test: fix the two flaky integration tests

[fylorn]

Summary

successful_login_decays_subnet_failure_counter (about 1 run in 100)

• Cause: the test client's PoW grinder gave up after 10M nonces. At difficulty 21 a solve takes 2^21 ≈ 2.1M tries on average, and the count is geometric, so 10M is exceeded with probability e^-4.77 ≈ 0.85%.
• Fix: the cap is now 32 × 2^difficulty, so the chance of hitting it is about e^-32.
• A difficulty above 26 is refused up front as a misconfiguration. The server's highest tier is 23.

drain_drops_row_after_max_attempts

• Cause: the test raced the server's own outbox drain, which ticks every 10s. When that tick claimed the due row first, the test's drain_once found nothing, and at the assertion the row was still leased mid-delivery.
• Fix: new TestApp::drain_outbox(forwarder_id). It makes the forwarder's rows due, drives one pass, then waits until each row has been attempted (delivered, dropped, or rescheduled with one more attempt), whichever drain claimed it.
• All four outbox call sites use it, including the redelivery test in webhook_signature.rs.
• The outbox tests now also reach the loopback receiver. Before, their 500s came from the SSRF guard refusing loopback, not from the receiver.

Test plan

• fmt, clippy (-D warnings)
• webhook_outbox + webhook_signature ×5, auth_pow ×3: all green
• A throwaway probe with a rival drain claiming the row the instant it was due: the row was retired and the receiver saw exactly one delivery

🤖 Generated with Claude Code