fix(backend): distributed-lock fencing, metrics on RPC failure, webhook retry consolidation, yield history tests

[dikachim060-wq]

Note: this contribution was authored with AI assistance (Claude Code).

Summary

This batch PR addresses four backend issues assigned in Wave 6: atomic distributed-lock release, suppressing false-healthy metrics on RPC failure, wiring the correct webhook retry implementation, and adding missing test coverage for yield history share/cost-basis logic.

closes #1163
closes #1201
closes #1167
closes #1171

Changes

• [Backend] defaultChecker releaseLock() deletes the Redis lock unconditionally, letting a long run delete another instance's lock #1163 — Atomic lock release in defaultChecker: acquireLock now stores the unique fencing token it set in Redis. releaseLock calls a new cacheService.deleteIfEqual method that runs an atomic Lua compare-and-delete (GET + DEL in one round-trip). A run that outlives the 10-minute LOCK_TTL can no longer delete a lock acquired by a subsequent instance, preventing the overlapping check_defaults submissions the lock was designed to prevent.

• [Backend] RPC failure to fetch chain tip records indexer_lag_ledgers=0, masking outages and defeating the lag alert #1201 — Skip lag/chain-tip metrics on RPC tip-fetch failure: getLatestLedgerSequence now returns null instead of 0 on error, and increments a new indexer_rpc_fetch_errors_total Prometheus counter. pollOnce skips recordIndexerLedgers when the tip is null, so an RPC outage no longer resets indexer_lag_ledgers to 0 and silently defeats the behind-chain-tip alert.

• [Backend] startWebhookRetryProcessor is dead and the wired scheduler uses a different backoff than webhookService stores #1167 — Wire webhookRetryProcessor as the single live retry path: index.ts now starts startWebhookRetryProcessor (which drives WebhookService.processRetries) instead of startWebhookRetryScheduler. This makes the retry schedule single-sourced: processRetries respects the next_retry_at timestamps stored by sendToWebhook (RETRY_DELAYS_MS: 5m/15m/45m, MAX_RETRY_ATTEMPTS=4). The old scheduler's independent BACKOFF=[60,300,1800]s that ignored stored next_retry_at is no longer the live path.

• [Testing] yieldHistoryService withdraw cost-basis and XDR share-decoding paths are uncovered #1171 — Test coverage for Withdraw/EmergencyWithdraw cost-basis and XDR share-decoding: Three new tests in yieldHistoryService.test.ts:

  1. Withdraw with value=null — asserts cost-basis is reduced proportionally via shareRatio
  2. Deposit + Withdraw with a base64 XDR ScVec value — asserts parseDepositWithdrawAmounts decodes shares as BigInt and converts to Number correctly
  3. EmergencyWithdraw encoded in XDR — follows the same cost-basis reduction path as Withdraw

Test plan

• Existing webhookRetryProcessor.test.ts covers WebhookService.processRetries (the now-live path)
• New tests in yieldHistoryService.test.ts cover the three acceptance criteria for [Testing] yieldHistoryService withdraw cost-basis and XDR share-decoding paths are uncovered #1171
• defaultChecker.test.ts covers lock acquisition; the new deleteIfEqual is a thin Lua wrapper tested via integration

---

Not built/tested locally (dependencies not installed in the Wave workspace); relying on CI for compilation and full test-suite validation.