feat(store): watermark persistence + enabled-repo enumeration (RIG-2883 T4) - #705

…83 T4)

The board webhook lane's durable state, additive on the pre-live `0001_init.sql`.

- `forge_repo_subscriptions` gains `swept_updated_at` (the per-repo updated-at watermark; NULL = never swept) and `list_etag` (the conditional-GET etag for the repo LIST walk), plus `LoadForgeRepoWatermark`/`StoreForgeRepoWatermark` (advance-after-sink; unknown coordinate reads zero/empty, store on a missing row is `ErrNotFound`), `ListEnabledForgeRepos` (ascending cross-coordinate enumeration for the reconciler), and `IsEnabledForgeRepo` (the point membership check the webhook arm gates on).
- `issues` gains `forge_updated_at TIMESTAMPTZ`, INERT this slice — the OQ-6(a) recency-guard column. T4a threads its write path and makes the upsert conditional; the bare column is a no-op until then.

The `forge_list_cursors` table and its store methods (`ForgeListCursor`/`UpsertForgeListCursorPage`/`PruneForgeListCursorPages`) stay: they retire atomically with their `serve.go` poll-driver consumer in T5, so this slice never breaks the tree build. pgtest covers the watermark round-trip, coordinate isolation, enabled-repo enumeration + point membership, and that the edited migration applies on a fresh database with the three new columns present.

Spec-impact: none. Refs RIG-2883
Co-authored-by: Matt Wilkinson <matt@rigel.build>

…cks (RIG-2883 T4)

ListEnabledForgeRepos / IsEnabledForgeRepo key on repo alone, matching
the frozen repo-keyed ingest seam, while the watermark methods are
coordinate-keyed (provider, host, repo). Unambiguous in a github.com-only
deployment; documents the latent cross-coordinate ambiguity so a future
multi-host enablement threads (provider, host) through the seam first.

Refs RIG-2883
Co-authored-by: Matt Wilkinson <matt@rigel.build>