feat: builder 简历生成 + 轻度优化 + 简历导入交付副本

自内部仓库剥离深度优化与 RAG 知识库后的交付版本：
- Builder 对话式简历生成（FSM + 意图路由 LLM 兜底增强）
- 条目级轻度优化：事实覆盖门禁 + STAR/bullet 修复链，功能/简介/成果与技术栈同级保护
- 简历导入：DOCX/PDF 解析、结构归一、手机号脱敏
- PostgreSQL 运行时 + Alembic 迁移链

Co-Authored-By: Claude <noreply@anthropic.com>

backend/.env.example

@@ -0,0 +1,43 @@
+# ===== LLM — production uses Volcengine Ark over the OpenAI-compatible protocol =====
+RESUME_AGENT_LLM_PROVIDER=volcengine
+VOLCENGINE_API_KEY=
+VOLCENGINE_BASE_URL=https://ark.cn-beijing.volces.com/api/v3
+# A plain model name (as below) or an Ark inference endpoint ID (ep-xxxxxxxx) both work.
+VOLCENGINE_MODEL=deepseek-v4-flash-260425
+
+# Alternative: any OpenAI-compatible gateway. With provider=auto the LLM is used
+# only when OPENAI_API_KEY is non-empty; otherwise requests fall back to rules.
+# RESUME_AGENT_LLM_PROVIDER=openai
+# OPENAI_API_KEY=
+# OPENAI_BASE_URL=https://api.openai.com/v1
+# OPENAI_MODEL=gpt-4o-mini
+
+# The Ark gateway requires json_object; switch back to json_schema if your
+# upstream supports it.
+OPENAI_STRUCTURED_OUTPUT_MODE=json_object
+OPENAI_TIMEOUT_SECONDS=30
+OPENAI_MAX_RETRIES=2
+OPENAI_STRUCTURED_OUTPUT_RETRIES=1
+# false = LLM failures surface as errors (current production value);
+# true = silently fall back to rule-based output.
+RESUME_AGENT_LLM_FALLBACK_TO_RULES=false
+
+# Builder chat intent routing: off = rules only, shadow = LLM logs but does not
+# act, on = LLM rescue enabled (current production value).
+RESUME_AGENT_INTENT_ROUTER_MODE=on
+# RESUME_AGENT_INTENT_MODEL=  # dedicated intent-classifier model; defaults to the main model
+
+# ===== Database — PostgreSQL is required at runtime (create_app fails without it) =====
+DATABASE_URL=postgresql+psycopg://resume_agent:change-me@127.0.0.1:5435/resume_agent
+RESUME_AGENT_TEST_DATABASE_URL=postgresql+psycopg://resume_agent:change-me@127.0.0.1:5435/resume_agent_test
+# Schema the tables live in; defaults to resume_agent. Set per environment when
+# several deployments share one database.
+# RESUME_AGENT_DATABASE_SCHEMA=resume_agent
+
+# ===== Optional =====
+# RESUME_AGENT_CORS_ORIGINS=http://localhost:5173,http://127.0.0.1:5173
+# Light-optimization rate limit: 20 requests per user per 3600 s window (defaults).
+# RESUME_AGENT_LIGHT_OPT_RATE_LIMIT=20
+# RESUME_AGENT_LIGHT_OPT_RATE_WINDOW_SECONDS=3600
+
+# Keep secrets only in .env; .env is ignored by Git.