Stop treating any error containing "unknown" as an XDR failure #183

Background

fetchEventsSafe decides whether to skip a ledger by substring-matching the error message: if (msg.includes("XDR") || msg.includes("unknown")) at src/rpc.ts:209, and again at :222.

"unknown" matches far too much — getaddrinfo ENOTFOUND … unknown host, a bare unknown error, provider 5xx bodies. When it matches, the ledger is silently skipped and the cursor advances past it (src/rpc.ts:211). A transient network blip is therefore indistinguishable from a genuinely unindexable ledger, and the data is gone with nothing but a console.warn.

Acceptance criteria

• Classification moves to a named predicate covering the real XDR cases and not bare "unknown"
• A transient Error("unknown error") propagates and is retried instead of skipping the ledger — fails before, passes after
• A genuine XDR error still bisects and skips; existing fetchEventsSafe tests stay green
• Every skipped ledger increments a metric, so skips are visible rather than only logged

Drips Wave · Complexity: Intermediate · 150 points

How this repo is laid out

Wraith is a Soroban incoming-transfer indexer: it fills the gap Horizon leaves for SAC/SEP-41 token events by recipient address. Roughly 14k lines under src/, REST + GraphQL + WebSocket, Postgres via Prisma, deployed on Render.

• Unit tests are jest (src/**/__tests__, tests/*.test.ts), with a coverage gate.
• Integration tests are vitest (tests/integration/, vitest.integration.config.ts, Docker).
• Don't mix the two runners in one file.

Known state of main before you start

tsc --noEmit and the unit suite are green. The Integration tests CI job is red, and so is Load test — both predate this batch. Open PRs #177 and #178 address them. If the integration job is red when you open a PR, that is not your fault; say so in the PR and carry on.

Four modules are shipped, closed as done, and imported by nothing: src/indexer/checkpoint.ts, src/ingest/reorg.ts, src/middleware/opa.ts, and src/api/candles.ts + src/workers/ohlc-refresh.ts. So the live indexer today has no reorg handling and no atomic checkpointing, despite both features being closed. Two issues below start unpicking that — don't assume a closed issue means a wired feature.

Ground rules

• Every PR needs a test that fails before the change and passes after, unless the issue says otherwise.
• Don't widen a Prometheus label to something unbounded — addresses and contract ids must never become label values.
• Nothing may log or return a raw database error, a connection string, or a provider URL.
• Don't reformat a file you're changing. A diff that is 90% whitespace hides the 10% that matters.

Required: Before submitting, join the contributor Telegram so your work can be tracked and counted toward the Stellar Wave: https://t.me/+fxHXq8f1SwlkZDBk