fix(tests): use URL-based mock routing to fix flaky tests

[la14-1]

Why

Two tests (digitalocean-token.test.ts and hetzner-cov.test.ts) failed intermittently in the full suite but passed in isolation. Both relied on sequential callCount to route mock fetch responses, but Bun runs test files in the same process — concurrent tests sharing global.fetch caused the count to drift, producing wrong responses and assertion failures.

What changed

• digitalocean-token.test.ts: Replaced count-based routing with URL-based routing (api.digitalocean.com vs cloud.digitalocean.com), returning harmless {} for unrelated concurrent fetches.
• hetzner-cov.test.ts: Replaced count-based routing with URL+method routing (POST /servers vs GET /servers, GET /ssh_keys, GET /primary_ips, DELETE /primary_ips/:id), using a serverPostCount to track only the relevant POST retries.

Test results

Full suite: 2138 pass, 0 fail (verified across multiple runs).

[la14-1]

Duplicate of #3378. Both PRs rewrite the same two test files (digitalocean-token.test.ts and hetzner-cov.test.ts) from sequential callCount mocking to URL-based mock routing. They modify the exact same lines and will conflict on merge.

#3378 was opened ~2 hours earlier and has a more detailed root-cause analysis in its description. This PR (#3379) adds HTTP method routing in the hetzner test which is slightly more precise, but the difference is minor — both approaches solve the problem.

Recommendation: This PR supersedes/duplicates #3378. Only one should be merged. Both are also complementary to #3376 which fixes the root cause (telemetry singleton leaking _enabled across parallel test files).

-- refactor/pr-maintainer

[la14-1]

Status check (day 3): All CI checks passing (lint, unit tests, mock tests, shellcheck, macOS compat). The earlier duplicate #3378 has been closed, so this is now the sole PR for the flaky callCount fix. Ready for review.

-- refactor/pr-maintainer