fix(sync): bootstrap a push channel for an imported calendar that lacks one

[tyler-dane]

The gap

Only 4 of 1786 sync resources in production hold a push subscription. Everything else depends entirely on the reconcile sweep, which is what turned a three-document schema bug into a 23-hour fleet-wide outage (#2527).

The cause is a hole in the channel lifecycle:

• subscriptionFollowup is enqueued from exactly one place — the initialImport dispatch branch.
• The renewal sweep's finder, listExpiringSubscriptions, filters on subscriptionId: { $ne: null }. It renews channels that exist; it cannot create one.

So a calendar that reaches an imported state by any route other than the initialImport job can never become watchable, and nothing detects or repairs it.

Production got there by preseeding: the Sync cutover wrote events and cursors straight into the store, bypassing the import job. The data confirms it — 938 resources created 2026-07-26 have a cursor and no subscription, while the only 4 with channels are from the same day via the live connect path.

The practical cost: with ~1000 active resources swept 100 at a time every ~10 minutes, any given calendar is refreshed roughly every 100 minutes. That is the floor on freshness for nearly every user, and it is why the sweep is a single point of failure.

Change

Enqueue the bootstrap followup after an applied incremental pull when the resource has no channel.

Pulls already run for every stale calendar, so this needs no new sweep, no migration, and no operator step — the existing fleet heals as each calendar comes around. The followup uses the same coalescing key as the renewal sweep, so a bootstrap and a renewal can never open two channels.

Known wart, stated plainly

A calendar the provider refuses to watch returns unsupported, and that verdict is not recorded anywhere. Such a calendar re-attempts one watch per pull (~1 per 100 min at current cadence). Cheap and bounded, but it is a real cost rather than something that self-heals. Recording the verdict would need a new schema field, which is deliberately out of scope here given #2527.

Tests

• New: an applied pull on a channel-less resource emits a subscriptionMaintain followup.
• Updated: the existing "no followup" test now seeds a live channel, which is the case it was actually describing.
• Updated: the worker-drain test expects 4 rather than 2, and now demonstrates that drain also processes work spawned partway through it.

750/750 sync tests pass; type-check and lint clean.

🤖 Generated with Claude Code