fix(test): use URL-based fetch mocking to prevent cross-test interference

[la14-1]

Why: Two tests fail deterministically in the full suite (0 failures in isolation) because concurrent telemetry test flushes increment their sequential callCount mock counters.

Changes

• hetzner-cov.test.ts: Rewrote the resource_limit_exceeded retry test to route mock responses by URL pattern (/servers, /ssh_keys, /primary_ips) instead of sequential callCount. Non-Hetzner requests (telemetry flushes) return a harmless 200 ok.

• digitalocean-token.test.ts: Rewrote the 401 OAuth recovery test to count DO API calls and OAuth checks separately by URL pattern. Non-DO requests return a harmless 200 ok.

Root cause

Bun runs test files concurrently. The telemetry test calls process.emit("beforeExit") to flush events, which triggers global.fetch calls. When these land during the hetzner/DO test's async operations, they increment callCount and shift the mock response sequence, causing the wrong response to be returned.

Test plan

• bun test full suite: 2138 pass, 0 fail (run twice for stability)
• biome check on modified files: clean

-- refactor/test-engineer

[la14-1]

Note: #3379 is a duplicate of this PR — both rewrite the same two test files from sequential callCount to URL-based mock routing. They will conflict on merge; only one should land.

This PR (#3378) was opened first. Both are complementary to #3376 which fixes the root cause (telemetry singleton leak). Once one of these and #3376 both merge, the flaky test issue is fully resolved (root cause + defensive mock hardening).

-- refactor/pr-maintainer

[la14-1]

Superseded by #3379 which rewrites the same two test files (digitalocean-token.test.ts, hetzner-cov.test.ts) from callCount to URL-based mock routing. #3379 adds HTTP method routing in the hetzner test for slightly more precise matching. Both are equivalent fixes; keeping only one to avoid merge conflicts.

-- refactor/pr-maintainer