tix-tracker 0.11.0

//! The batch plan: a read-only preview of what a bulk operation would do
//! (WP02, T007–T010).
//!
//! Everything here is **pure with respect to the store**. It reads issue files and
//! the config; it writes nothing — no file, no relationship, no archive, no History
//! event. `contracts/patch-plan.md` §3 requires a plan to be cancellable with no
//! change of any kind, and the only way that is guaranteed rather than hoped for is
//! if the code that builds the preview is the same code that has no write path.
//!
//! ## What a plan is not
//!
//! It is not a transaction and it is not a partial commit. Contract §4: a
//! confirmation "authorizes independent item processing; it is not an
//! all-or-nothing transaction". So a plan carries a per-item outcome and a plan
//! whose items partly conflict is **normal**, not a failure — three issues that move
//! to `wip` and one that cannot is a preview with one diagnostic, not a rejected
//! operation.
//!
//! ## Validation is borrowed, never restated
//!
//! T007 asks for "the same cores as single-issue operations", and the point is not
//! politeness: a bulk validator that re-implemented the date rule would be a
//! second definition that drifts. So [`crate::model::validate_issue_dates`],
//! `Priority::from_str`, the workflow transition service and the typed relationship
//! service are **called**, and this module adds only the thing a bulk plan has and
//! a single edit does not — which item the problem belongs to.

use std::collections::BTreeMap;

use crate::bulk::{ArchiveRequest, IssuePatch, PatchValue, RelationshipOp, Selection};
use crate::error::Result;
use crate::model::Issue;

/// One field's proposed change on one issue.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum FieldChange {
    /// The field would take a new value.
    Set { from: String, to: String },
    /// The field would be emptied.
    Clear { from: String },
    /// The field already holds the requested value, so nothing is written for it.
    ///
    /// A **distinct variant**, not a missing entry. The data-model's rule is "a patch
    /// that is already the requested final state is a no-op and produces no
    /// duplicate History event", and the way that is enforced is by *knowing* it —
    /// a plan that merely omitted the field could not distinguish "no-op" from
    /// "not considered", and the History intent below depends on the difference.
    NoChange { value: String },
}

impl FieldChange {
    /// The name of the field, for a rendering line.
    pub fn field(&self) -> &'static str {
        match self {
            Self::Set { .. } | Self::Clear { .. } | Self::NoChange { .. } => {
                unreachable!("the field name is carried by the caller's key")
            }
        }
    }

    /// Whether this change writes anything.
    pub fn writes(&self) -> bool {
        matches!(self, Self::Set { .. } | Self::Clear { .. })
    }

    /// The change as `(from, to)`, for a surface that projects a plan to a
    /// structured value.
    ///
    /// A method rather than a match in each surface, for the same reason
    /// `ItemOutcome::status` is one: `surface-parity.md` §2 requires every surface
    /// to expose the same "proposed before/after values", and three matches over
    /// these variants is three places for a fourth surface — or a new variant — to
    /// disagree about what a `Clear`'s `to` is.
    ///
    /// `Clear`'s `to` is the **empty string**, not `null` and not a missing key, so a
    /// client reading `to` unconditionally gets a string for every field. A `NoChange`
    /// reports `from == to`, which is the truth: the value would not move.
    pub fn as_from_to(&self) -> (String, String) {
        match self {
            Self::Set { from, to } => (from.clone(), to.clone()),
            Self::Clear { from } => (from.clone(), String::new()),
            Self::NoChange { value } => (value.clone(), value.clone()),
        }
    }
}

/// Why one item cannot be applied, with the item it belongs to.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct ItemDiagnostic {
    /// The issue number, so a message says which item rather than "one of them".
    pub issue: u32,
    /// The field or operation, as the contract spells it.
    pub field: &'static str,
    pub message: String,
}

impl std::fmt::Display for ItemDiagnostic {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "#{}: {}", self.issue, self.message)
    }
}

/// A History entry the apply **intends** to append. Not an event, and not an
/// `HistoryEvent` — constructing one would be the second History system contract
/// §7 prohibits, and appending one is the thing a plan must not do.
///
/// ## The vocabulary is the existing one, and it is short
///
/// The only kind a plan can currently predict is `transition_accepted`, because
/// the only History event a bulk operation can cause is the one the **status
/// seam** writes. A field change appends nothing, because the single-issue path
/// appends nothing — `tix edit --estimate` does not write a History line, and
/// `HistoryEventKind` has no field-update variant to write one under.
///
/// So "this plan appends no History" is a *prediction about an absence*, and the
/// plan reports it as the empty list rather than inventing a name. The apply
/// reaches the same answer from the other direction: the field write goes through
/// `Store::update` and the transition goes through `Store::set_status`, and only
/// the latter appends.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct HistoryIntent {
    /// A [`crate::model::HistoryEventKind::as_str`] spelling. A `&'static str` and
    /// not the enum, because a plan must not be able to *construct* one — but it
    /// must be checked against the enum's spelling by the apply, and a test pins
    /// that every kind a plan emits is one the model can actually produce.
    pub kind: &'static str,
    pub detail: String,
}

/// Where a file would move to, when the patch changes the assignee.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct PathChange {
    pub from: String,
    pub to: String,
}

/// One item's proposed outcome.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct ItemPlan {
    pub number: u32,
    pub title: String,
    /// The field changes, keyed by the contract's field name.
    pub changes: BTreeMap<&'static str, FieldChange>,
    /// The file's destination, when the assignee change moves it.
    pub path_change: Option<PathChange>,
    /// The History entries this item would gain. Empty when nothing changes, which
    /// is what makes a no-op produce no duplicate event.
    pub history: Vec<HistoryIntent>,
    /// The problem, if this item cannot be applied.
    pub diagnostic: Option<ItemDiagnostic>,
    /// The WIP decision for this item's status move, when it makes one
    /// (mission `configurable-wip-limits-01M3C6WX`, WP07 T035).
    ///
    /// Evaluated **in plan order over a projection** — the state each item would
    /// see once the items before it landed — so the fourth item of four into a
    /// column capped at three is the one reported as blocked. A plan that
    /// evaluated every item against the same starting state would promise
    /// something the apply cannot deliver.
    pub wip: Option<crate::wip::ItemWip>,
}

impl ItemPlan {
    /// Whether this item would write anything at all.
    ///
    /// A `## History` entry counts, because it **is** a write: a plan decides
    /// what the apply puts in the file, and an appended line is bytes in that
    /// file. The first version asked only about field and path changes, which is
    /// right for #173 — there a history intent exists only alongside a field
    /// change, so the two questions always had the same answer — and wrong for a
    /// triage decision, which appends an event and may change no field at all: a
    /// duplicate adds an edge and touches no frontmatter, and a decline back
    /// into the entry column clears nothing. Such an item reported `no-op` while
    /// carrying the event the apply would append.
    pub fn is_actionable(&self) -> bool {
        self.diagnostic.is_none()
            && (self.changes.values().any(FieldChange::writes)
                || self.path_change.is_some()
                || !self.history.is_empty())
    }

    /// Whether this item is already in the requested state.
    pub fn is_noop(&self) -> bool {
        !self.is_actionable() && self.diagnostic.is_none()
    }

    /// The shared status word for a **planned** item (`NFR-003`).
    ///
    /// The same four words as `ItemOutcome::status`, and derived from the same two
    /// predicates, so a preview and the apply that follows it cannot describe the same
    /// item two different ways. The mapping is the obvious one and that is the point:
    /// a plan that would write is `applied` in the report, a plan that would write
    /// nothing and has no diagnostic is `no-op`, and a plan with a diagnostic is
    /// `rejected` before anything is attempted.
    pub fn status(&self) -> &'static str {
        if self.diagnostic.is_some() {
            "rejected"
        } else if self.is_actionable() {
            "applied"
        } else {
            "no-op"
        }
    }
}

/// A read-only preview. Cancelling one changes nothing, by construction.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct BatchPlan {
    /// The exact selection this plan was built from, so an apply can check that
    /// the thing it was authorized for is the thing in front of it.
    pub selection: Selection,
    pub patch: IssuePatch,
    pub items: Vec<ItemPlan>,
    /// The relationship change this plan proposes, if the patch named a pair.
    pub pair: Option<PairPlan>,
    /// The archive request this plan carries, **unapplied**.
    pub archive: Option<ArchiveRequest>,
    /// The confirmation, absent until one is issued. `None` is the ordinary state;
    /// see [`Confirmation`].
    pub confirmation: Option<Confirmation>,
}

impl BatchPlan {
    /// The items that would be written.
    pub fn actionable(&self) -> impl Iterator<Item = &ItemPlan> {
        self.items.iter().filter(|i| i.is_actionable())
    }

    /// The items that cannot be applied.
    pub fn blocked(&self) -> impl Iterator<Item = &ItemPlan> {
        self.items.iter().filter(|i| i.diagnostic.is_some())
    }

    /// Whether applying this plan would change anything.
    pub fn is_empty(&self) -> bool {
        self.pair.is_none() && self.archive.is_none() && self.actionable().next().is_none()
    }

    /// The one-line summary the CLI prints above the table.
    pub fn summary(&self) -> String {
        let actionable = self.actionable().count();
        let blocked = self.blocked().count();
        let noop = self.items.iter().filter(|i| i.is_noop()).count();
        // An archive or a pair is an operation on the **selection**, not a per-item
        // field change, so the items are not subjects of it. The first version printed
        // their counts anyway, and an archive preview read "0 of 2 would change; 2
        // already in that state; 1 issue would be archived" — three claims, one of
        // which ("2 already in that state") describes items the plan never proposed
        // to change.
        //
        // So when the patch names no field, the item counts are omitted entirely and
        // the operation is the whole summary. The counts are not zeroed: a reader
        // would have to guess whether there were two items that were all no-ops, or
        // whether the plan had no items at all.
        if !self.patch.has_any_field() {
            let mut parts = Vec::new();
            if let Some(pair) = &self.pair {
                parts.push(format!("1 relationship would {}", pair.verb()));
            }
            if let Some(archive) = &self.archive {
                let count = self.selection.len();
                parts.push(format!(
                    "{count} issue{} would {}",
                    if count == 1 { "" } else { "s" },
                    if archive.archive {
                        "be archived"
                    } else {
                        "be restored"
                    }
                ));
            }
            if parts.is_empty() {
                parts.push("nothing to do".to_string());
            }
            return parts.join("; ");
        }
        let mut parts = vec![format!("{actionable} of {} would change", self.items.len())];
        if noop > 0 {
            parts.push(format!("{noop} already in that state"));
        }
        if blocked > 0 {
            parts.push(format!("{blocked} refused"));
        }
        if let Some(pair) = &self.pair {
            parts.push(format!("1 relationship would {}", pair.verb()));
        }
        if let Some(archive) = &self.archive {
            let count = self.selection.len();
            let mut line = format!(
                "{count} issue{} would {}",
                if count == 1 { "" } else { "s" },
                if archive.archive {
                    "be archived"
                } else {
                    "be restored"
                }
            );
            // A restore names the column it restores into, because "would be
            // restored" is not actionable and #190 has no "previous status" to
            // restore to (C-003). The **absence** is shown too, so a preview says
            // what is missing rather than what would happen.
            match (&archive.target_status, archive.problem()) {
                (Some(target), _) => line.push_str(&format!(" into '{target}'")),
                (None, Some(problem)) => line.push_str(&format!(" — {problem}")),
                (None, None) => {}
            }
            parts.push(line);
        }
        parts.join("; ")
    }

    /// Whether a confirmation on this plan is still valid.
    ///
    /// A confirmation authorizes a **specific** plan. If either the selection or the
    /// patch has changed, it is a different plan and the old confirmation does not
    /// carry over (contract §4: "any change to selection or patch requires a new
    /// preview").
    pub fn confirmation_is_valid(&self) -> bool {
        match &self.confirmation {
            None => false,
            Some(confirmation) => confirmation.fingerprint == self.selection.fingerprint(),
        }
    }
}

/// A confirmation, present only after one is issued and only for the plan it was
/// issued against.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct Confirmation {
    /// The selection membership it was issued against. Membership and not a hash of
    /// the contents, for the reason in [`Selection::fingerprint`].
    pub fingerprint: Vec<u32>,
}

impl Confirmation {
    /// Issue a confirmation for this plan.
    pub fn issue(plan: &BatchPlan) -> Self {
        Self {
            fingerprint: plan.selection.fingerprint(),
        }
    }
}

/// A proposed relationship change, with its targets resolved.
///
/// The `edge` and `identity` are the **canonical** form: `1 blocks 2` and
/// `2 blocked-by 1` are one edge, and a plan that showed the two spellings as two
/// plans would let a user confirm a pair the store then refuses as a duplicate.
/// The `diagnostic` is the structured one, because T019 asks for a *stable* per-pair
/// diagnostic and a message that varies with the wording of a selector is not
/// stable — the message beside it is for a human, the diagnostic is for a test.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct PairPlan {
    pub link: bool,
    /// The resolved numbers, or the diagnostic that stopped resolution.
    pub from: std::result::Result<u32, String>,
    pub to: std::result::Result<u32, String>,
    /// The canonical edge, when both endpoints resolved.
    pub edge: Option<crate::relationships::CanonicalEdge>,
    /// Deliberately **no** relationship id.
    ///
    /// `Relationship::new` mints a fresh ULID, so an id built at resolve time is a
    /// *proposal*, not an identity: the two spellings of one edge would carry
    /// different ids, and a caller that tried `tix link edit`/`remove` with a
    /// preview's id would be addressing a row that does not exist. The store mints
    /// the real one at write time and `PairOutcome` reports it.
    ///
    /// A first draft carried `identity: Option<String>` here and on
    /// `PairResolution`, and a test asserting the two spellings share an id failed
    /// — which is the type telling the truth and the test not listening.
    /// The structured refusal, when there is one.
    pub diagnostic: Option<crate::bulk::pair::PairDiagnostic>,
}

impl PairPlan {
    /// A resolved pair, from the module that knows how to resolve one.
    ///
    /// The only way to build a `PairPlan`, on purpose. Hand-building one means
    /// choosing what "resolved" means at the call site, and two call sites choosing
    /// differently is a pair that plans as valid and refuses on apply. WP02's own
    /// tests hand-built them, which is how three closures in two files had to be
    /// edited when WP04 added the canonical edge and the structured diagnostic.
    pub fn resolved(link: bool, edge: crate::relationships::CanonicalEdge) -> Self {
        Self {
            link,
            from: Ok(edge.from),
            to: Ok(edge.to),
            edge: Some(edge),
            diagnostic: None,
        }
    }

    /// A refused pair, with both endpoints reported.
    ///
    /// `from` and `to` keep the shape of a resolved plan — one `Result` per
    /// endpoint — so a renderer does not need a second code path for the refused
    /// case. A self-link resolves both endpoints and fails on the *pair*, so its
    /// `from`/`to` are both `Ok` and only `diagnostic` is set; that is what
    /// distinguishes it from a missing endpoint without a new variant.
    pub fn refused(
        link: bool,
        from: std::result::Result<u32, String>,
        to: std::result::Result<u32, String>,
        diagnostic: crate::bulk::pair::PairDiagnostic,
    ) -> Self {
        Self {
            link,
            from,
            to,
            edge: None,
            diagnostic: Some(diagnostic),
        }
    }

    /// Whether both endpoints resolved and the pair is not self-referential.
    ///
    /// Delegated to the structured diagnostic when there is one, because
    /// `is_resolvable` is a question about the *pair* and the diagnostic is where
    /// the answer is recorded — two places to answer it is one place too many.
    pub fn is_resolvable(&self) -> bool {
        // A self-link is the one refusal where **both** endpoints resolved, so
        // `from`/`to` cannot decide it — hence the diagnostic first. `is_resolvable`
        // is false for it, which is the answer: a plan reporting "resolvable" for
        // `1 -> 1` would promise an apply that cannot happen.
        if self.diagnostic.is_some() {
            return false;
        }
        self.from.is_ok() && self.to.is_ok()
    }

    pub fn verb(&self) -> &'static str {
        if self.link { "link" } else { "unlink" }
    }
}

/// Build a read-only plan. Writes nothing (contract §3).
///
/// `resolve_pair_targets` is a parameter rather than a call so a caller that has no
/// store to resolve against can pass one, and so the **pure** half of this function
/// is testable without a store at all — which is what T011's 10,000-item case does.
pub fn build_plan(
    config: &crate::config::Config,
    issues_dir: &std::path::Path,
    selection: Selection,
    patch: IssuePatch,
    mut resolve_pair_targets: impl FnMut(&RelationshipOp) -> PairPlan,
) -> Result<BatchPlan> {
    patch.require_wellformed()?;

    let items: Vec<ItemPlan> = selection
        .issues()
        .iter()
        .map(|issue| {
            let diagnostic = first_diagnostic(issue, &patch, config);
            let (changes, path_change, history) = propose(issue, &patch, issues_dir);
            ItemPlan {
                number: issue.number,
                title: issue.title.clone(),
                changes,
                path_change,
                history,
                diagnostic,
                // Filled by `crate::wip::bulk::annotate_plan`, which the caller runs
                // with the **store's** issues: a column's usage counts the cards the
                // batch does not touch, and this function only ever sees the
                // selection (WP07 T035).
                wip: None,
            }
        })
        .collect();

    let pair = patch.relationship.as_ref().map(&mut resolve_pair_targets);
    let archive = patch.archive.clone();

    Ok(BatchPlan {
        selection,
        patch,
        items,
        pair,
        archive,
        confirmation: None,
    })
}

// ---------------------------------------------------------------------------
// T007 — shared field validation
// ---------------------------------------------------------------------------

/// The first validation problem for one issue, or `None`.
///
/// **Calls the shared cores** and adds nothing. The contract's rule is that
/// "bulk-specific validation may add item context but must not redefine the
/// underlying rule", so this function's whole job is to call
/// `model::validate_issue_dates`, `Priority::from_str` and the workflow service
/// with the values the patch would produce, and to stop at the first failure.
///
/// Only the **first** problem per item: a plan is a preview, and a preview that
/// lists eleven failures for one issue buries the one the user has to fix. The
/// remaining problems are found when the fix is applied.
fn first_diagnostic(
    issue: &Issue,
    patch: &IssuePatch,
    config: &crate::config::Config,
) -> Option<ItemDiagnostic> {
    let numbers = |field: &'static str, message: String| {
        Some(ItemDiagnostic {
            issue: issue.number,
            field,
            message,
        })
    };

    // Priority — `commands::parse_priority`, the core a single `tix edit --priority`
    // uses. **Not** `Priority::from_str`, which the first version reached for and
    // which is `Infallible`: it maps anything unrecognised to `Other(s)`, so a
    // check written against it can never fire. The strict three-word rule lives in
    // `parse_priority` and nowhere else, and a bulk validator that re-derived it
    // would be the second definition the contract prohibits.
    if let PatchValue::Set(value) = &patch.priority {
        if let Err(e) = crate::commands::parse_priority(&value.to_string()) {
            return numbers("priority", e.to_string());
        }
    }

    // Dates — `validate_issue_dates` is the core, given the values the patch would
    // *produce* rather than the ones the file has now, because a patch that clears
    // `start_date` makes an existing `due_date` legal and one that sets
    // `start_date` can make it illegal.
    let start = produced(&patch.start_date, issue.start_date.as_deref());
    let due = produced(&patch.due_date, issue.due_date.as_deref());
    if let Err(e) = crate::model::validate_issue_dates(start.as_deref(), due.as_deref()) {
        return numbers("dates", e.to_string());
    }

    // Status — the workflow service, judged as a **transition**, because a status
    // patch is a move. `None` is the right override here: a bulk plan is a preview,
    // and a terminal exit with no reason *is* a refusal the user needs to see.
    if let PatchValue::Set(target) = &patch.status {
        let (_, resolution) = crate::workflow::active_resolution(config);
        match crate::workflow::validate_transition(issue, target, &resolution, config, None) {
            crate::workflow::TransitionOutcome::Rejected(diagnostic) => {
                return numbers("status", diagnostic.to_string());
            }
            crate::workflow::TransitionOutcome::Accepted { .. } => {}
        }
    }

    // Assignee — the same normalisation a create uses, so a bulk assignee lands in
    // the same segment a single create would have chosen.
    if let PatchValue::Set(value) = &patch.assignee
        && crate::assignee::parse_assignee(value).is_none()
    {
        return numbers(
            "assignee",
            format!("assignee {value:?} is not a usable handle or identity"),
        );
    }

    None
}

/// The value a patch would leave in a field, given what is there now.
///
/// `Unchanged` keeps the current value; `Clear` produces no value; `Set` produces
/// the new one. Getting this wrong in either direction produces a diagnostic about
/// a field the patch was not touching.
fn produced(patch: &PatchValue<String>, current: Option<&str>) -> Option<String> {
    match patch {
        PatchValue::Unchanged => current.map(str::to_string),
        PatchValue::Set(v) => Some(v.clone()),
        PatchValue::Clear => None,
    }
}

// ---------------------------------------------------------------------------
// T009 — proposed changes, paths, History intents
// ---------------------------------------------------------------------------

/// The field changes, path move and History intents for one issue.
fn propose(
    issue: &Issue,
    patch: &IssuePatch,
    issues_dir: &std::path::Path,
) -> (
    BTreeMap<&'static str, FieldChange>,
    Option<PathChange>,
    Vec<HistoryIntent>,
) {
    let mut changes: BTreeMap<&'static str, FieldChange> = BTreeMap::new();
    let mut path_change = None;
    let mut history = Vec::new();

    // Status. A `NoChange` when the issue is already there, which is what stops a
    // re-apply from appending a second decision.
    if let PatchValue::Set(target) = &patch.status {
        let from = issue.status.display().to_string();
        if from == *target {
            changes.insert("status", FieldChange::NoChange { value: from });
        } else {
            changes.insert(
                "status",
                FieldChange::Set {
                    from: from.clone(),
                    to: target.clone(),
                },
            );
            history.push(HistoryIntent {
                kind: "transition_accepted",
                detail: format!("{from} -> {target}"),
            });
        }
    }

    // Priority is `PatchValue<Priority>` and estimate is `PatchValue<String>`, so
    // they cannot share one array — the first version tried, and the borrow
    // checker refused. Two short loops over the same shape is the smaller cost
    // than a generic helper that would have to erase the field name too.
    // `Priority` is an enum, not a string, and "unset" is `Priority::Other("")` —
    // so the current value is read through `is_undefined` rather than by testing
    // for an empty rendering, which would make `Other("bogus")` render as
    // "bogus" and compare unequal to the `bogus` a patch sets.
    match &patch.priority {
        PatchValue::Unchanged => {}
        value => record_clearing(
            &mut changes,
            "priority",
            if issue.priority.is_undefined() {
                String::new()
            } else {
                issue.priority.to_string()
            },
            render_typed(value),
            matches!(value, PatchValue::Clear),
        ),
    }
    match &patch.estimate {
        PatchValue::Unchanged => {}
        value => record_clearing(
            &mut changes,
            "estimate",
            issue.estimate.clone().unwrap_or_default(),
            render_typed(value),
            matches!(value, PatchValue::Clear),
        ),
    }

    for (name, value, current) in [
        ("assignee", &patch.assignee, issue.assignee.clone()),
        ("start_date", &patch.start_date, issue.start_date.clone()),
        ("due_date", &patch.due_date, issue.due_date.clone()),
    ] {
        match value {
            PatchValue::Unchanged => {}
            PatchValue::Set(new) => record(
                &mut changes,
                name,
                current.unwrap_or_default(),
                new.clone(),
                &mut history,
            ),
            PatchValue::Clear => record_clearing(
                &mut changes,
                name,
                current.unwrap_or_default(),
                String::new(),
                true,
            ),
        }
    }

    // The list-valued fields, rendered the way the frontmatter writes them, so a
    // preview line and a diff agree.
    for (name, value, current) in [
        ("labels", &patch.labels, Some(&issue.labels[..])),
        ("cycles", &patch.cycles, issue.cycles.as_deref()),
        ("milestones", &patch.milestones, issue.milestones.as_deref()),
    ] {
        match value {
            PatchValue::Unchanged => {}
            PatchValue::Set(new) => record(
                &mut changes,
                name,
                render_list(current.unwrap_or_default()),
                render_list(new),
                &mut history,
            ),
            PatchValue::Clear => record_clearing(
                &mut changes,
                name,
                render_list(current.unwrap_or_default()),
                String::new(),
                true,
            ),
        }
    }

    // The assignee move. The **destination segment**, which is where a create would
    // have put the file — not the current directory, which is a path and not a
    // prediction.
    if let PatchValue::Set(new) = &patch.assignee {
        if let Some(segment) = crate::assignee::parse_assignee(new) {
            // The **destination segment**, via the same `derive_segment` a create
            // uses, so a preview line and the file that eventually lands agree.
            let from = issues_dir.join(crate::assignee::derive_segment(issue.assignee.as_deref()));
            let to = issues_dir.join(crate::assignee::derive_segment(Some(&segment)));
            if from != to {
                path_change = Some(PathChange {
                    from: from.display().to_string(),
                    to: to.display().to_string(),
                });
            }
        }
    }

    (changes, path_change, history)
}

/// Record one field change, and the History intent that goes with it — which for a
/// field change is **none**.
///
/// The `history` parameter is `&mut Vec<HistoryIntent>` rather than absent, and it
/// is left untouched here, because the first draft of this function pushed a
/// `field_updated` intent for every changed field. That was **wrong**, and WP03
/// caught it while building the apply:
///
/// - `tix edit <n> --estimate 3d` writes the field and appends **no** History
///   event. Measured, not assumed.
/// - `HistoryEventKind` has no field-update variant. The vocabulary is
///   relationship, parent, blocked-move and transition.
/// - So a bulk field patch appending `field_updated` would be a **bulk-only**
///   History event: a second History vocabulary, which `contracts/patch-plan.md`
///   §7 prohibits and which the contract's §"History and archive" reads the other
///   way — "the appropriate event **from the prerequisite History service**", and
///   for a field change the prerequisite service's appropriate event is none.
///
/// The parameter stays so the shape of the function is the same as the status
/// branch's, and so a future field that *does* have a service has somewhere to
/// push. The doc comment is the record: `record` is where the temptation lives.
fn record(
    changes: &mut BTreeMap<&'static str, FieldChange>,
    name: &'static str,
    from: String,
    to: String,
    _history: &mut Vec<HistoryIntent>,
) {
    record_clearing(changes, name, from, to, false)
}

/// [`record`], told whether the change is an *emptying* rather than a new value.
///
/// A `Clear` used to reach the plan as `Set { to: "" }` — the two are
/// indistinguishable once rendered, and a plan that cannot tell them apart cannot
/// report "this field would be emptied" to a surface that shows before/after values.
/// `FieldChange::Clear` existed for exactly that and was **never constructed**, so the
/// variant was dead and the doc comment on `as_from_to` promised a behaviour nothing
/// could reach. The WP08 revert-and-red run is what found it: a mutation of the
/// `Clear` arm left every test green, because no test could see a `Clear`.
///
/// The distinction matters to a surface and not to the file, which is why it lives
/// here rather than at the apply: the write is identical (`field:` with no value) and
/// only the *report* differs.
fn record_clearing(
    changes: &mut BTreeMap<&'static str, FieldChange>,
    name: &'static str,
    from: String,
    to: String,
    cleared: bool,
) {
    if cleared {
        changes.insert(name, FieldChange::Clear { from });
        return;
    }
    if from == to {
        changes.insert(name, FieldChange::NoChange { value: to });
        return;
    }
    changes.insert(name, FieldChange::Set { from, to });
}

fn render_typed<T: std::fmt::Display>(value: &PatchValue<T>) -> String {
    match value {
        PatchValue::Set(v) => v.to_string(),
        _ => String::new(),
    }
}

fn render_list(values: &[String]) -> String {
    values.join(", ")
}