test(e2e): revive the order lifecycle tests on the beta pipeline gate

[codyborn]

The post-deploy e2e gate runs 4 of 28 tests; order.test.ts contributes zero. This PR revives the order-lifecycle coverage — the tests that would have caught order tracking silently stopping — and fixes the reliability problems that got them skipped in the first place.

How they died

Two generations, both under misleading titles:

• ff32a8f (Nov 2025) — Goerli was deprecated, so every on-chain block (checking expiry, + attempt to fill, orders endpoint sanity checks) was skipped during the mainnet migration. The in-code comment says it plainly: "GOERLI chain is deprecated. 1. change RPC_1 2. Deploy contracts 3. fund wallets."
• 7be9bb9 (Apr 2026) — the last living test was skipped inside chore: bump lambda VERSION env vars, a one-character diff with no explanation in the PR body.

Why simply un-skipping would have failed

The old expiry tests posted 5000-UNI-input orders. The test wallet holds 1.0001 UNI (verified on-chain). The status tracker quotes orders on-chain, so an underfunded order resolves to insufficient-funds — never open → expired. Any naive revival would have gone red for a reason unrelated to expiry, which is likely why nobody did it.

Reliability fixes

1. Funded-but-unfillable orders. Dust UNI input the wallet actually covers (on-chain verified: 1.0001 UNI, standing max Permit2 allowance, tx nonce 1), output of 100k-ETH scale nobody will fill. CI posts orders but never sends a transaction.
2. Poll for the outcome, don't race the scheduler. waitAndGetOrderStatus slept a fixed period and checked exactly once, racing the SFN's ~12s jittered cadence — the classic flake. pollOrderStatusUntil polls to a 90s budget past the deadline and returns early on any terminal status, so a failure's diff shows what the order actually became.
3. beforeAll no longer transacts, and is actually awaited. checkApprovals(uni, alice) was a floating promise that auto-sent mainnet approval txs — the source of the old "timeouts in beforeAll when priority fee moves" note. It's now awaited and read-only: it throws with instructions if the standing allowance regresses.

Where they run

Lifecycle tests gate beta only (RUN_LIFECYCLE_TESTS, set per stage in addIntegTests). Two reasons:

• A beta e2e failure already blocks prod promotion — that's where the CD-blocking value is.
• Prod's OrderExpiryRate alarm is a ratio (SEV2 ≥ 20%). Deploy-time dust expiries could dominate a quiet 15-minute window on chain 1 and page for nothing.

Pure-HTTP blocks (orders endpoint sanity checks, 4xx) run on both gates.

Note: the beta gate's own expiry-rate alarms will see these dust expiries — same as the pre-Nov-2025 steady state, and beta alarms are visibility, not paging. Flagging rather than hiding it.

Still skipped, with reasons now written down

• + attempt to fill — the filler wallet has 0 ETH / 0 UNI on mainnet. Reviving it means spending real gas every pipeline run or repointing at a cheap chain with a reactor and funded wallets; that's an ops decision, not a code fix. Fill coverage lives in the Anvil-based suites meanwhile.
• quote-API post — depends on trading-api → this service → GPA /hard-quote agreeing, plus COSIGNER_ADDRESS matching GPA's signer. Skipped in chore: bump lambda VERSION env vars #649 without a recorded reason; deserves its own deliberate revival.

Verification — run for real against beta

Executed locally with the pipeline's exact secrets and env (RUN_LIFECYCLE_TESTS=true):

PASS test/e2e/order.test.ts (166s)
PASS test/e2e/get-orders.test.ts (10s)
Tests: 6 skipped, 22 passed, 28 total

22 passing vs the gate's previous 4, including all three SFN-tracked expiries against real mainnet orders, with zero retries burned. The 6 skips are exactly the documented ones (fill ×4, quote-API, nonce).

Found in the process: the integ-test RPC secret was dead

all/gouda-service/integ-test/rpc pointed at ssv6cpgn9i.execute-api…, whose DNS no longer resolves — the gateway was deleted. Nobody noticed because the only tests reading it were skipped. Its last-changed date is 2026-04-25 — the same day the final live e2e test was skipped in 7be9bb9, which is almost certainly the causal story: RPC infra changed, the test broke the pipeline, and it was skipped rather than fixed.

Remediation, included here / done alongside:

• The secret now holds the entry-gateway URL the service itself uses (old value was unreachable, so this is strictly recoverable; prior version ssv6cpgn9i…, new version id d1b04cf9…).
• That gateway authenticates via x-internal-service-secret, so the buildspec now passes RPC_HEADER_SECRET (from gouda-service-rpc-urls-2, same account) into .env. RPC_HEADERS already attaches it when set — no test-code change.

Also verified: tsc clean, eslint no new errors, both RUN_LIFECYCLE_TESTS modes collect, unit suite at baseline.

Related: #698 (alarms + synth tests for the same incident), #697 (the fix itself).

🤖 Generated with Claude Code