure is an attachment import — the one `WriteOperation` whose one-shot
    /// ranks in the last precedence class rather than with the other one-shots.
    var isAttachmentImport: Bool {
        if case .importAttachment = error.operation { true } else { false }
    }
}