[majesticwizardcat]

Why

Clients read pod through a load balancer across eventually-consistent fullnodes, so a lagging node can answer after a leading one and roll client state backwards. Ticks execute in lockstep on every node, so a tick deadline is a watermark comparable across nodes. The node now stamps every account read and push with timestamp (deadline of the last executed tick, micros), serves {withdrawals, timestamp} on new /v2/bridge/withdrawals routes (v1 unchanged), and returns error code 996 for retryable nonce rejections (node PR pod/pod#191 on Forgejo, ADR 0062). This PR is the SDK side.

TypeScript SDK

• One helper keeps the highest watermark seen per entity and drops older pages and pushes: positions, balances, triggers, markets (stats page + push). Equal is accepted, later wins.
• Balances are built from the pod_positions push (it carries every balances field); the REST re-fetch is gone.
• Orders re-seed no longer overwrites rows the stream already advanced.
• since: 0 subscriptions resubscribe with bounded backoff on rejection/close; one backoff formula shared by the sync modules.
• Candle replays after a subscription error are deduped.
• waitForReceipt treats a non-JSON or non-2xx response as "no receipt yet".
• Withdrawals read /v2/bridge/withdrawals[/{account}] ({withdrawals, timestamp}), subscribe with since from the page that ended the backfill, and keep that resume point across rejections.
• Delegated submit: no nonce ownership. Attempts wait 250ms, 500ms, then 1s apart, each send is cut off after 5s, the whole submit gives up after 10s. Retried: transport errors, cut-off sends, non-JSON/non-2xx, and node error code 996 (past/future nonce, account pending). Terminal errors are thrown at once. Re-signs with a fresh nonce only when no attempt could have been delivered; otherwise waits for the receipt and ends with a delivery-unknown error.
• Every REST read sends the highest watermark the entity holds as the node's optional timestamp query parameter (minTimestampUs on the PodRestClient methods); the node waits for its state to reach it. A 503 with Retry-After throws PodBehindError, retried once, then thrown to the resource's own retry path; a read never falls back to no minimum.
• pod_positions, pod_triggers and pod_markets (re)subscribe with since set to the held watermark, read when each frame is sent (ws.subscribe accepts a params function). The node waits for it before accepting and rejects with code 995 otherwise; a 995 is resubscribed after 1s without spending the retry budget.
• The keep-highest check stays for a different reason than before: two requests for one entity are in flight together (a poll and a push, a reconnect seed and a resubscribe snapshot) and land in either order.
• mint encodes the faucet's current mint(to, tokens, uint128 deadline) ABI; the old bool form no longer parses on the node.
• Live test npm run test:live (gated by POD_LIVE_RPC_A/POD_LIVE_RPC_B): round-robin fetch across two fullnodes with one made deaf; asserts no store ever moves backwards and the delegated send succeeds. Run by scripts/e2e-node-switch-test.sh in the node repo.

Rust SDK

• Simple nonce manager: every send fetches pending.
• send_transaction retries by the same rule (backoff, 5s per-attempt cut-off, 10s budget, code 996); new PodSendError::DeliveryUnknown { tx_hash } (survives From<RpcError>); re-sign only when the caller did not fix the nonce.
• wait_past_perfect_time returns errors instead of swallowing them, and decodes the node's null result.

Public API changes

• TS: optional trailing minTimestampUs on the PodRestClient read methods; new PodBehindError; submit throws PodDeliveryUnknownError with txHash.
• TS: rest.bridgeWithdrawals returns WithdrawalsPage { withdrawals, timestampUs } (was an array); optional timestampUs on PositionsSnapshot/Balances; sendRawTransaction throws on non-2xx / missing result, accepts a signal, and PodTxRevertError carries code; MintParams/client.mint lose depositToClob (the faucet always credits the CLOB balance).
• Rust: RecommendedFillers names NonceFiller<SimpleNonceManager>; on_url bounds changed; PodSendError gains DeliveryUnknown (non-exhaustive matches need an arm; the bridge relayer gets one with the rev bump).
• Nonce-rejection retries need a node that returns code 996 (pod/pod#191).

Tests

vitest 173 tests (+1 live, skipped without nodes), cargo test -p pod-sdk 18 + 3 doctests, fmt and clippy clean. Live test passed 3/3 against a two-fullnode devnet.

🤖 Generated with Claude Code

[0zzy-o]

Looks good, just a comment picked up from the bot probably worth looking at.