Adds the schema and queries needed to auto-archive inactive chats
from dbpurge, following the plan in /tmp/plans/chat-auto-archive.md
(not wired up yet — that's Phase 3).

* New migration 000471 creates the per-owner digest dedupe log.
* New chats.sql query AutoArchiveInactiveChats drives the whole
  show from a LATERAL MAX(chat_messages.created_at) over each
  family. All roles count; soft-deleted messages do not.
* New siteconfig.sql queries for the 'agents_chat_auto_archive_days'
  knob, default 90, 0 disables. Mirrors the existing retention knob.
* New chat_auto_archive_digest_log.sql for batch read + upsert of
  the per-owner dedupe rows.

The actual dbpurge wiring, HTTP endpoints, notification template,
and dispatch logic land in subsequent commits.

…igest_log

TestMigrateUpWithFixtures asserts that every table has at least one
row after running through all migrations. Migration 000473 creates
chat_auto_archive_digest_log without seeding any data, so the test
flagged it as empty. Seed a single dedupe row against the first
available user so the table is covered.

…ions into one

Collapse 000473 (digest log table), 000474 (digest notification
template), and 000475 (partial index on chats) into a single
000473_chat_auto_archive.{up,down}.sql. Reviewers only have to
reason about one file; the three pieces are tightly coupled to the
same feature anyway.

notification_messages already deduplicates on (template_id, user_id,
method, payload, day) via an MD5 hash trigger (migration 000245),
which is exactly the guarantee our custom chat_auto_archive_digest_log
table was providing - and users who find the digest noisy can disable
the template entirely from their notification preferences.

The custom 24h sliding window was only measurably better than native
dedupe during a large initial backfill where the same owner gets hit
across many ticks with payloads that happen to differ. That scenario
needs a deployment with thousands of stale chats to matter; under
early-access adoption (feature off by default, customers hop on
deliberately), it does not.

- Drop migration of chat_auto_archive_digest_log from 000473.
- Drop GetChatAutoArchiveDigestLogsForOwners and
  UpsertChatAutoArchiveDigestLog queries, their dbauthz wrappers,
  and the dbauthz test cases.
- Simplify dispatchChatAutoArchive to just audit + enqueue per owner;
  drop the unused db parameter.
- Drop chatAutoArchiveDigestDedupe constant and the skipped_deduped /
  skipped_dedupe_error Prometheus labels.
- Drop the migration fixture (table no longer exists).
- Drop the DedupesDigestWithin24h dbpurge subtest and the dedupe-log
  assertion inside ArchivesInactiveRoot.
- Update the auto-archive docs to describe native per-day dedupe and
  the per-template opt-out.

Net: -288 LoC.

- Drop coderd_chat_auto_archive_digests_sent_total. The native
  coderd_notifications_inflight_dispatches and
  coderd_notifications_retry_count are both labeled with
  notification_template_id, so filtering by the chat auto-archive
  template ID gives the same data without an extra per-feature
  metric. Only coderd_chat_auto_archive_records_archived_total
  stays because nothing else counts chats archived per tick.
- Fix the duplicate greeting in the digest template. The SMTP and
  webhook wrappers already prepend 'Hi {{.UserName}},', so the
  body_template must not repeat it.
- Drop the 'View archived chats' action button. The
  /chats?archived=true route isn't supported in the frontend yet
  and adding that view is out of scope for this PR.

…count

Deep review round 2 follow-up.

- Drop stale dedupe-log references from constant, function, and
  dispatch-block comments. The custom dedupe table was removed in
  5831e0d; comments now say what the code actually does.
- Remove unused int32 parameters from buildDigests. They were
  retained as blank identifiers during simplification but never
  used; moved to buildDigestData which is the real consumer.
- Compress verbose comments on constants, helpers, and types
  without losing meaningful content. dbpurge.go: 566 -> 544 LoC.
- Render additional_archived_count in the digest body template.
  buildDigestData already computes the overflow count, but the
  template never consumed it, so a user archiving 100 chats at
  once would silently see only the first 25 titles with no hint
  more existed. Now the body ends with "...and N more." when
  applicable.

Deep review round 3.

- Rename chatsAutoArchiveBatchSize to chatAutoArchiveBatchSize so
  all four chatAutoArchive* constants share a consistent naming
  convention (flagged by go-architect as a minor inconsistency).
- Add TestAutoArchiveInactiveChats/DigestOverflowCap. The overflow
  branch of buildDigestData (cap titles at 25, surface extras via
  additional_archived_count) was untested; this archives 27 roots
  and asserts the digest payload caps at 25 titles with a "2"
  overflow count.

… auth

Address review feedback:

- Drop swagger annotations from the auto-archive-days handlers; document
  the endpoints in docs/ai-coder/agents/chats-api.md instead. Swagger will
  pick this up when the API stabilizes.
- Grant UPDATE+READ on rbac.ResourceChat to subjectDBPurge and drop
  AsSystemRestricted from the auto-archive call; the ticker's AsDBPurge
  ctx now has everything it needs.
- Replace the custom labelsFromRaw helper with StringMap.Scan, which
  already handles []byte and nil.
- Compress the SQL comment block on AutoArchiveInactiveChats into a
  three-line header plus short inline annotations on the non-obvious
  predicates.
- Trim verbose Go docstrings across dbpurge.go: New, the dispatch
  block, chatFromAutoArchiveRow, dispatchChatAutoArchive,
  buildDigests, buildDigestData, and the constant block.

…ctiveChats

The comment referenced subjectSystemRestricted, but dbpurge now runs
with subjectDBPurge (which was granted ResourceChat read+update in
the same round).

…s retention clock

DeleteOldChats and DeleteOldChatFiles both key off chats.updated_at,
which AutoArchiveInactiveChats bumps to NOW() on archive (mirroring
manual archive). Users/admins deserve to know the clock restarts.

Follow-up tracked in CODAGT-222: add an archived_at column so the
purge queries can stop using updated_at as a proxy.

Adds three subtests to TestAutoArchiveInactiveChats:

- MultipleOwners: two owners in one org each get their own digest
  with only their chats; audit count matches archived roots.
- SecondTickIdempotent: after tick 1 archives two roots, a third
  root seeded between ticks archives on tick 2 without re-archiving
  or re-notifying the first two.
- BatchSizePagination: with batch size 20 and 27 stale roots, tick 1
  archives 20, tick 2 archives the remaining 7, tick 3 is a no-op.

To drive multi-tick tests deterministically, introduce a tickDriver
helper that installs quartz traps *before* dbpurge.New so the forced
initial tick and each loop iteration can be observed without racing
with asynchronous ticker.Reset calls. tickDriver is separate from
awaitDoTick because awaitDoTick conflates the forced tick and the
first loop iteration, which is fine for idempotent single-tick tests
but hides per-tick state when batch size is limiting.

chatAutoArchiveBatchSize becomes a var (from const) so
SetChatAutoArchiveBatchSizeForTest in export_test.go can shrink it
for pagination tests without inserting thousands of rows.

… landed 000474_drop_chat_model_config_provider_fk