[Bug] Anthropic API: 50% timeout errors causing token waste

Bug Description
Roughly 50% of API requests timeout regardless of context size. Even tiny classifier requests (used internally to check if a Bash command is safe) are timing out. Each timeout burns quota because Claude Code retries with the full context (60k+ cached tokens). Over a weekend with documented Anthropic outages (Jun 22, 23, 27), this consumed 34% of my weekly limit with little productive output.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.195
• Feedback ID: 3584f641-3868-40c3-a29b-37260d888804

Errors

Evidence of server-side issue (not network/context):

• Traceroute shows clean 25ms latency to api.anthropic.com
• Anthropic's own telemetry pipeline is timing out from my client:

[ERROR] 1P event logging: 58 events failed to export (code=ECONNABORTED)
[ERROR] Failed to flush logs to Datadog: AxiosError: timeout of 5000ms exceeded

Timeout examples on main requests:

18:28:13 [WARN] Slow first byte: no stream chunk 30s after request (attempt 1)
18:28:43 [WARN] [TTFB] no response headers after 60000ms — aborting
18:28:43 [ERROR] API error (attempt 1/11): Request timed out.
18:28:46 [DEBUG] Stream started — first byte after 2091ms  ← next attempt succeeds instantly

Even the permission classifier (tiny request) times out:

18:43:53 [WARN] classifier_request_progress ageMs=45000
18:44:08 [INFO] classifier_request_finished outcome=aborted durationMs=60004
18:44:08 [WARN] Auto mode classifier unavailable, denying with retry guidance (fail closed)

This is a ~1k token request to check if npx vitest run is safe to execute. It timed out after 60 seconds.