test(playwright): stabilize flaky full-flow E2E — root-cause 3 races (ClickUp 86caj0zzq)

[AquiGorka]

What

Stabilize the flaky full-flow Playwright browser E2E (playwright/tests/full-flow.spec.ts) so it's a reliable gate. The flakes moved between steps per run (observed 13, 5, 12); each is root-caused to a specific race/setup bug and replaced with a condition-based wait on the real signal. No app/product code changes — the diff is the spec only, plus a harness version bump.

ClickUp 86caj0zzq.

Race inventory + fixes

Step 13 — async settlement read race. The POS #pay-instant-btn calls /pay/instant/execute, which writes the merchant's IN/COMPLETED transaction row only after the bundle settles on chain (provider-platform executor 5s + verifier 10s ticks; the endpoint polls up to 120s). Step 12 returns in ~5s because it reads #pos-status mid-flight ("Processing payment…", non-error) — it never waits for settlement. The old Step 13 did a fixed waitForTimeout(5000) + a single-shot fetchCompletedInboundStroops read, which raced the projection and saw 0 stroops.
→ Replaced with a bounded expect.poll(fetchCompletedInboundStroops › 0) (150s timeout / 2s interval — just above the endpoint's own 120s cap). UI smoke made condition-based (expect(#balance-display,#tx-list).toBeVisible).

Step 5 — post-Step-4 navigation handoff race. Step 4's "Next" runs registerPp (network) then navigate("/setup/join"). Step 4 ended on a flaky waitForLoadState("networkidle"), which can resolve mid-navigation; under load the nav outran Step 5's 5s isVisible check, so Step 5 fell back to the fragile /home path (listPps()→badge→modal #jc-url async chain) and timed out. Trace confirmed the provider page went /setup/fund → /setup/join → /home; DB confirmed registration succeeded (pure nav-timing).
→ Step 4 now waits for the join view's #council-url to render (the real post-condition), so Step 5 deterministically starts on the join view.

Step 4 — under-funded bundle fee-payer (deterministic, was masked). The PP root account is the fee-payer for every bundle (executor.process.ts: feePayerPubkey = ppPublicKey). provider-platform's pre-flight requires the fee-payer to hold ≥ NETWORK_FEE (1_000_000_000 stroops = 100 XLM, canonical across all docker suites) + resource fee after reserves. Step 4 funded it with only 10 XLM, so every bundle terminal-FAILed with Insufficient fees on fee-payer account and the payment never settled — previously hidden by Step 13's old loose UI-text assertion. The green e2e/pos-instant suites fund the PP via friendbot (~10k XLM).
→ Step 4 funds the PP with 500 XLM (well above the 100 XLM floor). NETWORK_FEE left canonical; no product/config change.

No masking

No test.retries, test.skip/quarantine, .only, or blanket timeout: inflation used as the fix — grep-clean. Each fix waits on an actual condition/signal.

Stability evidence

./test.sh playwright run 10/10 consecutive green, 0 flakes (each ~6.3–6.9m, all 16/16 including Steps 13b + 14 which were previously blocked by the Step 13 failure).

Notes

• Diff: playwright/tests/full-flow.spec.ts only + playwright/package.json version bump (0.1.0 → 0.1.1). No app-code files touched.
• deno fmt/deno lint don't apply to playwright/ (Node/Playwright TS project, extensionless imports, excluded from the fmt-lint CI matrix); the suite is compiled by Playwright's own tooling.