feat(indexer): N-block reorg detection + rewind

[satyakwok]

Summary

Tier 3 of the indexer audit — first arch-level fix lands. Builds on top of #45 (Tier 1) + #46 (Tier 2).

Pre-Tier-3 `sync.ts` only checked the immediate parentHash of the new block — that catches a 1-block reorg but silently loses data on any reorg that lands a different block at a height already indexed. BFT chains rarely reorg deep but the indexer should be correct against the worst case (validator re-org during binary swap or chain.db rsync recovery).

What lands

New module `apps/indexer/src/reorg.ts` exports `checkAndRewindReorg`:

1. Read `last_synced_height` from `_meta`.
2. `SELECT` block hashes from `blocks` for `[synced - DEPTH, synced]`.
3. Refetch the canonical hash for each height via chain RPC. With viem batch transport (Tier 1) these collapse to one HTTP request.
4. Walk forward, find first divergence.
5. On divergence: `DELETE FROM blocks` (FK cascade clears `transactions` + `logs`) + `DELETE FROM token_transfers` + rewind `last_synced_height` to `divergedAt - 1`, all in one SQL transaction. Bumps the `reorg_count` observability counter in `_meta`.

Wired into `apps/indexer/src/index.ts` via `setInterval`:

Env	Default	Purpose
`INDEXER_REORG_INTERVAL_MS`	60000	Cadence between checks
`INDEXER_REORG_CHECK_DEPTH`	16	How many blocks back to verify

Cleared on graceful shutdown alongside the `stats_daily_mv` refresh timer.

Deferred Tier 3 items (separate follow-up PRs)

The remaining Tier 3 work is each a multi-day refactor and lands in its own PR:

• Declarative event handlers (feat(indexer): is_contract auto-detection + faster testnet backfill #9) — Ponder-style `indexer.on(event)` registry to replace the procedural `sync.ts` ERC-20/721/1155 dispatch hardcode.
• GraphQL surface (Feat/native tx shape adapter #8) — Mercurius plugin with auto-generated schema from Drizzle types, exposing the same data model as REST.
• Table partitioning (feat(api): /contracts/recent endpoint #10) — `pg_partman` by `block_height` range for `transactions` / `logs` / `token_transfers` once row counts cross 50M+. Requires plan + data-migration window; runbook lands separately.

Test plan

• `pnpm turbo build` passes
• On a staging DB, manually mutate one `blocks.hash` to a wrong value and verify the next check rewinds correctly + bumps `reorg_count`
• Verify `token_transfers` rows are cleaned up alongside cascaded `transactions` / `logs`
• On a quiescent indexer (no new tips), `reorg_count` stays 0 across multiple intervals