如何接手：不用 AI 也能继续How to continue by hand, without AI

1. 机器与分工1. Machines and roles

来源：cognition-os/history/08_RUNNING_until_2026-10-21.md §2（lang/lang2/Virgo 限制原文）。Source: cognition-os/history/08_RUNNING_until_2026-10-21.md §2 (the lang/lang2/Virgo limits as written).

2. 数据、代码与结果在哪2. Where data, code and results live

来源：README.md；cognition-os/01_PROJECT_INDEX.md；AGENTS.md。Source: README.md; cognition-os/01_PROJECT_INDEX.md; AGENTS.md.

3. 可复制命令3. Copy-paste commands

以下命令都在 2026-10-21 实际运行过，输出照抄。普查脚本没有 --help（是普通脚本，必须在含 data/ 的目录里运行），所以先复制输入再运行。临时目录用 $TMPDIR。Every command below was actually run on 2026-10-21 and the output is copied verbatim. The census scripts have no --help (plain scripts that must run from a directory containing data/), so copy the inputs first. Scratch goes to $TMPDIR.

3.1 重跑覆盖普查3.1 Re-run the coverage census

export TMPDIR=/u/Lang2/ai_tmp/hermes/handbook_2026-10-21-handbook-build
cd $TMPDIR && mkdir -p byhand/data && cd byhand
cp /u/Lang2/ai_tmp/claude/CGMStacking/census_2026-10-21/data/chanmaster.ecsv \
   /u/Lang2/ai_tmp/claude/CGMStacking/census_2026-10-21/data/xmmmaster.ecsv \
   /u/Lang2/ai_tmp/claude/CGMStacking/census_2026-10-21/data/hecate_m10.ecsv \
   /u/Lang2/ai_tmp/claude/CGMStacking/census_2026-10-21/data/sdss_mw.csv data/
cp /u/Lang3/program/CGMStacking/cognition-os/artifacts/scripts/chandra_census/count.py .
/home/huangrui/software/miniconda3/envs/ciao-4.18/bin/python count.py

实际输出（SDSS Zhang 式样品的行）：Actual output (the SDSS Zhang-like sample row):

## SDSS z0.01-0.2 logM*10.5-11.0 (Zhang-like, all, not only centrals): N=251561  (eRASS1-DE half: 108981)
   theta(R=50..250kpc) median arcmin: [0.4, 0.8, 1.2, 1.6, 2.0]
   ACIS-I  50kpc:3472(1585|152t)  100kpc:3159(1442|152t)  150kpc:2875(1305|152t)  200kpc:2603(1169|151t)  250kpc:2334(1040|148t)

括号里是曝光 \(\ge 20\) ks 的星系数与视场内目标数。count2.py 把 ACIS-I 与 ACIS-S 合并并去掉星系团目标场，给出 I007 (c) 表里的数：Inside the brackets: galaxies with exposure \(\ge 20\) ks and targets in the field. count2.py merges ACIS-I with ACIS-S and drops cluster-target fields, giving the numbers in I007 (c):

cp /u/Lang3/program/CGMStacking/cognition-os/artifacts/scripts/chandra_census/count2.py .
/home/huangrui/software/miniconda3/envs/ciao-4.18/bin/python count2.py

SDSS z0.01-0.2 logM*10.5-11.0 (Zhang-lik 50:4969(2116) 100:4341(1858) 150:3753(1608) 200:3267(1413) 250:2823(1216)

3.2 Vrot 选样普查3.2 The Vrot-selected census

cp /u/Lang2/ai_tmp/claude/CGMStacking/census_2026-10-21/data/leda_try.txt \
   /u/Lang2/ai_tmp/claude/CGMStacking/census_2026-10-21/data/alfalfa.ecsv \
   /u/Lang2/ai_tmp/claude/CGMStacking/census_2026-10-21/data/durbala.ecsv \
   /u/Lang2/ai_tmp/claude/CGMStacking/census_2026-10-21/data/fashi.ecsv data/
cp /u/Lang3/program/CGMStacking/cognition-os/artifacts/scripts/chandra_census/vrot.py .
/home/huangrui/software/miniconda3/envs/ciao-4.18/bin/python vrot.py

## HyperLEDA Vrot 150-300, i>=45: N=9566 (eRASS1-DE 4638)  median D_A=114 Mpc
   Chandra  50:203(75) 100:101(32) 150:49(15) 200:23(8) 250:13(3)
   XMM      50:570(411) 100:447(319) 150:361(256) 200:271(194) 250:217(154)

3.3 读零场结果表3.3 Read the null-field result tables

head -1 /u/Lang3/program/CGMStacking/null_pilot/fits/results_claude/stablesky/0600090401/mos_pn_table.md
grep "N+S" /u/Lang3/program/CGMStacking/null_pilot/fits/results_claude/stablesky/0600090401/mos_pn_table.md

# NULL0600090401: MOS vs pn excess amplitude A (1e-6 APEC norm per arcmin^2, kT 0.3; = K units)
| N+S | 1_2 | 1.50 | -0.00 ± 0.26 (1.0/1) | +0.13 ± 0.16 (0.0/1) | -0.13 (-0.4σ) | ...

D005 汇总（九个零场、两种天空模型）：The D005 summary (nine nulls, two sky models):

head -8 /u/Lang3/program/CGMStacking/null_pilot/fits/results_claude/stablesky/d005_summary.json

{
 "10p1_30p0": {
  "stable": [
   -0.267643793079895,
   0.07927339444890803,
   8.074421287104768,
   6

3.4 查询 Chandra/XMM 档案表（HEASARC TAP）3.4 Query the Chandra/XMM archive tables (HEASARC TAP)

export TMPDIR=/u/Lang2/ai_tmp/hermes/handbook_2026-10-21-handbook-build
cat > $TMPDIR/tap_query.py <<'EOF'
import pyvo
h = pyvo.dal.TAPService("https://heasarc.gsfc.nasa.gov/xamin/vo/tap")
t = h.run_async("select top 5 obsid, status, name, ra, dec, detector, exposure from chanmaster "
                "where status='archived' and grating='NONE' and detector like 'ACIS%'").to_table()
print(t)
EOF
/home/huangrui/software/miniconda3/envs/ciao-4.18/bin/python $TMPDIR/tap_query.py

obsid  status          name             ra       dec    detector exposure
                                       deg       deg                s
----- -------- -------------------- --------- --------- -------- --------
21421 archived          ESO005-G004  91.42333 -86.63194   ACIS-S    23060
10143 archived 1RXSJ200924.1-853911 302.30417 -85.64633   ACIS-S     2030
22326 archived          RE J0317-853  49.31604 -85.54043   ACIS-I    37930

需要 pyvo（pip install pyvo）。普查脚本 fetch.py 用同一 TAP 服务拉全表；它写的是相对路径 data/，所以也要在含 data/ 的目录运行。Needs pyvo (pip install pyvo). The census script fetch.py pulls the full tables from the same TAP service; it writes to the relative path data/, so it also must run from a directory containing data/.

来源：cognition-os/artifacts/scripts/chandra_census/{count.py,count2.py,vrot.py,fetch.py}；null_pilot/fits/results_claude/stablesky/。命令于 2026-10-21 由本手册构建实际运行。Source: cognition-os/artifacts/scripts/chandra_census/{count.py,count2.py,vrot.py,fetch.py}; null_pilot/fits/results_claude/stablesky/. The commands were actually run by this handbook build on 2026-10-21.

4. 必须遵守的规则4. Rules to keep

1. lang 与 Virgo 整机 CPU 和内存 \(\le 60\%\)；看住磁盘空间。Whole-machine CPU and memory \(\le 60\%\) on lang and Virgo; watch the disks.
2. 临时文件只写 /u/Lang2/ai_tmp/<agent>/；永不写 /tmp、/var/tmp、~。Scratch only in /u/Lang2/ai_tmp/<agent>/; never /tmp, /var/tmp or ~.
3. 还原（reduction）在 Virgo；拟合在 lang（lang2 也可）；永不在 Virgo 拟合。Reduction on Virgo; fitting on lang (lang2 also); never fit on Virgo.
4. ../XMMproposal2026/ 从本项目只读。../XMMproposal2026/ is read-only from this project.
5. 校准问题（xsherpa、ESAS 引擎、QPB 标定）报告给 DiffuseXrayCalibration 的 cognition-os/tool_iteration.md，不在本地修。Calibration problems (xsherpa, the ESAS engine, QPB normalisation) are reported to DiffuseXrayCalibration's cognition-os/tool_iteration.md, not fixed locally.
6. 每个触碰 cognition-os/ 的回合在 06_AGENT_LOG.md 追加一行。Every turn that touches cognition-os/ appends one row to 06_AGENT_LOG.md.
7. 远程 stdin 脚本一律 timeout（ssh virgo timeout 600 python -B - <<EOF）。Remote stdin scripts always under timeout (ssh virgo timeout 600 python -B - <<EOF).
8. xsherpa 用锁定版本 /u/Lang3/program/xsherpa/releases/current，永不 allow_deprecated；QPB 缩放只作为 pn-only 测试（DXC D016）。xsherpa via the pinned release /u/Lang3/program/xsherpa/releases/current, never allow_deprecated; QPB scaling only as a pn-only test (DXC D016).

来源：AGENTS.md；cognition-os/PITFALLS.md。Source: AGENTS.md; cognition-os/PITFALLS.md.

5. 下一步（编号计划）5. Next steps (numbered plan)

每一步给出：目标、输入、方法、成功/失败判据、时间估计、先读哪个文件。运行中的任务（P1–P4、S1–S2、RT-01/RT-02）结果待定；分诊它们的报告是第 1–2 步。Each step gives: goal, inputs, method, how to tell success from failure, time estimate, and which file to read first. The running tasks (P1–P4, S1–S2, RT-01/RT-02) are pending; triaging their reports is steps 1–2.

步骤 1：分诊 Chandra S/N 预报（P1）Step 1: triage the Chandra S/N forecast (P1)

• 目标：知道 Zhang 式 Chandra 叠加在 50–250 kpc 每个半径档的可达 S/N。Goal: know the reachable S/N per radial bin (50–250 kpc) for a Zhang-like Chandra stack.
• 输入：P1 报告（.agent/jobs/2026-10-21-P1-chandra-snr/）。Input: the P1 report (.agent/jobs/2026-10-21-P1-chandra-snr/).
• 方法：读报告，把每个 (N, 曝光, 半径) 的 S/N 抄进信任表。Method: read the report and copy the S/N for each (N, exposure, radius) into the trust table.
• 成功：50–150 kpc 的 S/N \(\ge 3\) 在现实筛选下可达 \(\to\) 走步骤 3。失败：S/N 不够 \(\to\) Chandra 只做点源深度检验（P4）。Success: S/N \(\ge 3\) at 50–150 kpc reachable with realistic cuts \(\to\) go to step 3. Failure: S/N insufficient \(\to\) Chandra does only the point-source-depth test (P4).
• 时间：约 1 小时（读 + 记录）。先读：08_RUNNING.md，再读 P1 的 REPORT。Time: about 1 h (read + record). Read first: 08_RUNNING.md, then the P1 REPORT.

步骤 2：分诊 Vrot 检验设计与文献（P2、S1、S2）Step 2: triage the Vrot test design and literature (P2, S1, S2)

• 目标：知道 eRASS1 公共数据能否裁决 Vrot vs M*，需要多大的 N。Goal: know whether eRASS1 public data can decide Vrot vs M*, and the N needed.
• 输入：P2 报告（.agent/jobs/2026-10-21-P2-erass1-vrot-test/）、S1/S2 检索报告。Inputs: the P2 report (.agent/jobs/2026-10-21-P2-erass1-vrot-test/), the S1/S2 search reports.
• 成功：eRASS1 足够 \(\to\) 运行该叠加（公共数据，lang）。失败：不够 \(\to\) Vrot 检验推迟到 eRASS:4。Success: eRASS1 suffices \(\to\) run that stack (public data, on lang). Failure: insufficient \(\to\) the Vrot test waits for eRASS:4.
• 时间：约 1–2 小时。先读：P2 的 REPORT，再读 S2 的文献清单。Time: about 1–2 h. Read first: the P2 REPORT, then the S2 reading list.

步骤 3：Chandra 还原试点（约 20 个场）Step 3: Chandra reduction pilot (~20 fields)

• 目标：Zhang 式远距离 M* 选样中约 20 个场的真实数据走通 Chandra 还原。Goal: run the Chandra reduction end-to-end on ~20 fields of the Zhang-like distant M* sample.
• 输入：P3 的现实选样（真实 aimpoint + roll，去星系团成员）。Input: the P3 realistic sample (real aimpoint + roll, group/cluster members dropped).
• 方法：CIAO 标准还原（lang 单线程，限 60%）；Chandra 还原主机与 PI 商定（Virgo 只做 XMM）。Method: standard CIAO reduction (lang single-threaded, 60% cap); the Chandra reduction host is to be decided with the PI (Virgo is for XMM only).
• 成功：每个场产出可拟合的谱与图像；失败：某一步在 \(> 25\%\) 的场失败 \(\to\) 停下写坑记录。Success: fittable spectra and images for every field; failure: any step fails in \(> 25\%\) of fields \(\to\) stop and write a pitfall.
• 时间：数天（机器时间为主）。先读：08_RUNNING.md §7 第 1 条。Time: several days (mostly machine time). Read first: 08_RUNNING.md §7 item 1.

步骤 4：点源深度检验（P4 设计）Step 4: the point-source-depth test (P4 design)

• 目标：eRASS1（与 eRASS:4）探测限和 Chandra 极限之间的点源通量，占叠加超额的多大比例。Goal: what fraction of the stacked excess is point-source flux between the eRASS1 (and eRASS:4) detection limit and the Chandra limit.
• 输入：CSC 2.1 点源表 + Chandra 覆盖的星系 + Zhang/He & Li 叠加 SB。Inputs: the CSC 2.1 point-source catalogue, the Chandra-covered galaxies, and the Zhang/He & Li stacked SB.
• 成功：得到一个带误差的占比数字；失败：CSC 覆盖不足 \(\to\) 改用 Chandra 自己的源表。Success: a fraction with an error bar; failure: CSC coverage insufficient \(\to\) use Chandra's own source list instead.
• 时间：约 1 天。先读：P4 的 BRIEF.md。Time: about 1 day. Read first: P4's BRIEF.md.

步骤 5：红队分诊（RT-01 普查、RT-02 \(0.17 \pm 0.10\)）Step 5: triage the red teams (RT-01 census, RT-02 the \(0.17 \pm 0.10\) number)

• 目标：普查数字与 \(0.17 \pm 0.10\) 这个承载全局的数经过独立复核后定级。Goal: after independent re-analysis, set the trust of the census numbers and of the load-bearing \(0.17 \pm 0.10\).
• 方法：读两份红队报告；有推翻性发现就更新 03_TRUST_TABLE.md 与本手册。Method: read both reports; if a finding overturns a number, update 03_TRUST_TABLE.md and this handbook.
• 时间：约 1 小时。先读：REDTEAM_QUEUE.md。Time: about 1 h. Read first: REDTEAM_QUEUE.md.

步骤 6（停泊）：XMM 零场的 MOS 偏置Step 6 (parked): the XMM null-field MOS offset

• 两个 PI 同意的检验（固定 CXB 天空；NGC 5746 多历元）保持停泊，直到 DiffuseXrayCalibration 校准完成；在此之前，MOS 科学值与零值必须来自同一拟合变体（D006 规则）。The two PI-agreed tests (fixed-CXB sky; NGC 5746 multi-epoch) stay parked until DiffuseXrayCalibration is done; until then a MOS science value and its null must come from the same fit variant (the D006 rule).

来源：cognition-os/08_RUNNING.md §7；cognition-os/RESEARCH_PLAN.md；cognition-os/04_DECISION_LOG.md D006。步骤 1–5 的顺序是本手册的建议（未经评审）。Source: cognition-os/08_RUNNING.md §7; cognition-os/RESEARCH_PLAN.md; cognition-os/04_DECISION_LOG.md D006. The order of steps 1–5 is this handbook's suggestion (not reviewed).