refactor(api): TOML 配置 SSOT、统一错误契约、Auth/事务加固与可观测性 (#33)

配置 SSOT（TOML + .env）
统一错误契约
Auth 与事务边界
Redis / Celery 可靠性:业务 Redis（DB/0）与 Celery broker/backend（DB/1）显式拆分；连接池、sync client
可观测性（OpenTelemetry + LGTM）

api/tests/test_judge_service.py

@@ -4,6 +4,7 @@ import pytest
 
 from app.core.config import settings
 from app.core.llm_call import LLMCallError
+from app.features.evaluation.constants import eval_cfg
 from app.features.evaluation.conversation_compare_summary import (
     build_conversation_compare_summary,
 )
@@ -91,8 +92,8 @@ async def test_judge_conversation_wrapper_keeps_legacy_shape(
 def test_eval_judge_transcript_budget_exceeds_legacy_8192(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
-    monkeypatch.setattr(settings, "eval_judge_max_transcript_chars", 0)
-    monkeypatch.setattr(settings, "eval_judge_context_window_tokens", 200_000)
+    monkeypatch.setattr(eval_cfg, "judge_max_transcript_chars", 0)
+    monkeypatch.setattr(eval_cfg, "judge_context_window_tokens", 200_000)
     n = eval_judge_conversation_transcript_max_chars()
     assert n > 90_000
     each = eval_judge_compare_transcript_each_max_chars()
@@ -127,7 +128,7 @@ def test_trim_compare_transcript_pair_prefers_trimming_longer_side() -> None:
 def test_eval_judge_transcript_budget_respects_explicit_cap(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
-    monkeypatch.setattr(settings, "eval_judge_max_transcript_chars", 12_000)
+    monkeypatch.setattr(eval_cfg, "judge_max_transcript_chars", 12_000)
     assert eval_judge_conversation_transcript_max_chars() == 12_000