fix(backends): make streamed replies and session state consistent

- Validate stage prefixes before forwarding model tokens. Invalid prefixes
  may use the existing repair budget; once output starts, validation failure
  aborts the turn without appending a regenerated reply. Bound the token
  queue and close inner streams when consumption stops.
- Replace delete-time lock eviction with reference-counted session locks
  covering holders and waiters. Serialize set/get with chat/stream/delete
  and release cancelled waiters without replacing a live lock.
- Store person synonyms under workflow-native driver keys and expose
  consistent alias views to single/group reads and graph prompt injection.
- Separate committed field patches from public formUpdate visibility.
  Commit callbacks keep repository fields synchronized even when updates
  are hidden or the consumer closes after graph completion.
- Disable OpenAI SDK automatic retries. Own bounded retry attempts in the
  application, use 1/2/4 second backoff, retry streams only before text is
  emitted, enforce attempt timeouts, and close upstream/client resources.
- Translate FastGPT SDK max_retries (total attempts) from the configured
  extra-retry count, and verify the real SDK loop uses four attempts.
- Document the per-LLM budget and streaming failure semantics without
  claiming the whole multi-call HTTP turn has the same timeout.

Validation: 267 offline tests passed, including actual SDK requests through
MockTransport, stream repair/abort, timeout/close, lock races, aliases and
formUpdate visibility; existing four dependency deprecation warnings.

.env.example

@@ -17,8 +17,9 @@ VOICE_CONFIG=config/voice.json
 # 每组 LLM_<名>_BASE_URL / _API_KEY / _MODEL 必填；
 # 可选 _TIMEOUT_SECONDS / _MAX_RETRIES / _TEMPERATURE / _EXTRA_BODY（JSON 字符串，
 # 如思考开关 KIMI 网关为 {"reasoning_effort":"none"}）。
-# 超时/重试缺省 60 秒 / 3 次，与 FastGPT 客户端 SLA 对齐
-# （契约测试 test/backends/test_sla_alignment.py）。
+# 超时/重试缺省：每次尝试 60 秒，最多额外重试 3 次，退避 1、2、4 秒。
+# SDK 自动重试关闭；流式发出首个文本 token 后不重试。
+# 单次 LLM 调用预算最多 247 秒（另加响应关闭清理）；不是整轮聊天总超时。
 # AGENT_HISTORY_MESSAGES=50：注入模型的对话历史上限（最近 N 条，
 # 双向计数）；调大支持长对话；0 = 不截断（对齐 FastGPT 全量历史）。
 # AGENT_HISTORY_MESSAGES=0