feat(all): NaN 伪收敛否决与 nl_tight 能量回退、emflux 检验 4π 修正、nst 行宽与种子边界修复、gfATO 谱线表接通与网格加密 9216 点、阶段分项统计与跳板机部署

物理修复(400 失败点归因,见 docs/failed400_nan_pseudo_convergence_2026_08_17.md):
- runner: 阶段 converged 后复查 fort.7,含 NaN/Inf 即否决(fort.9 全零伪收敛,
  曾致 261 点误跳过 nl_direct 回退);否决阶段不产出种子,阻断污染传播
- runner: nl_tight 回退——仅能量边际失败时以 CHMAX 收紧 10× 从自身模型续迭代,
  残差降幅 ~10×;execute_tlusty_stage 抽取供主链与回退共用
- conv_check: fort.14 为 Eddington 通量 Hλ,积分需乘 4π 再比 σTeff⁴
  (旧版 ratio 稳定 0.0796=1/4π,全点系统性假阳性)+ 回归测试
- nst_writer: 单行超 80 字符被 TLUSTY 静默截断,IFALI/JALI/TRAD 等从未生效;
  按 75 字符自动换行
- seed_finder: Teff 容忍度改含边界 <=,相邻 5000K 档恢复互为种子 + 回归测试

谱线表与网格:
- 默认线表 gfVIS99 → gfATO(全波段 18-23000Å),TaskSpec.linelist 支持工作流
  级覆盖,节点按需下载(进程互斥锁防并发重复下载 238MB)
- sdB_cno Teff 加密至 5000K 步长,432 → 9216 点;tlusty/synspec 静态二进制更新

统计与部署:
- grid 汇总改按 tlusty_status/synspec_status 分项计数,新增 tlusty_failed/
  synspec_failed/synspec_pending,前端详情页双视图适配
- deploy/fetch_results 支持跳板机 ProxyJump 与 SSH 主连接复用,fetch 新增 --force;
- Docker 构建支持 CARGO_MIRROR/USE_MIRRORS 国内镜像参数;移除 tools/ 拷贝
- 新增 tlusty-synspec-test skill 与 6 篇根因分析/验证文档
This commit is contained in:
fmq
2026-08-17 23:55:26 +08:00
parent 43b82b1ae2
commit b058e66722
54 changed files with 5624 additions and 347 deletions
+8 -2
View File
@@ -216,12 +216,16 @@ batch_deploy_profiles() {
[ -z "$r_port" ] && r_port="${REMOTE_PORT}"
local sock="${SSH_CONTROL_DIR}/${r_user}@${r_ip}:${r_port}"
if ssh -p "${r_port}" -o ControlPath="${sock}" -O check "${r_user}@${r_ip}" >/dev/null 2>&1; then
# 跳板机支持:profile 的 REMOTE_PROXY_JUMP(仅建立主连接时需要;-O check/exit 只操作本地 mux socket
local r_jump jump_opt=""
r_jump=$(grep -E '^REMOTE_PROXY_JUMP=' "$p" 2>/dev/null | cut -d'=' -f2- | tr -d '"' | tr -d "'" || true)
[ -n "${r_jump}" ] && jump_opt="-o ProxyJump=${r_jump}"
if ssh -p "${r_port}" -o ControlPath="${sock}" ${jump_opt} -O check "${r_user}@${r_ip}" >/dev/null 2>&1; then
echo " -> [${pre_idx}/${total}] ${r_user}@${r_ip}:${r_port} (连接已存在,复用)"
continue
fi
echo " -> [${pre_idx}/${total}] 正在连接 ${r_user}@${r_ip}:${r_port} ..."
ssh -p "${r_port}" -o ControlMaster=yes -o ControlPath="${sock}" -o ControlPersist=1800 -fN "${r_user}@${r_ip}" \
ssh -p "${r_port}" -o ControlMaster=yes -o ControlPath="${sock}" -o ControlPersist=1800 ${jump_opt} -fN "${r_user}@${r_ip}" \
&& echo -e " ${GREEN}[√] 连接已建立${NC}" \
|| echo -e " ${YELLOW}[!] 连接失败,将在部署该节点时重试${NC}"
done
@@ -699,6 +703,8 @@ setup_ssh_control() {
local socket_key="${REMOTE_USER}@${REMOTE_IP}:${REMOTE_PORT}"
SSH_CONTROL_PATH="${SSH_CONTROL_DIR}/${socket_key}"
SSH_OPTS="-o ControlMaster=auto -o ControlPath=${SSH_CONTROL_PATH} -o ControlPersist=1800"
# 非直连节点:profile 提供 REMOTE_PROXY_JUMP 时经跳板机转发(对齐 fetch_results.sh
[ -n "${REMOTE_PROXY_JUMP:-}" ] && SSH_OPTS="${SSH_OPTS} -o ProxyJump=${REMOTE_PROXY_JUMP}"
mkdir -p "${SSH_CONTROL_DIR}"
# 若已有活跃的主连接 (如批量部署父进程已建立),直接复用,避免重复认证。