P2-02 (agent side) + P2-03: agent scheduler + schedule.fire dispatch

Closes the schedule reconciliation loop end-to-end.

* New `internal/agent/scheduler` package wraps robfig/cron/v3 with
  the lifecycle the agent needs:
  - Apply(ScheduleSetPayload, Sender) stops the prior cron (waiting
    for in-flight entries to return), rebuilds from scratch, starts,
    and emits schedule.ack with the version we just applied.
  - Disabled entries skipped silently; bad cron exprs (which
    shouldn't reach us — the server validates — but defensive)
    log a warn and skip.
  - On each cron tick the entry sends a new schedule.fire envelope
    to the server with {schedule_id, scheduled_at}. The scheduler
    itself never builds CommandRunPayloads — server is the source
    of truth for jobs.
  - tx is swapped on every Apply, so reconnect is handled
    naturally: cron entries that fire against a dropped tx log
    "no active connection" and skip the tick.
  - Stop() is idempotent and waits for the cron's in-flight
    workers via cron.Stop().Done().

* New wire message api.MsgScheduleFire + api.ScheduleFirePayload
  for the agent → server "I just fired locally" RPC.

* Server-side dispatch (schedule_push.go: dispatchScheduledJob):
  looks up the schedule by id, validates ownership + that it's
  enabled, builds args from kind (paths for backup; other kinds
  are still arg-less in Phase 2 and grow as those job kinds land
  in P2-05..08), persists a jobs row with actor_kind=schedule +
  scheduled_id, and writes command.run back on the same conn so
  the agent runs through its existing dispatch path.

* store.CreateJob now writes scheduled_id. This column was in the
  schema since 0001 but never populated — the original P1 path
  only had operator-driven jobs, so actor_kind was always 'user'
  and scheduled_id was always nil.

* cmd/agent/main.go integration: dispatcher gains a
  *scheduler.Scheduler; the MsgScheduleSet case now hands the
  payload to scheduler.Apply (in a goroutine so the WS read loop
  keeps draining other messages).

* WS dispatcher gains OnScheduleFire alongside OnScheduleAck.

* Tests:
  - scheduler unit tests (4): ack-on-apply, cron tick fires
    schedule.fire envelope, disabled entries don't fire, replace-
    prior-state stops the old cron.
  - Server-side end-to-end: schedule.fire → command.run with the
    right job_id / kind / args, plus jobs row with actor_kind=
    "schedule" and scheduled_id linking back to the schedule.

Persistence of next-fire times across agent restarts is
deliberately deferred. A missed fire window during downtime
simply fires once on reconnect — that's the desirable behaviour
(the operator wants the missed backup to run, not be silently
skipped because we lost track of when it was due).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

tasks.md

@@ -98,8 +98,8 @@ Sizes: **S** = under a day, **M** = 1–3 days, **L** = 3–7 days.
 ## Phase 2 — Scheduling, retention, repo operations
 
 - [x] **P2-01** (M) Schedule schema + CRUD API. `schedules` table was already laid down in 0001; this slice adds `store.Schedule`/`RetentionPolicy`/`ScheduleOptions` types, `CreateSchedule` / `GetSchedule` / `ListSchedulesByHost` / `UpdateSchedule` / `DeleteSchedule` / `GetHostScheduleVersion` / `SetHostAppliedScheduleVersion` (mutations bump `host_schedule_version` atomically in-tx), and REST endpoints `GET|POST /api/hosts/{id}/schedules` + `PUT|DELETE /api/hosts/{id}/schedules/{sid}`. Validation: cron-expr parses via `robfig/cron/v3` (same parser the agent will use, so anything that validates here will fire there); kind ∈ {backup, forget, prune, check} (init/unlock are operator-only); backup schedules require ≥1 path; hooks rejected on non-backup kinds (spec §14.3). Mutations audit-logged. Server + store tests cover the happy path, validation, and version bumps.
-- [~] **P2-02** (L) Server-pushed schedule reconciliation. Server side complete: `pushScheduleSet*` helpers (one for the on-hello window, one for the post-CRUD async flavour), wiring in `onAgentHello` (always pushes, even when the host has no repo creds yet), `pushScheduleSetAsync` called from Create/Update/Delete handlers (no-op when the host is offline; on-hello catches up). `MsgScheduleAck` is now handled in the WS dispatcher: `OnScheduleAck` callback persists `applied_schedule_version`. Server-side end-to-end test covers (a) on-hello push of an already-created schedule + ack round-trip + applied_version write-through, and (b) connect-then-CRUD push. **Remaining:** agent-side `schedule.set` handler that applies the new state to its local cron + emits `schedule.ack` — lands in P2-03.
-- [ ] **P2-03** (M) Agent local scheduler (`robfig/cron/v3`); persists next-fire times across restarts
+- [x] **P2-02** (L) Server-pushed schedule reconciliation. `pushScheduleSet*` helpers (on-hello + async post-CRUD flavours), wiring in `onAgentHello` (always pushes, even when the host has no repo creds yet), `pushScheduleSetAsync` called from Create/Update/Delete handlers (no-op when the host is offline; on-hello catches up). `MsgScheduleAck` handled in the WS dispatcher: `OnScheduleAck` callback persists `applied_schedule_version`. Agent-side `schedule.set` handler ships in P2-03; the server side now has parity tests.
+- [x] **P2-03** (M) Agent local scheduler. New `internal/agent/scheduler` package wraps `robfig/cron/v3` — `Apply(ScheduleSetPayload, Sender)` stops the prior cron (waits for in-flight entries), rebuilds from scratch (skipping disabled entries + skipping bad cron exprs with a warn log), starts, and emits `schedule.ack`. On a tick the entry sends a new `schedule.fire` envelope to the server with `{schedule_id, scheduled_at}`. The server's `OnScheduleFire` callback (`dispatchScheduledJob`) looks up the schedule, builds args from kind, persists a `jobs` row with `actor_kind=schedule` + `scheduled_id`, and ships `command.run` back on the same conn — agent runs the job through the existing dispatcher. Tx is swapped on every Apply so reconnect is handled naturally (cron entries that fire against a dropped tx log + skip the tick). `CreateJob` now writes `scheduled_id`; this column was in the schema since 0001 but never populated. Tests: scheduler unit tests cover ack-on-apply, cron tick → fire envelope, disabled-entries silent, replace-prior-state stops the old cron. Server-side end-to-end test covers fire → command.run with the right job_id/kind/args, plus jobs row with actor_kind=schedule + scheduled_id linking back. **Deferred:** persistence of next-fire times across agent restarts (a missed fire window during downtime simply fires once on reconnect — desirable behaviour).
 - [ ] **P2-04** (M) Schedule editor UI (paths, excludes, tags, cron, retention)
 - [ ] **P2-05** (M) `forget` command with retention policy (keep-last/daily/weekly/monthly/yearly)
 - [ ] **P2-06** (M) `prune` command (admin-only, uses non-append-only credential)