Require Claude Code 2.1.248 and remove gateway auth workaround

[andy-xu-db]

Summary

• require Claude Code 2.1.248+, where gateway model discovery works with apiKeyHelper as the only credential
• stop setting ANTHROPIC_AUTH_TOKEN and stop creating a refresh cache for native gateway model discovery
• keep the model-alias proxy and preserve the existing relayed-subscription token refresh path

Testing

• uv run pytest tests/test_agent_claude.py tests/test_anthropic_model_discovery_proxy.py tests/test_gateway_proxy.py -q (167 passed)
• uv run pytest tests/test_lint.py -q (3 passed)
• uv run ruff check .
• GitHub CI: unit/lint and workspace e2e jobs passed

Live end-to-end verification

Verified on eng-ml-inference-staging with Claude Code 2.1.251:

• ucode claude --enable-model-discovery --skip-preflight launched through the PR model-alias proxy
• the Claude child used a loopback ANTHROPIC_BASE_URL and had no ANTHROPIC_AUTH_TOKEN
• /model displayed gateway models (system.ai.claude-sonnet-4-6[1m], system.ai.claude-haiku-4-5, plus 13 more), proving /v1/models authenticated via apiKeyHelper
• a non-interactive inference returned the exact marker UCODE_PR_413_OK with exit code 0

Local full-suite note: the locally installed Claude Code 2.1.251 returned a canned response in TestClaudeUserAgent::test_user_agent_arrives_at_gateway without contacting its localhost capture server; the same test passed in GitHub CI.

[lilly-luo]

the model discovery proxy is now completely separate from the relayed auth proxy now right? model discovery is responsible for patching the model name where as relayed auth swaps headers. in that case, i think it makes sense to fully decouple them.

[andy-xu-db]

the model discovery proxy is now completely separate from the relayed auth proxy now right? model discovery is responsible for patching the model name where as relayed auth swaps headers. in that case, i think it makes sense to fully decouple them.

makes sense, i updated it so model discovery and relayed auth use separate proxies now