trueppm-api 0.4.0__py3-none-any.whl

trueppm_api/apps/projects/seed/replay.py

@@ -0,0 +1,1695 @@
+"""Replay a v2 seed event timeline with backdated history (ADR-0114).
+
+The v1 importer materializes final state: one ORM write per row, every
+``django-simple-history`` row dated ``now()`` by the importer. A v2 seed instead
+authors the program's *life* — an ordered ``events`` timeline — and this module
+replays it so the demo reads like a program that has run for months:
+
+- Tasks are created at their base column and walked forward through real status
+  transitions; each transition writes a history row dated to the event, attributed
+  to a named persona, so a COMPLETE task's History tab shows dated moves by people.
+- A deterministic synthesizer fills the unauthored "boring middle": any task whose
+  final state implies it passed through earlier columns gets synthetic transitions
+  (it always traverses IN_PROGRESS, never NOT_STARTED→COMPLETE, so burndowns are
+  not hollow — VoC/Alex). Authored events always win over synthesis.
+- A day-by-day sim clock writes one ``SprintBurnSnapshot`` per active-sprint day,
+  so burndown curves and multi-sprint velocity are real history, not a single
+  fabricated number.
+
+Replay runs inside the importer's existing ``transaction.atomic()`` and under the
+``seed_replay`` context flag, so the live side effects a real edit would trigger —
+today-dated burndown, board broadcasts, notifications, webhooks — are suppressed.
+A single CPM recalc per project is enqueued by the importer after commit.
+"""
+
+from __future__ import annotations
+
+import functools
+import logging
+import random
+from collections.abc import Callable, Sequence
+from dataclasses import dataclass, field
+from datetime import date, datetime, timedelta
+from typing import TYPE_CHECKING, Any
+from zoneinfo import ZoneInfo
+
+from django.db.models import F, Q
+from django.utils import timezone
+
+from trueppm_api.apps.projects.actual_date_rules import check_actual_dates
+from trueppm_api.apps.projects.models import (
+    AcceptanceCriterion,
+    Baseline,
+    BaselineTask,
+    CommentAcknowledgement,
+    CommentReaction,
+    EstimateStatus,
+    RetroActionItem,
+    Risk,
+    RiskComment,
+    ScopeChangeStatus,
+    Sprint,
+    SprintRetro,
+    SprintScopeChange,
+    SprintState,
+    Task,
+    TaskComment,
+    TaskNote,
+    TaskStatus,
+)
+from trueppm_api.apps.projects.seed.reldates import WorkingCalendar, resolve_timestamp
+from trueppm_api.apps.projects.seed.replay_ctx import seed_replay
+from trueppm_api.apps.projects.services import upsert_burndown_for_sprint
+from trueppm_api.apps.timetracking.models import TimeEntry
+
+if TYPE_CHECKING:
+    from trueppm_api.apps.notifications.services import ParsedMention
+
+logger = logging.getLogger(__name__)
+
+# Seed-script beat action key for board-column moves.
+_TASK_STATUS_ACTION = "task.status"
+
+# Forward progression of the board. The synthesizer walks a task from its base
+# column up to its authored final column through this sequence, so every
+# in-flight/done task passes through IN_PROGRESS (a hollow burndown is the bug
+# Alex called out in VoC). BACKLOG/ON_HOLD are off the progression spine.
+_PROGRESSION = [
+    TaskStatus.NOT_STARTED,
+    TaskStatus.IN_PROGRESS,
+    TaskStatus.REVIEW,
+    TaskStatus.COMPLETE,
+]
+_PROGRESSION_INDEX: dict[str, int] = {str(s): i for i, s in enumerate(_PROGRESSION)}
+
+# Upper bound on the day-by-day snapshot window (~10 years). The schema already
+# caps a single date offset at 4 digits; this bounds the *span* between the
+# earliest and latest beat so no crafted timeline drives an oversized loop.
+MAX_REPLAY_SPAN_DAYS = 3660
+
+
+@dataclass
+class ReplayContext:
+    """Symbol tables + resolved settings the importer hands to the replay pass."""
+
+    anchor: date
+    program_code: str
+    default_actor: Any  # the importing owner; fallback when an event names no actor
+    users: dict[str, Any]  # account slug -> User (or None on the generic path)
+    tasks: dict[tuple[str, str], Task]  # (project_slug, wbs_path) -> Task
+    sprints: dict[tuple[str, str], Sprint]  # (project_slug, sprint_slug) -> Sprint
+    projects: dict[str, Any]  # project_slug -> Project
+    project_calendars: dict[str, WorkingCalendar]  # project_slug -> calendar facts
+    risks: dict[str, Risk]  # risk slug -> Risk
+    final_status: dict[tuple[str, str], str]  # desired end column per task
+    final_sprint: dict[tuple[str, str], dict[str, Any]]  # desired end state per sprint
+    # Authored in-flight progress per task, restored after the timeline runs
+    # (#3486). Only carries the keys the document actually wrote, so a dated
+    # ``task.points`` beat is never overwritten by a value nobody authored.
+    final_progress: dict[tuple[str, str], dict[str, Any]] = field(default_factory=dict)
+    tz: Any = None  # program timezone for synthesized timestamps (defaults to UTC)
+    # Scope-change rows opened during replay, so a later resolve can close them.
+    open_scope: dict[Any, SprintScopeChange] = field(default_factory=dict)
+    # v2.1 (#3603): acceptance criteria per task in position order, so a
+    # ``task.ac_met`` beat can tick one by index; and comments by the slug their
+    # ``task.comment`` beat declared, so a reply/reaction/ack can find them.
+    criteria: dict[tuple[str, str], list[AcceptanceCriterion]] = field(default_factory=dict)
+    comments: dict[str, TaskComment] = field(default_factory=dict)
+
+
+@dataclass(order=True)
+class _Beat:
+    """One resolved timeline beat, sortable by (when, order)."""
+
+    when: datetime
+    order: int
+    action: str = field(compare=False)
+    target: str = field(compare=False)
+    actor: Any = field(compare=False, default=None)
+    data: dict[str, Any] = field(compare=False, default_factory=dict)
+    synthetic: bool = field(compare=False, default=False)
+
+
+def replay_timeline(payload: dict[str, Any], ctx: ReplayContext) -> None:
+    """Replay the document's events (+ synthesized fill) against ``ctx``.
+
+    Idempotent within an import: the caller wipes-and-recreates the program, so
+    the timeline is always replayed from a clean base. Determinism comes from a
+    seed derived from ``program_code`` + the task wbs path.
+
+    Also seeds every persona's notification preferences, then — beside the
+    per-beat notifications individual handlers below emit — synthesizes the two
+    event types with no single triggering beat: a ``task.stale`` sweep and the
+    sponsor persona's ``program.health_digest`` (#3489).
+    """
+    with seed_replay():
+        _seed_notification_preferences(payload, ctx)
+        beats = _resolve_authored(payload, ctx)
+        beats.extend(_synthesize(ctx, beats))
+        beats.sort()
+        _run_sim_clock(beats, ctx)
+        _synthesize_stale_notifications(ctx)
+        _synthesize_program_digest(payload, ctx)
+        _apply_notification_realism(ctx)
+
+
+# --- resolution --------------------------------------------------------------
+
+
+def _resolve_authored(payload: dict[str, Any], ctx: ReplayContext) -> list[_Beat]:
+    """Turn authored event dicts into sortable beats with concrete datetimes."""
+    out: list[_Beat] = []
+    for i, ev in enumerate(payload.get("events", [])):
+        when = resolve_timestamp(ev["at"], anchor=ctx.anchor)
+        actor = ctx.users.get(ev["actor"]) if ev.get("actor") else None
+        out.append(
+            _Beat(
+                when=when,
+                order=i,
+                action=ev["action"],
+                target=ev.get("target", ""),
+                actor=actor or ctx.default_actor,
+                data=ev,
+            )
+        )
+    return out
+
+
+def _resolve_task(ctx: ReplayContext, target: str) -> Task | None:
+    """Resolve a ``task:<project>:<wbs>`` target to a Task in the symbol table."""
+    _, _, ref = target.partition(":")
+    project_slug, _, wbs = ref.partition(":")
+    return ctx.tasks.get((project_slug, wbs))
+
+
+def _resolve_sprint(ctx: ReplayContext, target: str) -> Sprint | None:
+    _, _, ref = target.partition(":")
+    project_slug, _, sprint_slug = ref.partition(":")
+    return ctx.sprints.get((project_slug, sprint_slug))
+
+
+def _task_window(ctx: ReplayContext, key: tuple[str, str], task: Task) -> tuple[date, date]:
+    """The [start, end] working window a synthesized task progresses across.
+
+    Agile tasks ride their sprint window; scheduled tasks ride
+    planned_start..+duration; everything else rides the project start..anchor.
+    The end is clamped to the anchor — a demo never has work completing in the
+    future relative to "today".
+    """
+    project_slug = key[0]
+    project = ctx.projects[project_slug]
+    sprint = task.sprint
+    if sprint is not None:
+        start, end = sprint.start_date, sprint.finish_date
+    elif task.planned_start is not None:
+        start = task.planned_start
+        end = start + timedelta(days=max(task.duration, 1))
+    else:
+        start = project.start_date
+        end = ctx.anchor
+    # Clamp both ends to the anchor: a completed task whose nominal window sits
+    # in the future (relative to import day) still gets past-dated history, never
+    # future-dated — the demo's "today" is the anchor.
+    start = min(start, ctx.anchor)
+    end = min(end, ctx.anchor)
+    if end < start:
+        end = start
+    return start, end
+
+
+def _synthesize(ctx: ReplayContext, authored: list[_Beat]) -> list[_Beat]:
+    """Fill the unauthored timeline: sprint ceremonies + task progressions."""
+    return _synthesize_sprints(ctx, authored) + _synthesize_tasks(ctx, authored)
+
+
+def _synthesize_sprints(ctx: ReplayContext, authored: list[_Beat]) -> list[_Beat]:
+    """Synthesize activate/close beats for sprints the timeline left implicit.
+
+    A sample that merely declares a sprint's end state + points (the v1 shape)
+    still gets a real run: activate at start, close at finish, so the sim clock
+    writes a burndown curve and a velocity number that is actual history.
+    """
+    authored_targets = {b.target for b in authored if b.action.startswith("sprint.")}
+    out: list[_Beat] = []
+    for key, sprint in ctx.sprints.items():
+        target = f"sprint:{key[0]}:{key[1]}"
+        if target in authored_targets:
+            continue
+        final_state = ctx.final_sprint.get(key, {}).get("state", SprintState.PLANNED)
+        if final_state not in (SprintState.ACTIVE, SprintState.COMPLETED):
+            continue
+        out.append(
+            _Beat(
+                when=_aware_dt(ctx, sprint.start_date, 9, 0),
+                order=0,
+                action="sprint.activate",
+                target=target,
+                actor=ctx.default_actor,
+                data={},
+                synthetic=True,
+            )
+        )
+        if final_state == SprintState.COMPLETED:
+            out.append(
+                _Beat(
+                    when=_aware_dt(ctx, sprint.finish_date, 17, 0),
+                    order=2_000_000,  # close sorts after that day's task moves
+                    action="sprint.close",
+                    target=target,
+                    actor=ctx.default_actor,
+                    data={"goal_outcome": ctx.final_sprint.get(key, {}).get("goal_outcome")},
+                    synthetic=True,
+                )
+            )
+    return out
+
+
+def _synthesize_tasks(ctx: ReplayContext, authored: list[_Beat]) -> list[_Beat]:
+    """Generate status transitions for tasks the authored timeline left implicit.
+
+    For each task whose final column is past NOT_STARTED and which has no authored
+    ``task.status`` beat, walk it forward through every intermediate column on
+    evenly-spaced, seeded dates within its window. Deterministic: a fixed seed per
+    (program, wbs) means re-import and round-trip reproduce the same history.
+    """
+    authored_status_targets = {b.target for b in authored if b.action == _TASK_STATUS_ACTION}
+    out: list[_Beat] = []
+    for key, task in ctx.tasks.items():
+        project_slug, wbs = key
+        target = f"task:{project_slug}:{wbs}"
+        if target in authored_status_targets:
+            continue  # author owns this task's progression
+        final = ctx.final_status.get(key, TaskStatus.NOT_STARTED)
+        final_idx = _PROGRESSION_INDEX.get(final)
+        if final_idx is None or final_idx == 0:
+            continue  # BACKLOG / ON_HOLD / still NOT_STARTED — nothing to walk
+        start, end = _task_window(ctx, key, task)
+        steps = _PROGRESSION[1 : final_idx + 1]  # columns to move into, in order
+        rng = random.Random(f"{ctx.program_code}:{wbs}")
+        actor = task.assignee or ctx.default_actor
+        span = (end - start).days or 1
+        # Place each transition at a stable fraction of the window, jittered.
+        for n, status in enumerate(steps, start=1):
+            frac = n / (len(steps) + 1)
+            jitter = rng.uniform(-0.5, 0.5) / (len(steps) + 1)
+            day_offset = round(min(max(frac + jitter, 0.02), 0.98) * span)
+            when = _aware_dt(
+                ctx, start + timedelta(days=day_offset), rng.randint(9, 16), rng.randint(0, 59)
+            )
+            out.append(
+                _Beat(
+                    when=when,
+                    order=1_000_000 + n,  # synthetic beats sort after same-instant authored
+                    action=_TASK_STATUS_ACTION,
+                    target=target,
+                    actor=actor,
+                    data={"to": status},
+                    synthetic=True,
+                )
+            )
+    return out
+
+
+# --- sim clock ---------------------------------------------------------------
+
+
+def _run_sim_clock(beats: list[_Beat], ctx: ReplayContext) -> None:
+    """Apply beats in time order; snapshot active-sprint burndown each day.
+
+    Iterating every calendar day (not only event days) means an active sprint
+    gets a snapshot on quiet days too, so the burndown curve has no gaps.
+    """
+    if not beats:
+        # No timeline (e.g. a v2 doc with only relative dates) — still finalize
+        # sprint states and in-flight progress so the seed's intended end state
+        # holds.
+        _finalize(ctx, ctx.anchor)
+        return
+
+    by_day: dict[date, list[_Beat]] = {}
+    for b in beats:
+        by_day.setdefault(b.when.date(), []).append(b)
+
+    first = min(by_day)
+    last = max(by_day)
+    # Bound the per-day snapshot loop. The schema caps any single offset at ~27
+    # years, but a seed could still author a multi-decade span; clamp the
+    # snapshotted window so a crafted timeline can't drive an oversized loop.
+    # Beats before the window are still applied (so task states entering the
+    # window are correct) — only their day-by-day snapshots are skipped.
+    window_start = first
+    if (last - first).days > MAX_REPLAY_SPAN_DAYS:
+        window_start = last - timedelta(days=MAX_REPLAY_SPAN_DAYS)
+        logger.warning(
+            "seed replay: span %sd exceeds %sd cap; snapshotting recent window only",
+            (last - first).days,
+            MAX_REPLAY_SPAN_DAYS,
+        )
+        for d in sorted(by_day):
+            if d >= window_start:
+                break
+            for beat in by_day[d]:
+                _apply(beat, ctx)
+
+    day = window_start
+    while day <= last:
+        for beat in by_day.get(day, []):
+            _apply(beat, ctx)
+        _snapshot_active_sprints(ctx, day)
+        day += timedelta(days=1)
+
+    _finalize(ctx, last)
+
+
+def _snapshot_active_sprints(ctx: ReplayContext, day: date) -> None:
+    for sprint in ctx.sprints.values():
+        if sprint.state == SprintState.ACTIVE and sprint.start_date <= day <= sprint.finish_date:
+            try:
+                upsert_burndown_for_sprint(sprint, snapshot_date=day)
+            # TODO(#4421): swallows SoftTimeLimitExceeded; the drain-loop guard
+            # is ineffective until this is fixed
+            except Exception:  # burndown is a secondary observation; never fail the import
+                logger.exception(
+                    "seed replay: burndown snapshot failed sprint=%s day=%s", sprint.pk, day
+                )
+
+
+def _finalize(ctx: ReplayContext, day: date) -> None:
+    """Restore the end states the replay had to walk up to, once it has.
+
+    Two passes, both backstops for the same structural fact: replay births an
+    entity at the *base* of its progression so the timeline has somewhere to walk
+    from, and the walk does not always land on every authored field.
+    """
+    _finalize_sprints(ctx, day)
+    _finalize_tasks(ctx, day)
+
+
+def _finalize_sprints(ctx: ReplayContext, day: date) -> None:
+    """Force every sprint to the seed's authored end state after replay.
+
+    Authored/synthesized activate+close beats normally drive this; this is the
+    backstop for sprints the timeline never touched, so the seed's declared
+    state still holds.
+    """
+    for key, sprint in ctx.sprints.items():
+        desired = ctx.final_sprint.get(key, {}).get("state")
+        if desired and sprint.state != desired:
+            sprint.state = desired
+            _save(sprint, _aware(ctx, day), ctx.default_actor, ["state"])
+
+
+def _finalize_tasks(ctx: ReplayContext, day: date) -> None:
+    """Restore the authored progress of every task that ends in flight or in review.
+
+    A task whose end column is IN_PROGRESS or REVIEW is born at 0% with its full
+    points remaining so the timeline has room to walk it forward, and nothing
+    ever puts the authored numbers back on its own: ``_apply_task_status``
+    writes only the column and the actual dates. For IN_PROGRESS neither field
+    is rescued elsewhere (#3486). For REVIEW, ``Task._coerce_signoff_percent``
+    fixes ``percent_complete`` to 100 by design ("work done, awaiting
+    sign-off") but has no opinion on ``remaining_points`` — a story in review
+    with points still open is a real, authorable state (#3518). Without this
+    pass every in-flight or in-review task in every v2 seed loads holding its
+    full ``story_points``.
+
+    That is not cosmetic. ``percent_complete`` sets the *remaining* duration CPM
+    schedules (ADR-0132/0136), so a 0% in-flight task is billed at its full
+    duration and both the utilization heat map and the burndown's "remaining"
+    read high. The pass therefore runs before the importer's post-commit
+    ``enqueue_recalculate``, so the recompute sees the corrected values.
+
+    Only fields the document actually authored are written, which is what keeps
+    a dated ``task.points`` beat authoritative: a task that declares no
+    ``remaining_points`` of its own is left holding whatever the timeline gave
+    it. The importer never hands back ``percent_complete`` for a REVIEW task
+    (see its ``final_progress`` construction), so the coerced 100 is never
+    fought here even though this loop treats both statuses uniformly.
+    """
+    for key, task in ctx.tasks.items():
+        if ctx.final_status.get(key) not in (TaskStatus.IN_PROGRESS, TaskStatus.REVIEW):
+            continue
+        authored = ctx.final_progress.get(key)
+        if not authored:
+            continue
+        fields = [
+            name
+            for name in ("percent_complete", "remaining_points")
+            if name in authored and getattr(task, name) != authored[name]
+        ]
+        for name in fields:
+            setattr(task, name, authored[name])
+        if fields:
+            _save(task, _aware(ctx, day), ctx.default_actor, fields)
+
+
+# --- event handlers ----------------------------------------------------------
+
+
+def _apply(beat: _Beat, ctx: ReplayContext) -> None:
+    handler = _HANDLERS.get(beat.action)
+    if handler is None:  # pragma: no cover - validator restricts the action set
+        logger.warning("seed replay: no handler for action %r", beat.action)
+        return
+    handler(beat, ctx)
+
+
+def _apply_task_status(beat: _Beat, ctx: ReplayContext) -> None:
+    task = _resolve_task(ctx, beat.target)
+    if task is None:
+        return
+    new_status = beat.data.get("to")
+    if not new_status or new_status == task.status:
+        return
+    task.status = new_status
+    fields = ["status"]
+    # Actual dates make baseline-vs-actual slip visible (VoC/Sarah): first move
+    # into IN_PROGRESS stamps actual_start; reaching COMPLETE stamps finish.
+    if new_status == TaskStatus.IN_PROGRESS and task.actual_start is None:
+        task.actual_start = beat.when.date()
+        fields.append("actual_start")
+    if new_status == TaskStatus.COMPLETE:
+        task.actual_finish = beat.when.date()
+        task.remaining_points = 0
+        fields += ["actual_finish", "remaining_points"]
+    # Defense-in-depth (ADR-1153, #3709): beats are deterministic,
+    # developer-authored data and should never trip these rules, but this is a
+    # direct model write like the MS Project importer's, so it runs through the
+    # same shared check rather than trusting that by construction. A violation
+    # here means the beat script itself is wrong — log it and drop the
+    # offending field rather than persisting a pair the engine would reject at
+    # the project's next recompute.
+    violation = check_actual_dates(
+        actual_start=task.actual_start,
+        actual_finish=task.actual_finish,
+        status=task.status,
+        project=task.project,
+    )
+    if violation is not None:
+        logger.warning(
+            "seed replay: beat %r on task %s would violate ADR-1153 (%s); dropping %s",
+            beat.action,
+            task.pk,
+            violation.message,
+            violation.field,
+        )
+        setattr(task, violation.field, None)
+        if violation.field not in fields:
+            fields.append(violation.field)
+    _save(task, beat.when, beat.actor, fields)
+    # Task.save force-stamps status_changed_at=now(); correct it to the beat time
+    # so cycle-time and "in column since" read as history, not import time.
+    Task.objects.filter(pk=task.pk).update(status_changed_at=beat.when)
+    task.status_changed_at = beat.when
+
+
+def _apply_task_assign(beat: _Beat, ctx: ReplayContext) -> None:
+    task = _resolve_task(ctx, beat.target)
+    if task is None:
+        return
+    previous_assignee_id = task.assignee_id
+    assignee = ctx.users.get(beat.data["assignee"]) if beat.data.get("assignee") else None
+    task.assignee = assignee
+    _save(task, beat.when, beat.actor, ["assignee"])
+    _notify_task_assigned(task, previous_assignee_id, assignee, beat)
+
+
+def _notify_task_assigned(
+    task: Task, previous_assignee_id: Any, assignee: Any, beat: _Beat
+) -> None:
+    """``task.assigned`` to the new assignee (#638/#639), mirroring
+    ``views._emit_assignee_change_events``'s notification half.
+
+    Fires on any change that lands a real assignee — a fresh None→user
+    assignment or a user→user reassignment alike — except a clear (assignee
+    set to None) and never to the actor who made the assignment.
+    """
+    if assignee is None or assignee.pk == previous_assignee_id:
+        return
+    if assignee.pk == getattr(beat.actor, "pk", None):
+        return
+    from trueppm_api.apps.notifications.models import NotificationEventType
+    from trueppm_api.apps.notifications.services import create_event_notifications
+
+    subject = f"You were assigned to {task.name}"
+    body = f'You were assigned to the task "{task.name}" in TruePPM.'
+    _dispatch_and_backdate(
+        beat.when,
+        event_type=NotificationEventType.TASK_ASSIGNED.value,
+        project_id=task.project_id,
+        recipient_ids=[assignee.pk],
+        dispatch=lambda: create_event_notifications(
+            event_type=NotificationEventType.TASK_ASSIGNED.value,
+            recipient_ids=[assignee.pk],
+            subject=subject,
+            body=body,
+            project_id=task.project_id,
+            task_id=task.pk,
+        ),
+    )
+
+
+def _apply_task_estimate(beat: _Beat, ctx: ReplayContext) -> None:
+    task = _resolve_task(ctx, beat.target)
+    est = beat.data.get("estimate")
+    if task is None or not est:
+        return
+    task.optimistic_duration = est["optimistic"]
+    task.most_likely_duration = est["most_likely"]
+    task.pessimistic_duration = est["pessimistic"]
+    task.estimate_status = EstimateStatus.ACCEPTED
+    _save(
+        task,
+        beat.when,
+        beat.actor,
+        ["optimistic_duration", "most_likely_duration", "pessimistic_duration", "estimate_status"],
+    )
+
+
+def _apply_task_points(beat: _Beat, ctx: ReplayContext) -> None:
+    task = _resolve_task(ctx, beat.target)
+    if task is None:
+        return
+    fields = []
+    if "points" in beat.data:
+        task.story_points = beat.data["points"]
+        fields.append("story_points")
+    if "remaining_points" in beat.data:
+        task.remaining_points = beat.data["remaining_points"]
+        fields.append("remaining_points")
+    if fields:
+        _save(task, beat.when, beat.actor, fields)
+
+
+def _apply_task_ac_met(beat: _Beat, ctx: ReplayContext) -> None:
+    """Tick one acceptance criterion (v2.1 ``criterion``), else mark the story ready.
+
+    With ``criterion`` the beat records the review trail on that row —
+    ``met_by``/``met_at`` dated to the beat — and leaves Definition of Ready alone:
+    ticking a criterion during sprint review is not a readiness decision. Without
+    it the beat keeps its v2.0 meaning.
+    """
+    task = _resolve_task(ctx, beat.target)
+    if task is None:
+        return
+    index = beat.data.get("criterion")
+    if index is not None:
+        _, _, ref = beat.target.partition(":")
+        project_slug, _, wbs = ref.partition(":")
+        rows = ctx.criteria.get((project_slug, wbs), [])
+        if index < len(rows):
+            row = rows[index]
+            row.met = True
+            row.met_by = beat.actor
+            row.met_at = beat.when
+            row.save(update_fields=["met", "met_by", "met_at"])
+        return
+    task.dor = "ready"
+    _save(task, beat.when, beat.actor, ["dor"])
+
+
+def _apply_task_block(beat: _Beat, ctx: ReplayContext) -> None:
+    """Raise the explicit blocker flag, dated to the beat (#3094).
+
+    ``blocked_reason`` is the flag-of-record. ``Task.save()`` stamps
+    ``blocked_since`` with ``timezone.now()`` on the empty -> non-empty
+    transition, which on a backdated timeline would render every blocker as
+    "0d blocked" — and age is the entire triage signal. So the stamp is
+    overwritten with the beat time through a queryset update, the same way
+    ``_apply_scope_inject`` corrects its ``auto_now_add`` column.
+    """
+    task = _resolve_task(ctx, beat.target)
+    reason = beat.data.get("body")
+    if task is None or not reason:
+        return
+    task.blocked_reason = reason
+    task.blocker_type = beat.data.get("blocker_type", "")
+    task.blocked_by = beat.actor
+    fields = ["blocked_reason", "blocker_type", "blocked_by"]
+    blocking_ref = beat.data.get("blocking_task")
+    if blocking_ref:
+        blocking = _resolve_task(ctx, f"task:{blocking_ref}")
+        if blocking is not None and blocking.pk != task.pk:
+            task.blocking_task = blocking
+            fields.append("blocking_task")
+    _save(task, beat.when, beat.actor, fields)
+    Task.objects.filter(pk=task.pk).update(blocked_since=beat.when)
+    task.blocked_since = beat.when
+    _notify_task_blocked(task, beat)
+
+
+def _notify_task_blocked(task: Task, beat: _Beat) -> None:
+    """``task.blocked`` to the assignee + project leads (#855/#476/ADR-0124),
+    reusing the real recipient resolver and renderer so the content matches the
+    live path exactly (``views._emit_blocked_notification``).
+
+    The rendered body's age line reads "just now" for a freshly-flagged task —
+    it always does here too, since this fires in the same beat that sets
+    ``blocked_since``.
+    """
+    from trueppm_api.apps.notifications.models import NotificationEventType
+    from trueppm_api.apps.notifications.services import create_event_notifications
+    from trueppm_api.apps.projects.blocker_services import (
+        render_blocker_notification,
+        resolve_impediment_recipients,
+    )
+
+    recipients = resolve_impediment_recipients(task)
+    recipients.discard(getattr(beat.actor, "pk", None))
+    if not recipients:
+        return
+    subject, body = render_blocker_notification(task)
+    recipient_ids = list(recipients)
+    _dispatch_and_backdate(
+        beat.when,
+        event_type=NotificationEventType.TASK_BLOCKED.value,
+        project_id=task.project_id,
+        recipient_ids=recipient_ids,
+        dispatch=lambda: create_event_notifications(
+            event_type=NotificationEventType.TASK_BLOCKED.value,
+            recipient_ids=recipient_ids,
+            subject=subject,
+            body=body,
+            project_id=task.project_id,
+            task_id=task.pk,
+        ),
+    )
+
+
+def _apply_task_unblock(beat: _Beat, ctx: ReplayContext) -> None:
+    """Clear the blocker. ``Task.save()`` owns the cascade — emptying
+    ``blocked_reason`` also nulls ``blocked_since``, ``blocker_type``,
+    ``blocking_task`` and ``blocked_by`` — so this only empties the flag and
+    lets the model do the rest, keeping one definition of "unblocked"."""
+    task = _resolve_task(ctx, beat.target)
+    if task is None or not (task.blocked_reason or "").strip():
+        return
+    task.blocked_reason = ""
+    _save(task, beat.when, beat.actor, ["blocked_reason"])
+
+
+def _apply_risk_note(beat: _Beat, ctx: ReplayContext) -> None:
+    """Append a RiskComment, so a risk.status flip carries its reason (#3094).
+
+    Without this a risk walks OPEN -> MITIGATING -> RESOLVED with no artifact of
+    the work: the register records *that* it was mitigated and never *how*.
+    ``created_at`` is ``auto_now_add``, so it is backdated after insert.
+    """
+    _, _, slug = beat.target.partition(":")
+    risk = ctx.risks.get(slug)
+    body = beat.data.get("body")
+    if risk is None or not body:
+        return
+    comment = RiskComment.objects.create(risk=risk, author=beat.actor, message=body)
+    RiskComment.objects.filter(pk=comment.pk).update(created_at=beat.when)
+
+
+def _apply_task_comment(beat: _Beat, ctx: ReplayContext) -> None:
+    task = _resolve_task(ctx, beat.target)
+    body = beat.data.get("body")
+    if task is None or not body:
+        return
+    # v2.1 threading (#3493). Validation holds a reply to a top-level comment on
+    # the same task; the task check is repeated because a parent on another task
+    # would render the reply under a thread it is not part of.
+    parent = ctx.comments.get(beat.data["reply_to"]) if beat.data.get("reply_to") else None
+    if parent is not None and (parent.task_id != task.pk or parent.parent_id is not None):
+        parent = None
+    comment = TaskComment.objects.create(task=task, author=beat.actor, body=body, parent=parent)
+    # created_at is auto_now_add (stamped now() on insert); backdate it.
+    TaskComment.objects.filter(pk=comment.pk).update(created_at=beat.when)
+    if beat.data.get("slug"):
+        ctx.comments[beat.data["slug"]] = comment
+    # A reply is still a comment on the task, and the live product's own
+    # _notify_assignee_of_comment applies uniformly regardless of parent — so
+    # this fires for a reply beat exactly as it does for a top-level one (#3489).
+    parsed = _fan_out_mentions(comment, task, beat)
+    _notify_comment_on_assignee(task, comment, beat, parsed)
+
+
+def _fan_out_mentions(comment: TaskComment, task: Task, beat: _Beat) -> list[ParsedMention]:
+    """Create Mention + Notification rows for an @mention in a seeded comment.
+
+    Only the live view path parsed mentions, so a seeded ``@mei`` was plain text:
+    it *read* like a mention and produced no Mention row and no notification, and
+    a persona signing in landed on an empty notification list. That is the shape
+    of defect this seed audit keeps finding — a surface that looks populated and
+    is not.
+
+    Resolution filters user mentions to current **project members**, which is why
+    this only became possible once seeded accounts held ProjectMemberships
+    (#3092). ``now=beat.when`` keeps the notification contemporaneous with the
+    comment rather than stamping import day; a mention of a non-member resolves
+    to ``skipped_users`` and is silently dropped, exactly as the live path
+    reports it.
+
+    Both ``Mention.created_at`` and ``Notification.created_at`` are
+    ``auto_now_add`` — always real wall-clock time — so the rows this call
+    writes are backdated to the beat afterward, scoped to this exact comment's
+    (freshly-created, unique) pk so a concurrent import elsewhere can never be
+    touched (#3489).
+
+    Returns the parsed mentions so the caller can de-dup the sibling
+    ``comment_on_my_task`` notification against an assignee who was also
+    directly @mentioned.
+    """
+    from trueppm_api.apps.notifications.models import Mention, Notification
+    from trueppm_api.apps.notifications.services import (
+        create_mention_notifications,
+        parse_mentions,
+        resolve_parsed_mentions,
+    )
+
+    parsed = parse_mentions(comment.body)
+    if not parsed:
+        return []
+    resolved = resolve_parsed_mentions(parsed, task.project_id)
+    if not resolved.user_targets and not resolved.group_targets:
+        return list(parsed)
+    t0 = timezone.now()
+    create_mention_notifications(
+        task_comment=comment,
+        mentioner=beat.actor,
+        parsed_result=resolved,
+        project_id=task.project_id,
+        now=beat.when,
+    )
+    Mention.objects.filter(task_comment=comment, created_at__gte=t0).update(created_at=beat.when)
+    Notification.objects.filter(mention__task_comment=comment, created_at__gte=t0).update(
+        created_at=beat.when
+    )
+    return parsed
+
+
+def _notify_comment_on_assignee(
+    task: Task, comment: TaskComment, beat: _Beat, parsed: list[ParsedMention]
+) -> None:
+    """``comment_on_my_task`` to the task's assignee (#639), mirroring
+    ``views.TaskCommentViewSet._notify_assignee_of_comment``.
+
+    Skipped when there is no assignee, the assignee wrote the comment, or the
+    assignee was already @mentioned in it — the mention path already notified
+    them, and this de-dups so one comment never pings the same person twice.
+    """
+    assignee = task.assignee
+    if assignee is None or assignee.pk == getattr(beat.actor, "pk", None):
+        return
+    mentioned_usernames = {p.value for p in parsed if p.kind == "user"}
+    if assignee.username in mentioned_usernames:
+        return
+
+    from trueppm_api.apps.notifications.models import NotificationEventType
+    from trueppm_api.apps.notifications.services import create_event_notifications
+
+    actor = beat.actor
+    full_name = getattr(actor, "get_full_name", lambda: "")()
+    author_name = full_name or getattr(actor, "username", "") or "Someone"
+    subject = f"New comment on {task.name}"
+    body = f'{author_name} commented on your task "{task.name}" in TruePPM.'
+    _dispatch_and_backdate(
+        beat.when,
+        event_type=NotificationEventType.COMMENT_ON_MY_TASK.value,
+        project_id=task.project_id,
+        recipient_ids=[assignee.pk],
+        dispatch=lambda: create_event_notifications(
+            event_type=NotificationEventType.COMMENT_ON_MY_TASK.value,
+            recipient_ids=[assignee.pk],
+            subject=subject,
+            body=body,
+            project_id=task.project_id,
+            task_id=task.pk,
+        ),
+    )
+
+
+def _apply_sprint_activate(beat: _Beat, ctx: ReplayContext) -> None:
+    sprint = _resolve_sprint(ctx, beat.target)
+    if sprint is None:
+        return
+    sprint.state = SprintState.ACTIVE
+    sprint.activated_at = beat.when
+    # Snapshot the commitment at activation from the seed's authored end state.
+    key = _sprint_key(ctx, sprint)
+    committed = ctx.final_sprint.get(key, {}).get("committed_points")
+    if committed is not None:
+        sprint.committed_points = committed
+    _save(sprint, beat.when, beat.actor, ["state", "activated_at", "committed_points"])
+
+
+def _apply_sprint_close(beat: _Beat, ctx: ReplayContext) -> None:
+    """Route the close through the real close contract (ADR-0176, #3488).
+
+    The prior version set four columns directly and stopped, so every closed
+    sample sprint carried zero ``SprintTaskOutcome`` rows and produced no
+    ``VelocitySuggestion`` — the sprint-review "what didn't finish" panel, the
+    carry-over history, and the velocity-calibration prompt all read from
+    exactly those rows, so a fresh load of every sample reviewed as empty.
+    This now calls the same snapshot/outcome/velocity functions the live
+    ``close_sprint`` drain task calls, in the same order, so a replayed close
+    produces the audit trail a real one would.
+
+    Narrowed from the live path on purpose, following ``is_seed_replay_active()``'s
+    own contract (the burndown receiver in ``receivers.py`` is the established
+    precedent): the board broadcast, the ``sprint.closed`` webhook, the
+    carry-over assignee notification, and the milestone reforecast digest are
+    all "the live side effect a real edit would trigger" and are skipped
+    outright — none of them write a row the review panel, carry-over history,
+    or velocity flow reads, and firing them here would spam a workspace
+    nobody is watching during import. ``apply_pending_disposition`` is
+    likewise skipped: its "carry" branch calls ``record_sprint_scope_change``,
+    which fires the un-suppressed ``sprint_scope_changed`` notify signal — the
+    same reason ``_apply_scope_inject`` above writes its ``SprintScopeChange``
+    row directly instead of going through that path. ``was_pending`` on the
+    outcome row is unaffected: it is read by ``snapshot_sprint_task_outcomes``
+    below, before any disposition would run, exactly as the live close orders
+    it.
+
+    Every write here — the sprint's own state/closed_at/goal_outcome/
+    completed_*, and any task the carry-over moves — goes through this
+    module's ``_save()`` rather than a bare ``instance.save()``, so its history
+    row is backdated to the beat like every other write in this file. Calling
+    ``services.apply_carry_over`` directly instead of ``_replay_carry_over``
+    below was the first attempt: it produced the right FK moves but dated the
+    *moved task's own* history row to import time, which
+    ``test_authored_percent_restore_is_backdated_not_import_time`` catches
+    immediately — the fixture's own sprint closes with in-flight tasks still
+    on it. ``_replay_carry_over`` re-applies the identical two classification
+    tuples the real service uses (imported, not copied by value, so a future
+    change to which statuses carry stays in sync) through ``_save()`` instead.
+    """
+    sprint = _resolve_sprint(ctx, beat.target)
+    if sprint is None:
+        return
+
+    from trueppm_api.apps.projects.services import (
+        snapshot_completed_metrics,
+        snapshot_sprint_task_outcomes,
+    )
+    from trueppm_api.apps.scheduling.services import compute_velocity_suggestions
+
+    # ADR-0176 §2 / #3488: derive completed_*/the default goal_outcome from
+    # current task state before anything else moves, exactly like the live
+    # close. A synthesized close (no authored `sprint.close` beat) has no
+    # `ctx.final_sprint` goal_outcome — the importer never populates one for a
+    # synthesized beat (see `_synthesize_sprints`) — so without this every
+    # such sprint closed with `goal_outcome=None`; this derives a real
+    # MET/PARTIAL/MISSED verdict from committed vs. completed points instead.
+    snapshot_completed_metrics(sprint)
+    if beat.data.get("goal_outcome"):
+        # An authored value is the seed author's editorial call (e.g. "hit the
+        # points but missed the actual goal") and overrides the points-ratio
+        # default — mirroring the SCHEDULER+ override the live product allows
+        # after close.
+        sprint.goal_outcome = beat.data["goal_outcome"]
+    sprint.state = SprintState.COMPLETED
+    sprint.closed_at = beat.when
+    _save(
+        sprint,
+        beat.when,
+        beat.actor,
+        ["state", "closed_at", "goal_outcome", "completed_points", "completed_task_count"],
+    )
+
+    # ADR-0176 §2: snapshot task membership-at-close BEFORE the carry-over
+    # mutates Task.sprint — same order as the live close — otherwise the
+    # "what didn't ship" set would already be gone by the time this reads it.
+    carry_over_to = _implied_carry_over_target(ctx, sprint)
+    snapshot_sprint_task_outcomes(sprint, carry_over_to=carry_over_to)
+    carried_task_ids = _replay_carry_over(sprint, carry_over_to, beat.when, beat.actor)
+    _notify_carry_over(sprint, carry_over_to, carried_task_ids, beat)
+
+    compute_velocity_suggestions(sprint.pk)
+
+
+def _replay_carry_over(sprint: Sprint, carry_over_to: str, when: datetime, actor: Any) -> list[str]:
+    """Mirror ``services.apply_carry_over`` (ADR-0176) with a backdated write.
+
+    The real close's ``task.save()`` has no beat to backdate to — correct
+    there, timestamped wrong here (#3488): a replayed carry-over must date the
+    moved task's history row to the beat, not to import time, like every
+    other write in this file. The move rules are the same two classification
+    tuples the real service reads (imported, not copied by value, so this
+    stays in sync if the close contract's carry-over policy changes) — only
+    *how* the result is saved differs.
+
+    Returns the moved tasks' ids so the caller can fan out the
+    ``task.moved_sprint`` notification (#3489) against exactly the set this
+    call actually carried.
+    """
+    from trueppm_api.apps.projects.services import (
+        _CARRY_OVER_INCOMPLETE_STATUSES,
+        _DECOMMITTED_ON_CARRY_OVER_TO_BACKLOG,
+    )
+
+    if carry_over_to == "none":
+        return []
+
+    incomplete = Task.objects.filter(
+        sprint_id=sprint.pk, status__in=_CARRY_OVER_INCOMPLETE_STATUSES, is_deleted=False
+    )
+    carried_ids: list[str] = []
+    for task in incomplete:
+        fields = ["sprint"]
+        if carry_over_to == "backlog":
+            task.sprint = None
+            if task.status in _DECOMMITTED_ON_CARRY_OVER_TO_BACKLOG:
+                task.status = TaskStatus.BACKLOG
+                fields.append("status")
+        else:
+            task.sprint_id = carry_over_to
+        _save(task, when, actor, fields)
+        carried_ids.append(str(task.pk))
+    return carried_ids
+
+
+def _notify_carry_over(
+    sprint: Sprint, carry_over_to: str, carried_task_ids: list[str], beat: _Beat
+) -> None:
+    """``task.moved_sprint`` to each carried task's assignee (#1470/ADR-0232),
+    reusing the real grouping/copy helpers (``services._carried_tasks_by_assignee``
+    / ``_carryover_row``) so a demo persona with several carried tasks gets the
+    same one-row-per-assignee summary the live close produces.
+    """
+    if not carried_task_ids:
+        return
+
+    from trueppm_api.apps.notifications.models import NotificationEventType
+    from trueppm_api.apps.notifications.services import create_event_notifications_batch
+    from trueppm_api.apps.projects.services import (
+        _carried_tasks_by_assignee,
+        _carry_over_destination_label,
+        _carryover_row,
+    )
+
+    actor_pk = getattr(beat.actor, "pk", None)
+    by_assignee = _carried_tasks_by_assignee(carried_task_ids, actor_pk)
+    if not by_assignee:
+        return
+    destination = _carry_over_destination_label(carry_over_to)
+    rows = [
+        _carryover_row(assignee_id, tasks, origin=sprint.name, destination=destination)
+        for assignee_id, tasks in by_assignee.items()
+    ]
+    recipient_ids = list(by_assignee.keys())
+    _dispatch_and_backdate(
+        beat.when,
+        event_type=NotificationEventType.TASK_MOVED_SPRINT.value,
+        project_id=sprint.project_id,
+        recipient_ids=recipient_ids,
+        dispatch=lambda: create_event_notifications_batch(
+            event_type=NotificationEventType.TASK_MOVED_SPRINT.value,
+            project_id=sprint.project_id,
+            rows=rows,
+        ),
+    )
+
+
+def _apply_scope_inject(beat: _Beat, ctx: ReplayContext) -> None:
+    """Record a mid-sprint scope injection as the real audit row (VoC/Alex).
+
+    Creates the SprintScopeChange the drawer chip + Enterprise audit read, and
+    flags the task pending — the same row the live ``record_sprint_scope_change``
+    path writes — without firing the notify signal (suppressed during replay).
+
+    The injected sprint is ``task.sprint`` whenever the task still holds one;
+    see :func:`_injected_sprint` for the case a re-imported export creates,
+    where it no longer does.
+    """
+    task = _resolve_task(ctx, beat.target)
+    if task is None:
+        return
+    sprint = _injected_sprint(task, beat, ctx)
+    if sprint is None:
+        return
+    scope = SprintScopeChange.objects.create(
+        task=task,
+        sprint=sprint,
+        subtask_name=task.name,
+        added_by=beat.actor,
+        goal_impact=bool(beat.data.get("goal_impact", False)),
+        status=ScopeChangeStatus.PENDING,
+    )
+    # added_at is auto_now_add (stamped now() on insert); backdate it to the beat
+    # so the scope-injection audit reads chronologically and — for a still-PENDING
+    # injection — the exporter can reconstruct a deterministic scope_inject event
+    # from this row (the round-trip fixpoint depends on a stable timestamp).
+    SprintScopeChange.objects.filter(pk=scope.pk).update(added_at=beat.when)
+    ctx.open_scope[task.pk] = scope
+    task.sprint_pending = True
+    _save(task, beat.when, beat.actor, ["sprint_pending"])
+    _notify_scope_inject(task, sprint, beat)
+
+
+def _notify_scope_inject(task: Task, sprint: Sprint, beat: _Beat) -> None:
+    """``sprint.membership_changed`` to the project lead cohort (#1946/ADR-0412),
+    reusing the real recipient/copy helpers (``services._sprint_lead_recipient_ids``
+    / ``_sprint_change_body``).
+
+    Fires only when the injected sprint is currently ACTIVE, mirroring the live
+    trigger's "a live-commitment change" gate — a scope change into a
+    PLANNED/COMPLETED/CANCELLED sprint carries no accountability signal.
+    """
+    if sprint.state != SprintState.ACTIVE:
+        return
+
+    from trueppm_api.apps.notifications.models import NotificationEventType
+    from trueppm_api.apps.notifications.services import create_event_notifications_batch
+    from trueppm_api.apps.projects.services import _sprint_change_body, _sprint_lead_recipient_ids
+
+    actor_pk = getattr(beat.actor, "pk", None)
+    recipient_ids = _sprint_lead_recipient_ids(task.project_id, actor_pk)
+    if not recipient_ids:
+        return
+    actor_name = getattr(beat.actor, "username", "") or "Someone"
+    body = _sprint_change_body(
+        actor_name,
+        task.name,
+        None,
+        {"name": sprint.name},
+        entered_active=True,
+        left_active=False,
+    )
+    rows = [(rid, "Sprint scope changed", body, str(task.pk)) for rid in recipient_ids]
+    _dispatch_and_backdate(
+        beat.when,
+        event_type=NotificationEventType.SPRINT_MEMBERSHIP_CHANGED.value,
+        project_id=task.project_id,
+        recipient_ids=recipient_ids,
+        dispatch=lambda: create_event_notifications_batch(
+            event_type=NotificationEventType.SPRINT_MEMBERSHIP_CHANGED.value,
+            project_id=task.project_id,
+            rows=rows,
+        ),
+    )
+
+
+def _injected_sprint(task: Task, beat: _Beat, ctx: ReplayContext) -> Sprint | None:
+    """The sprint a ``sprint.scope_inject`` beat injects ``task`` into.
+
+    Normally read straight off ``task.sprint``: the documented contract (see
+    ``validation._EVENT_TARGET_KIND``) is that the beat targets the task and its
+    sprint is derived, matching ``record_sprint_scope_change(task, sprint)``,
+    where every live call site links the task to the sprint first and records
+    the audit row second. That derivation holds for every hand-authored seed,
+    because a document declares a task's ``sprint`` statically before it can
+    date an injection against it.
+
+    It stops holding for a *re-imported export* (#3488). The exporter writes each
+    task's ``sprint`` from its **current** membership, and a replayed sprint close
+    now carries an unfinished task back to the backlog — so a still-PENDING
+    ``SprintScopeChange``, which the exporter reconstructs as a
+    ``sprint.scope_inject`` beat, can name a task that by export time holds no
+    sprint at all. Born sprintless, the beat wrote no row, the re-export dropped
+    the event, and the #616 byte-identical round trip failed.
+
+    The fallback is a derivation, not a guess: an injection is by definition a
+    link into the sprint that is *running at that moment* (ADR-0102 §4), so the
+    beat's own instant identifies it — the one sprint in the task's project that
+    replay has already activated, has not yet closed, and whose dates contain the
+    beat. Ambiguity is refused rather than resolved: nothing in the schema stops a
+    seed authoring two overlapping active sprints in one project, and picking one
+    of them arbitrarily would file the audit row against the wrong sprint.
+    """
+    if task.sprint_id is not None:
+        return task.sprint
+    _, _, ref = beat.target.partition(":")
+    project_slug, _, _ = ref.partition(":")
+    day = beat.when.date()
+    candidates = [
+        sprint
+        for key, sprint in ctx.sprints.items()
+        if key[0] == project_slug
+        and sprint.state == SprintState.ACTIVE
+        and sprint.start_date <= day <= sprint.finish_date
+    ]
+    return candidates[0] if len(candidates) == 1 else None
+
+
+def _apply_scope_resolve(beat: _Beat, ctx: ReplayContext) -> None:
+    task = _resolve_task(ctx, beat.target)
+    if task is None:
+        return
+    scope = ctx.open_scope.pop(task.pk, None)
+    accepted = (beat.data.get("to") or "ACCEPTED").upper() != "REJECTED"
+    if scope is not None:
+        scope.status = ScopeChangeStatus.ACCEPTED if accepted else ScopeChangeStatus.REJECTED
+        scope.save(update_fields=["status"])
+    task.sprint_pending = False
+    fields = ["sprint_pending"]
+    if not accepted:
+        task.sprint = None
+        fields.append("sprint")
+    _save(task, beat.when, beat.actor, fields)
+
+
+def _apply_risk_status(beat: _Beat, ctx: ReplayContext) -> None:
+    _, _, slug = beat.target.partition(":")
+    risk = ctx.risks.get(slug)
+    if risk is None:
+        return
+    new_status = beat.data.get("to")
+    if not new_status or new_status == risk.status:
+        return
+    risk.status = new_status
+    _save(risk, beat.when, beat.actor, ["status"])
+
+
+def _apply_baseline_capture(beat: _Beat, ctx: ReplayContext) -> None:
+    """Capture a baseline of the project's current task dates at the beat time.
+
+    ``is_active`` (default False, #3495) collapses the real app's two-step
+    capture-then-activate flow (``BaselineViewSet.perform_create`` +
+    ``BaselineActivateView``) into one beat: when true, any baseline already
+    active for the project is deactivated first — the same swap
+    ``BaselineActivateView`` performs — so an authored rebaseline can supersede
+    the one it replaces in a single timeline entry instead of leaving the
+    project with two inactive baselines or an ``IntegrityError`` from
+    ``unique_active_baseline_per_project``.
+    """
+    _, _, project_slug = beat.target.partition(":")
+    project = ctx.projects.get(project_slug)
+    if project is None:
+        return
+    is_active = bool(beat.data.get("is_active", False))
+    if is_active:
+        Baseline.objects.filter(project=project, is_active=True).update(is_active=False)
+    baseline = Baseline.objects.create(
+        project=project,
+        name=beat.data.get("body") or f"Baseline {beat.when.date().isoformat()}",
+        is_active=is_active,
+    )
+    Baseline.objects.filter(pk=baseline.pk).update(created_at=beat.when)
+    rows = [
+        BaselineTask(
+            baseline=baseline,
+            task_id=task.pk,
+            task_name=task.name,
+            start=task.planned_start,
+            finish=task.planned_start + timedelta(days=task.duration)
+            if task.planned_start
+            else None,
+            duration=task.duration,
+            story_points=task.story_points,
+        )
+        for (slug, _), task in ctx.tasks.items()
+        if slug == project_slug
+    ]
+    BaselineTask.objects.bulk_create(rows)
+
+
+def _apply_retro_action(beat: _Beat, ctx: ReplayContext) -> None:
+    """Create a RetroActionItem on the target sprint's retro (ADR-0114 §7).
+
+    The parent ``SprintRetro`` is created lazily the first time an action item
+    attaches to the sprint (mirroring the live ``_get_or_create_retro`` path),
+    so a seed authors retro outcomes without a separate "open retro" event. An
+    optional ``notes`` on the beat sets the retro's own meeting-summary text
+    (distinct from the action item's ``body``) — a seed typically carries it on
+    the last action-item beat for that sprint, once the retro has something to
+    summarize (#3497).
+    """
+    sprint = _resolve_sprint(ctx, beat.target)
+    body = beat.data.get("body")
+    if sprint is None or not body:
+        return
+    retro, created = SprintRetro.objects.get_or_create(
+        sprint=sprint, defaults={"created_by": beat.actor}
+    )
+    if created:
+        # created_at is auto_now_add; backdate the retro to its first action so
+        # the demo's retro is dated to the ceremony, not import time.
+        SprintRetro.objects.filter(pk=retro.pk).update(created_at=beat.when)
+    notes = beat.data.get("notes")
+    if notes:
+        retro.notes = notes
+        _save(retro, beat.when, beat.actor, ["notes"])
+    assignee = ctx.users.get(beat.data["assignee"]) if beat.data.get("assignee") else None
+    item = RetroActionItem.objects.create(
+        retro=retro,
+        text=body,
+        assignee=assignee,
+        story_points=beat.data.get("points"),
+    )
+    RetroActionItem.objects.filter(pk=item.pk).update(created_at=beat.when)
+
+
+def _apply_retro_promote(beat: _Beat, ctx: ReplayContext) -> None:
+    """Promote a retro action item (matched by ``body``) to a backlog task.
+
+    Mirrors ``promote_retro_action_item`` but writes directly (no on_commit
+    broadcast / recalc enqueue) because replay is side-effect-suppressed — the
+    importer enqueues a single per-project recalc after commit. The resulting
+    Task is a project-backlog item (``status=BACKLOG``, ``sprint=None``) exactly
+    as the live promote path produces, so the demo shows the retro→task loop
+    closed with a real ``T-XXX`` link.
+    """
+    sprint = _resolve_sprint(ctx, beat.target)
+    body = beat.data.get("body")
+    if sprint is None or not body:
+        return
+    retro = SprintRetro.objects.filter(sprint=sprint).first()
+    if retro is None:
+        return
+    item = (
+        RetroActionItem.objects.filter(
+            retro=retro, text=body, promoted_task_id__isnull=True, is_deleted=False
+        )
+        .order_by("created_at")
+        .first()
+    )
+    if item is None:
+        return  # nothing (left) to promote for this text
+    short_id = sprint.short_id or str(sprint.pk)[:8]
+    task = Task(
+        project=sprint.project,
+        name=body[:255],
+        duration=1,
+        status=TaskStatus.BACKLOG,
+        sprint=None,
+        assignee=item.assignee,
+        story_points=item.story_points,
+        notes=f'source: "retrospective" (from Sprint {short_id} retro)',
+    )
+    _save_new_backdated(task, beat.when, beat.actor)
+    item.promoted_task_id = task.pk
+    item.save(update_fields=["promoted_task_id"])
+
+
+# --- notifications -------------------------------------------------------------
+#
+# #3489: fresh samples produced notifications only from the mention fan-out
+# above (`_fan_out_mentions`) — every other beat wrote through the ORM and
+# never touched `notifications.services`, so a persona signing in saw an
+# empty bell. The handlers above call the small `_notify_*` functions added
+# alongside them; this section holds the cross-cutting backdating helper plus
+# the two event types with no single triggering beat (a synthesized
+# `task.stale` sweep and the sponsor's `program.health_digest`) and the final
+# read/unread/archived/snoozed realism pass.
+
+
+def _dispatch_and_backdate(
+    when: datetime,
+    *,
+    event_type: str,
+    project_id: Any,
+    recipient_ids: Sequence[Any],
+    dispatch: Callable[[], Any],
+) -> None:
+    """Run a ``notifications.services`` dispatch, then backdate the rows it wrote.
+
+    Every dispatcher in that module stamps ``Notification.created_at`` via
+    ``auto_now_add`` — always real wall-clock time, correct for the live
+    product and wrong for a replayed beat. This captures the instant right
+    before dispatch and rewrites every matching row to ``when`` afterward.
+
+    The filter is scoped to ``(recipient, project, event_type, created_at >= t0)``.
+    ``project_id`` is always a project this same import just created (or
+    ``None`` for the one account-scoped digest, whose recipient is likewise a
+    user this import just created), so the window can never reach a row a
+    concurrent, unrelated import or request wrote in the same instant.
+    """
+    ids = [rid for rid in recipient_ids if rid is not None]
+    if not ids:
+        return
+
+    from trueppm_api.apps.notifications.models import Notification
+
+    t0 = timezone.now()
+    dispatch()
+    Notification.objects.filter(
+        recipient_id__in=ids,
+        project_id=project_id,
+        event_type=event_type,
+        created_at__gte=t0,
+    ).update(created_at=when)
+
+
+def _seed_notification_preferences(payload: dict[str, Any], ctx: ReplayContext) -> None:
+    """Backfill every persona's default ``NotificationPreference`` rows, plus one
+    ``ProjectNotificationPreference`` with a non-default quiet-hours window for
+    the program's OWNER persona (#3489's "preference pages render seeded rows,
+    not defaults" acceptance bar).
+    """
+    import datetime as _dt
+
+    from trueppm_api.apps.notifications.models import ProjectNotificationPreference
+    from trueppm_api.apps.notifications.services import get_or_create_default_preferences
+
+    for user in ctx.users.values():
+        if user is not None:
+            get_or_create_default_preferences(user)
+
+    owner_slug = next(
+        (a["slug"] for a in payload.get("accounts", []) if a.get("role") == "OWNER"), None
+    )
+    owner = ctx.users.get(owner_slug) if owner_slug else None
+    project = next(iter(ctx.projects.values()), None)
+    if owner is None or project is None:
+        return
+    ProjectNotificationPreference.objects.update_or_create(
+        project=project,
+        user=owner,
+        defaults={
+            "quiet_hours_enabled": True,
+            "quiet_hours_from": _dt.time(18, 0),
+            "quiet_hours_until": _dt.time(8, 0),
+        },
+    )
+
+
+def _synthesize_stale_notifications(ctx: ReplayContext) -> None:
+    """Seed-local ``task.stale`` sweep (ADR-0200), scoped to this program's own tasks.
+
+    Deliberately NOT ``notifications.services.create_stale_task_notifications``:
+    that function scans every project in the install against the real wall
+    clock, which would (a) sweep already-imported, unrelated samples and (b)
+    stamp ``created_at`` at real import time rather than a date this replay
+    can own. This mirrors its eligibility rule (every non-``COMPLETE`` status,
+    an assignee set, untouched past the project's ``stale_task_threshold_days``)
+    but dates each notification to when the task actually went stale.
+    """
+    from trueppm_api.apps.notifications.models import NotificationEventType
+    from trueppm_api.apps.notifications.services import (
+        DEFAULT_STALE_TASK_THRESHOLD_DAYS,
+        create_event_notifications_batch,
+    )
+
+    for (project_slug, _wbs), task in ctx.tasks.items():
+        if task.status == TaskStatus.COMPLETE or task.assignee_id is None:
+            continue
+        if task.status_changed_at is None:
+            continue
+        project = ctx.projects[project_slug]
+        threshold = project.stale_task_threshold_days or DEFAULT_STALE_TASK_THRESHOLD_DAYS
+        stale_since = task.status_changed_at.date() + timedelta(days=threshold)
+        if stale_since > ctx.anchor:
+            continue
+        display_name = task.name if len(task.name) <= 200 else task.name[:200].rstrip() + "…"
+        subject = f'"{display_name}" has gone stale'
+        body = (
+            f'Your task "{display_name}" has sat in the same status for more than '
+            f"{threshold} days. If it is still active, move it forward; "
+            f"otherwise update its status so the board reflects reality."
+        )
+        when = _aware_dt(ctx, stale_since, 8, 0)
+        rows = [(task.assignee_id, subject, body, str(task.pk))]
+        _dispatch_and_backdate(
+            when,
+            event_type=NotificationEventType.TASK_STALE.value,
+            project_id=task.project_id,
+            recipient_ids=[task.assignee_id],
+            # functools.partial (not a lambda closing over the loop variables)
+            # so each call is bound to THIS iteration's rows/project, not
+            # whatever the loop variable holds by the time it runs.
+            dispatch=functools.partial(
+                create_event_notifications_batch,
+                event_type=NotificationEventType.TASK_STALE.value,
+                project_id=task.project_id,
+                rows=rows,
+            ),
+        )
+
+
+def _synthesize_program_digest(payload: dict[str, Any], ctx: ReplayContext) -> None:
+    """One ``program.health_digest`` row for the sponsor persona — the seed's
+    VIEWER account, the read-only executive role every sample carries exactly
+    one of — dated to the most recent Sunday evening on/before the anchor,
+    matching the real weekly send slot (ADR-0663).
+
+    ``PROGRAM_HEALTH_DIGEST`` defaults OFF on both channels (a weekly send
+    nobody asked for is exactly the noise the default guards against), so the
+    sponsor is opted in explicitly here first — otherwise the dispatch below
+    would silently create nothing.
+    """
+    accounts = payload.get("accounts", [])
+    sponsor_slug = next((a["slug"] for a in accounts if a.get("role") == "VIEWER"), None)
+    if sponsor_slug is None:
+        return
+    sponsor = ctx.users.get(sponsor_slug)
+    if sponsor is None:
+        return
+
+    from trueppm_api.apps.notifications.digests import build_program_health_digest
+    from trueppm_api.apps.notifications.models import (
+        NotificationChannel,
+        NotificationEventType,
+        NotificationPreference,
+    )
+    from trueppm_api.apps.notifications.services import create_event_notifications
+
+    # `_seed_notification_preferences` already backfilled the sponsor's default
+    # rows earlier in `replay_timeline`; only the digest event's opt-in (below)
+    # needs an explicit override.
+    NotificationPreference.objects.update_or_create(
+        user=sponsor,
+        event_type=NotificationEventType.PROGRAM_HEALTH_DIGEST.value,
+        channel=NotificationChannel.IN_APP.value,
+        defaults={"enabled": True},
+    )
+
+    digest_day = ctx.anchor - timedelta(days=(ctx.anchor.weekday() + 1) % 7)
+    when = _aware_dt(ctx, digest_day, 18, 0)
+    subject, body = build_program_health_digest(sponsor, when)
+    _dispatch_and_backdate(
+        when,
+        event_type=NotificationEventType.PROGRAM_HEALTH_DIGEST.value,
+        project_id=None,
+        recipient_ids=[sponsor.pk],
+        dispatch=lambda: create_event_notifications(
+            event_type=NotificationEventType.PROGRAM_HEALTH_DIGEST.value,
+            recipient_ids=[sponsor.pk],
+            subject=subject,
+            body=body,
+            project_id=None,
+        ),
+    )
+
+
+def _apply_notification_realism(ctx: ReplayContext) -> None:
+    """Age the freshly-emitted inbox into a mix of read / unread / archived /
+    snoozed rows (#3489) — a batch created in one instant and left untouched
+    would all read as equally fresh, which is its own tell.
+
+    Rows backdated more than ~5 days before the anchor are marked read; a
+    deterministic handful of those are archived; one still-unread row is
+    snoozed into the demo's near future. Recent rows are left unread so the
+    bell has something to show. Scoped to this program's own projects (plus,
+    for the one account-scoped digest, this program's own users) so a
+    re-import of a different sample can never be touched.
+    """
+    from trueppm_api.apps.notifications.models import Notification
+
+    project_ids = [p.pk for p in ctx.projects.values()]
+    user_ids = [u.pk for u in ctx.users.values() if u is not None]
+    qs = Notification.objects.filter(
+        Q(project_id__in=project_ids) | Q(project_id__isnull=True, recipient_id__in=user_ids)
+    )
+
+    read_cutoff = _aware(ctx, ctx.anchor - timedelta(days=5))
+    qs.filter(created_at__lt=read_cutoff, is_read=False).update(
+        is_read=True, read_at=F("created_at") + timedelta(hours=2)
+    )
+
+    archive_ids = list(
+        qs.filter(is_read=True, is_archived=False)
+        .order_by("created_at")
+        .values_list("pk", flat=True)[:3]
+    )
+    if archive_ids:
+        Notification.objects.filter(pk__in=archive_ids).update(is_archived=True)
+
+    snooze_candidate = (
+        qs.filter(is_read=False, is_archived=False, snoozed_until__isnull=True)
+        .order_by("-created_at")
+        .first()
+    )
+    if snooze_candidate is not None:
+        Notification.objects.filter(pk=snooze_candidate.pk).update(
+            snoozed_until=_aware_dt(ctx, ctx.anchor + timedelta(days=2), 9, 0)
+        )
+
+
+# --- helpers -----------------------------------------------------------------
+
+
+def _save_new_backdated(instance: Any, when: datetime, user: Any) -> None:
+    """Insert a brand-new instance, dating its creation history row to ``when``.
+
+    Unlike ``_save`` (which uses ``update_fields`` for a narrow update), this is
+    for a first insert: simple-history honors ``_history_date`` / ``_history_user``
+    on the create so the row is attributed and backdated in one write.
+    """
+    instance._history_date = when
+    instance._history_user = user
+    instance.save()
+
+
+def _save(instance: Any, when: datetime, user: Any, fields: list[str]) -> None:
+    """Save ``instance`` writing a history row dated ``when`` by ``user``.
+
+    django-simple-history honors ``_history_date``/``_history_user`` set on the
+    instance before save (ADR-0114 — the first backdated-history writer in the
+    codebase). update_fields keeps the write narrow.
+    """
+    instance._history_date = when
+    instance._history_user = user
+    instance.save(update_fields=fields)
+
+
+def _aware(ctx: ReplayContext, day: date) -> datetime:
+    return _aware_dt(ctx, day, 12, 0)
+
+
+def _aware_dt(ctx: ReplayContext, day: date, hour: int, minute: int) -> datetime:
+    """Build an aware datetime in the program timezone (default UTC)."""
+    tz = ctx.tz or ZoneInfo("UTC")
+    return datetime(day.year, day.month, day.day, hour, minute, tzinfo=tz)
+
+
+def _sprint_key(ctx: ReplayContext, sprint: Sprint) -> tuple[str, str]:
+    for key, candidate in ctx.sprints.items():
+        if candidate.pk == sprint.pk:
+            return key
+    return ("", "")
+
+
+def _implied_carry_over_target(ctx: ReplayContext, sprint: Sprint) -> str:
+    """Infer this close's carry-over destination from the seed's own timeline (#3488).
+
+    The live close reads ``carry_over_to`` off the API request; a replayed
+    close has none, so it is read off the seed's intent instead: the next
+    sprint in the same project that the timeline actually runs (state ACTIVE
+    or COMPLETED in ``ctx.final_sprint`` — i.e. it gets an authored or
+    synthesized ``sprint.activate`` beat somewhere in the timeline) is where
+    an incomplete task would land. No such sprint — this closing sprint is
+    the project's last to actually run — falls back to ``"backlog"``, the
+    same default ``POST .../close/`` uses.
+    """
+    project_slug, _ = _sprint_key(ctx, sprint)
+    candidates = [
+        (candidate.start_date, candidate.pk)
+        for key, candidate in ctx.sprints.items()
+        if key[0] == project_slug
+        and candidate.pk != sprint.pk
+        and candidate.start_date > sprint.finish_date
+        and ctx.final_sprint.get(key, {}).get("state")
+        in (SprintState.ACTIVE, SprintState.COMPLETED)
+    ]
+    if not candidates:
+        return "backlog"
+    candidates.sort(key=lambda c: c[0])
+    return str(candidates[0][1])
+
+
+def _named_actor(beat: _Beat, ctx: ReplayContext) -> Any:
+    """The account a beat names, without the importing-owner fallback.
+
+    ``_Beat.actor`` falls back to the importing owner, which is right for a
+    status move or a comment but wrong for a fact whose attribution *is* the
+    fact: an hour logged, a reaction, an acknowledgement. On the generic import
+    path a pre-existing real account resolves to ``None`` (#1057); crediting its
+    hours to the importer instead would fabricate the importer's own timesheet.
+    """
+    slug = beat.data.get("actor")
+    return ctx.users.get(slug) if slug else None
+
+
+def _comment_for(ctx: ReplayContext, target: str) -> TaskComment | None:
+    _, _, slug = target.partition(":")
+    return ctx.comments.get(slug)
+
+
+def _apply_task_note(beat: _Beat, ctx: ReplayContext) -> None:
+    """Append a dated ``TaskNote`` — the task's why/decision log (ADR-0143, #3492).
+
+    ``decision`` is what the Decisions view reads. ``created_at`` is
+    ``auto_now_add``, so it is backdated after insert like a comment.
+    """
+    task = _resolve_task(ctx, beat.target)
+    body = beat.data.get("body")
+    if task is None or not body:
+        return
+    note = TaskNote.objects.create(
+        task=task,
+        author=beat.actor,
+        body=body,
+        pinned=bool(beat.data.get("pinned", False)),
+        decision=bool(beat.data.get("decision", False)),
+    )
+    TaskNote.objects.filter(pk=note.pk).update(created_at=beat.when)
+
+
+def _apply_task_react(beat: _Beat, ctx: ReplayContext) -> None:
+    """Add a reaction to a slugged comment (ADR-0075 §A.4). Never notifies."""
+    comment = _comment_for(ctx, beat.target)
+    user = _named_actor(beat, ctx)
+    emoji = beat.data.get("emoji")
+    if comment is None or user is None or not emoji:
+        return
+    reaction, created = CommentReaction.objects.get_or_create(
+        comment=comment, user=user, emoji=emoji
+    )
+    if created:
+        CommentReaction.objects.filter(pk=reaction.pk).update(created_at=beat.when)
+
+
+def _apply_task_ack(beat: _Beat, ctx: ReplayContext) -> None:
+    """Acknowledge a slugged comment — "I saw this / I'm on it" (ADR-0075 §A.3)."""
+    comment = _comment_for(ctx, beat.target)
+    user = _named_actor(beat, ctx)
+    if comment is None or user is None:
+        return
+    ack, created = CommentAcknowledgement.objects.get_or_create(comment=comment, user=user)
+    if created:
+        CommentAcknowledgement.objects.filter(pk=ack.pk).update(created_at=beat.when)
+
+
+def _apply_time_log(beat: _Beat, ctx: ReplayContext) -> None:
+    """Write a ``TimeEntry`` dated to the beat (ADR-0185, #3490).
+
+    Never forward-dated: a beat after the anchor is dropped, because logged time
+    records work that happened and the demo's "today" is the anchor. Validation
+    already rejects a relative ``A+N`` beat; this also covers an ISO literal.
+    """
+    task = _resolve_task(ctx, beat.target)
+    user = _named_actor(beat, ctx)
+    minutes = beat.data.get("minutes")
+    if task is None or user is None or not minutes:
+        return
+    entry_date = beat.when.date()
+    if entry_date > ctx.anchor:
+        return
+    entry = TimeEntry.objects.create(
+        task=task,
+        user=user,
+        minutes=minutes,
+        entry_date=entry_date,
+        note=beat.data.get("note", ""),
+        source=beat.data.get("source", "manual"),
+    )
+    TimeEntry.objects.filter(pk=entry.pk).update(created_at=beat.when)
+
+
+_HANDLERS = {
+    "task.note": _apply_task_note,
+    "task.react": _apply_task_react,
+    "task.ack": _apply_task_ack,
+    "time.log": _apply_time_log,
+    _TASK_STATUS_ACTION: _apply_task_status,
+    "task.assign": _apply_task_assign,
+    "task.estimate": _apply_task_estimate,
+    "task.points": _apply_task_points,
+    "task.ac_met": _apply_task_ac_met,
+    "task.block": _apply_task_block,
+    "task.unblock": _apply_task_unblock,
+    "task.comment": _apply_task_comment,
+    "sprint.activate": _apply_sprint_activate,
+    "sprint.close": _apply_sprint_close,
+    "sprint.scope_inject": _apply_scope_inject,
+    "sprint.scope_resolve": _apply_scope_resolve,
+    "risk.status": _apply_risk_status,
+    "risk.note": _apply_risk_note,
+    "baseline.capture": _apply_baseline_capture,
+    "retro.action": _apply_retro_action,
+    "retro.promote": _apply_retro_promote,
+}