fix(deepseek): explicit RuntimeError on missing env (Squad G / E2E Blocker Elimination)

[Ghenghis]

Summary

Discovery audit Agent #3 (2026-05-09) flagged this during the E2E Blocker Elimination Program audit pass.

build_probe_request and completion_caller previously used os.environ[config.api_key_env] which raises a bare KeyError when the env var is missing. The bare KeyError surfaced as an opaque HTTP 500 with the env variable NAME in the traceback (low-risk info disclosure).

Fix

Both functions now use os.environ.get(...) and explicitly raise RuntimeError("DeepSeek provider is not configured: ...") with an actionable operator-facing message that does NOT name the env var.

Tests (6/6 pass)

• test_build_probe_request_missing_env_raises updated to expect RuntimeError with match string
• New test_completion_caller_missing_env_raises pins the same guard for the completion path

Test plan

• py_compile: OK
• Focused tests: 6/6 pass
• Pre-push hook: passed
• CI on PR

Scope

• DeepSeek only. MiniMax already has explicit configured-flag handling (Agent 14 verified). No other providers touched.

Discovery provenance

Discovery audit Agent #3 of the E2E Blocker Elimination Program. Tracks as one of 2 security risks (the other = broad-except cleanup at 6 sites; separate PR follow-up).

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request replaces implicit KeyError exceptions with explicit RuntimeError messages when the DeepSeek API key is missing from the environment, preventing potential information disclosure of environment variable names in tracebacks. These changes were applied to both the probe request construction and the completion caller. Feedback was provided to move the configuration check in completion_caller from call-time to initialization-time to implement a 'fail-fast' pattern, along with corresponding test updates.

[gemini-code-assist]

This test should be updated to reflect the move of the configuration check from call-time to initialization-time in completion_caller. This ensures that the 'fail-fast' behavior is correctly verified.

Suggested change

	caller = deepseek.completion_caller(_config())
	request = LLMRequest(prompt="hello", max_completion_tokens=10, token_id="test-token-id")
	with pytest.raises(RuntimeError, match="DeepSeek provider is not configured"):
	caller(request)
	with pytest.raises(RuntimeError, match="DeepSeek provider is not configured"):
	deepseek.completion_caller(_config())