fix(anthropic): classify OAuth tool-use overage failures

[LeonSGP43]

Summary

• add a distinct classifier reason for native Anthropic OAuth requests that fail only when tools are present with the misleading "out of extra usage" 400
• pass OAuth/tool-presence context from the Anthropic run-loop into the classifier
• surface actionable guidance and immediate fallback behavior instead of treating this as generic auth/billing noise

Issue / overlap

Fixes #15080.

This is intentionally narrower than the larger Claude CLI bridge discussed in the issue. It also does not duplicate #10576, which handles prompt-text-triggered OAuth failures, or #19260, which removes CLAUDE_CODE_OAUTH_TOKEN from Anthropic auth and would break Max-only OAuth users.

Verification

• scripts/run_tests.sh tests/agent/test_error_classifier.py
• python -m py_compile agent/error_classifier.py run_agent.py
• git diff --check

[teknium1]

Thanks for the focused OAuth-overage recovery work. Current main now provides the requested user-visible behavior, so this is redundant. This is an automated hermes-sweeper review.

• ea9e8d6e8c80fbdd0b0b4864b833d2bfd3b429bb added Anthropic’s exact "out of extra usage" HTTP-400 wording to the billing classifier (agent/error_classifier.py:118); tests/agent/test_error_classifier.py:1357 verifies fallback and non-retryable recovery flags.
• agent/conversation_loop.py:3707 attempts configured fallback before aborting non-retryable errors.
• e00800fc89e761382326c3a5eef8f4cccfbbc205 added Anthropic-specific guidance at agent/conversation_loop.py:214, including claude.ai/settings/usage, billing-cycle reset, API-key, and provider-switch options.
• The implementation shipped in v2026.7.1.