perf(indexer): batch sync.ts + stats_daily_mv + RPC fallback transport

[satyakwok]

Summary

Tier 2 of the indexer audit pass. Three changes that reshape the hot loop. Builds on top of #45 (Tier 1).

1. Batch `indexBlock`

`apps/indexer/src/sync.ts` — pre-batch sync was N HTTP round-trips + N INSERTs per tx, sequential, inside one long-held DB transaction. On a 5000-tx block that meant ~10 s of held write lock and chain RPC saturation from one indexer instance alone.

Refactor to four phases:

• PHASE 1: `getNativeTransaction()` fan-out via `mapWithConcurrency` with a 25-task cap (env `INDEXER_TX_FETCH_CONCURRENCY`). HTTP happens outside the DB transaction.
• PHASE 2: build `TxRow[]` / `AddrRow[]` arrays in memory; dedupe addresses by primary key (Postgres rejects multi-row `ON CONFLICT` on the same target with `cardinality_violation`).
• PHASE 3: `getLogsRange` + decode → `LogRow[]` / `TransferRow[]`.
• PHASE 4: `db.transaction` wraps four batch `INSERT`s (txs, addresses, logs, transfers). Lock window now ms not seconds.

Expected: ~50–100x throughput on log-heavy blocks; backfill burst no longer monopolises the chain LB.

2. `stats_daily_mv` materialised view

Replaces the per-process 5-min in-memory cache that was lost on every restart and not shared across api processes.

• Migration `0005_stats_daily_mv.sql` creates the view with a UNIQUE INDEX on date so `REFRESH MATERIALIZED VIEW CONCURRENTLY` works.
• Indexer worker (`apps/indexer/src/index.ts`) fires REFRESH every 5 min (env `INDEXER_STATS_REFRESH_INTERVAL_MS`); initial seed is non-CONCURRENT since the unique index isn't valid until the first populate.
• API (`apps/api/src/routes/native.ts`) drops the in-memory cache and just `SELECT`s from the view (~700 rows for 2 yrs of chain history).

Combined with Tier 1's 60 s edge cache: worst-case freshness gap is ~6 min.

3. RPC HTTP fallback transport

`packages/chain/src/index.ts` — operator can now set `INDEXER_RPC_HTTP_URLS=primary,backup1,backup2` and viem's `fallback()` transport rolls over on connection error / 5xx / timeout. Health-checks each transport every 60 s and demotes bad ones automatically.

Legacy single-URL env (`INDEXER_RPC_HTTP_URL`) continues to work — it just becomes the only entry in the URL list. Single-URL deployments fall back to plain `http()` (no fallback wrapper overhead) so behaviour is byte-identical.

Out of scope (Tier 3)

Reorg buffer, declarative event handlers, GraphQL, partitioning.

Test plan

• `pnpm turbo build` passes
• `pnpm db:migrate` applies 0005 cleanly on staging DB
• `SELECT * FROM stats_daily_mv LIMIT 5` returns rows after worker first refresh
• Index a single block locally; verify all txs land via batch INSERT (one statement per table per block in PG `pg_stat_statements`)
• Set `INDEXER_RPC_HTTP_URLS=https://bad-url,https://rpc.sentrixchain.com\` and verify worker falls back successfully (chain logs report rank-demote of the bad endpoint)