前世今生：时间线与教训The history: a timeline and its lessons

1. 起点：AO-26 提案（2026-09/10）1. Origin: the AO-26 proposal (2026-09/10)

问题诞生在 AO-26 XMM 提案的准备期。提案的设想是：对几颗邻近的侧向旋涡星系做深度 XMM 指向，测它们的热晕，与 eROSITA 叠加对比。2026-09-26 团队在邮件里明确"旧提案思路作废"，新目标改为量化 eROSITA 叠加与深逐星测量之间的张力。提案最终选了 NGC 4565（165 ks）+ NGC 2613（85 ks）作为新的深场目标，样本 6–8 颗星系。The question was born while preparing the AO-26 XMM proposal. The idea was to take deep XMM pointings of a few nearby edge-on spirals, measure their hot halos, and compare with the eROSITA stacks. On 2026-09-26 the team agreed in email that "the old proposal idea is dead"; the new aim is to quantify the tension between the eROSITA stack and deep per-galaxy measurements. The proposal finally chose NGC 4565 (165 ks) + NGC 2613 (85 ks) as the new deep targets, with a sample of 6–8 galaxies.

这个阶段留下的关键事实：Edmund 的参考曲线是 Zhang 2024 的 MW 质量模型，3\sigma 上限，kT 0.2，用局部背景；他认为"源掩模比背景区更重要"；Edmund 的 NGC 5746 点其实是 Laktionov 等人 2025 发表的非风分量，他自己的测量是上限。这些后来都成了方法对比表的输入。Key facts left by this stage: Edmund's reference curve is the Zhang 2024 MW-mass model, 3\sigma upper limits, kT 0.2, local background; he thinks "source masking will matter more than the background region"; Edmund's NGC 5746 point is actually the non-wind value published by Laktionov et al. 2025, his own measurement being an upper limit. All of these later fed the method-comparison table.

来源：cognition-os/stages/S01_origin_in_ao26.md；XMMproposal2026 cognition-os/insights/I028_correspondence_2026-09-07_to_10-07.md（邮件史 E01–E42）。Source: cognition-os/stages/S01_origin_in_ao26.md; XMMproposal2026 cognition-os/insights/I028_correspondence_2026-09-07_to_10-07.md (the email history E01–E42).

2. ADAP 样本的复现（2026-10-09）2. The ADAP sample reproduction (2026-10-09)

2026-10-09 PI 问：如果按 eROSITA 的方式叠 XMM 档案会怎样？既然我们逐颗分辨，能不能说清张力的来源？第一步是复现 Li 2018 ADAP 提案的目标表（"XMM 档案里的热 CGM 遗产"），因为它是档案叠加的母表。On 2026-10-09 the PI asked: what if we stacked the XMM archive the eROSITA way, and since we resolve each galaxy, can we say what causes the tension? The first step was reproducing the Li 2018 ADAP proposal's target table (the "hot CGM as a legacy of the XMM-Newton archive"), because it is the parent list for an archival stack.

四个不同代理/模型各自独立复现。教训：总数对上了不等于名单对上了。codexnim 复现了提案的观测数与总曝光（2309 vs 2304，81.3 vs 81.3 Ms）并找到了曝光定义，但它的星系数比提案多 28%——因为 3′ 位置匹配把非星系指向附近的偶然星系也算进去了。四个代理的 \ge2 共识给出 1015 颗 / 82.6 Ms，被当作候选母表（不是提案的名单）。Four different agents/models reproduced it independently. Lesson: matching the totals does not mean matching the list. codexnim reproduced the proposal's observation count and exposure (2309 vs 2304, 81.3 vs 81.3 Ms) and found the exposure definition, but its galaxy count is 28% above the proposal — a 3′ position match also picks up serendipitous galaxies near non-galaxy pointings. The four agents' \ge2-agent consensus gives 1015 galaxies / 82.6 Ms, used as a candidate parent list (not the proposal's list).

来源：cognition-os/insights/I002_adap_sample_reproduction.md；XMMproposal2026 adap_sample_comparison/SUMMARY.md。Source: cognition-os/insights/I002_adap_sample_reproduction.md; XMMproposal2026 adap_sample_comparison/SUMMARY.md.

3. 立项目（2026-10-13）与零场试点3. The project starts (2026-10-13) and the null pilot

2026-10-13 PI 说"这部分讨论单独建一个项目吧"，CGMStacking 与它的 cognition-os 诞生。同天批准了两个第一步（D002）：NGC 5746 的 MOS vs pn 检验；零信号对照场试点。零场试点后来变成一部"排除法教科书"：On 2026-10-13 the PI said "let's give this discussion its own project", and CGMStacking with its cognition-os was born. The same day approved two first steps (D002): the NGC 5746 MOS vs pn test; and the null-signal control-field pilot. The null pilot became a textbook of elimination:

• 初选 10 个零场，其中 0135940101 两次撞上引擎的 XSA-vs-SUM.SAS 曝光一致性检查（pn U001 开始时间差 4.1 h）——教训：不要绕过检查，换掉这个场（PITFALLS）。The first 10 nulls included 0135940101, which twice hit the engine's XSA-vs-SUM.SAS exposure consistency check (pn U001 start time differs by 4.1 h) — lesson: do not bypass the check, replace the field (PITFALLS).
• 烟雾测试场 0111280101（SGP-1）是耀斑主导的：10 ks 名义曝光只剩 0.30/0.48 ks 干净 MOS。教训：档案标称曝光不等于干净曝光——重选时用 5XMM 流水线的 ONTIME 做干净代理（\(\ge 0.7\times\) 科学场，\(\ge 15\) ks），得到 9 个活动场（v2）。The smoke-test field 0111280101 (SGP-1) turned out flare-dominated: of 10 ks nominal, only 0.30/0.48 ks of clean MOS remained. Lesson: the archive's nominal exposure is not the clean exposure — re-select with the 5XMM pipeline ONTIME as a clean proxy (\(\ge 0.7\times\) the science field, \(\ge 15\) ks), giving 9 active fields (v2).
• 0822390601 在还原后被退役（真干净曝光 MOS1 13.7 ks < 15 ks），由 0804271301 顶替。教训：代理不能替代真筛选。0822390601 was retired after reduction (true clean exposure MOS1 13.7 ks < 15 ks), replaced by 0804271301. Lesson: a proxy does not replace the real screen.
• Virgo 上 ESAS 工人进程被钉在 CPU {0,40}（两个工人挤两个核，arfgen 单个跑 77 分钟）。教训：用 cap_affinity.py 重钉（PITFALLS）。On Virgo the ESAS worker processes were pinned to CPUs {0,40} (two cores for all workers; arfgen alone ran 77 min). Lesson: re-pin with cap_affinity.py (PITFALLS).
• 还原完成：9 个零场交付、逐场 sha256 校验。拟合按批次推进（batch 1–6），每批有执行者报告与自检。Reduction complete: 9 nulls delivered, sha256-verified per case. Fits proceeded in batches (1–6), each with an executor report and self-check.

来源：cognition-os/04_DECISION_LOG.md D002 及其附录（更正 1–11）；cognition-os/PITFALLS.md；null_pilot/selection/SELECTION_v2.md。Source: cognition-os/04_DECISION_LOG.md D002 and its addendum (corrections 1–11); cognition-os/PITFALLS.md; null_pilot/selection/SELECTION_v2.md.

4. MOS 偏置的故事与被拒绝的解释4. The MOS-offset story and the rejected explanations

零场拟合揭示了一个没有预料到的模式：MOS 的零场幅度系统性为负，pn 却为零。这个"偏置"的寻找过程本身是本项目最有教育意义的部分，一共四步排除：The null fits revealed an unexpected pattern: the MOS null amplitudes are systematically negative while pn is zero. The hunt for this "offset" is the most instructive part of the project — four steps of elimination:

1. 2026-10-15（六个零场）：以为是"共同偏置"，工作修正 +0.30。后来被推翻（R001）：九个零场平均 \(-0.19 \pm 0.06\)，但随场而变——常规场 \(-0.37 \pm 0.08\)，两个硬超额场 \(+0.06 \pm 0.10\)。不存在全局常数修正。2026-10-15 (six nulls): thought to be a "common offset", working correction +0.30. Later withdrawn (R001): nine nulls average \(-0.19 \pm 0.06\) but are field-dependent — normal fields \(-0.37 \pm 0.08\), two hard-excess fields \(+0.06 \pm 0.10\). No global constant exists.
2. 假设一：冻结的软质子谱指数。把 SP 指数放开重拟合三个常规场：SP-free \(-0.44 \pm 0.13\) vs 冻结 \(-0.43 \pm 0.11\)，不动。拒绝（R003）。Hypothesis 1: the frozen soft-proton index. Refit three normal fields with the SP index free: SP-free \(-0.44 \pm 0.13\) vs frozen \(-0.43 \pm 0.11\), no change. Rejected (R003).
3. 假设二：pn QPB 通过联合第一步天空泄漏进 MOS。这条线索来自 pn0 缩放实验（只缩放 pn0 时 MOS 会动）。预注册了判决规则（D003），然后 MOS-only 重拟合：\(-0.354 \pm 0.111\) vs 联合 \(-0.434 \pm 0.107\)，在阈值处拒绝（D004，R004；边界上，引用为"在阈值处拒绝"）。Hypothesis 2: pn QPB leaking into MOS through the joint step-1 sky. This lead came from the pn0 scaling experiment (MOS moves when only pn0 is rescaled). A verdict rule was pre-registered (D003), then the MOS-only refits: \(-0.354 \pm 0.111\) vs joint \(-0.434 \pm 0.107\), rejected at the threshold (D004, R004; on the boundary, quoted as "rejected at the threshold").
4. 假设三：天空模型在参照框里不稳定。PI 提示"QPB 与 SP 时变；天空与源大致稳定"。检查 (a)(b)(c) 找不到驱动；但发现天空在必须稳定的地方不稳定（同一场 N/S 两侧不同、同一目标四个历元不同）。稳定天空重拟合（每场一个天空）把常规场偏置移到 \(-0.268 \pm 0.079\)，按 D005 预注册规则裁决为不确定（D006，R005）。偏置的剩余部分原因仍开放。Hypothesis 3: the sky model is unstable in the reference box. The PI's guidance: "QPB and SP are time-variable; sky and source are roughly stable". Checks (a)(b)(c) found no driver, but the sky is indeed unstable where it must be stable (N/S sides of one field differ; four epochs of one target differ). The stable-sky refit (one sky per field) moves the normal-field offset to \(-0.268 \pm 0.079\); by the pre-registered D005 rule the verdict is ambiguous (D006, R005). The cause of the remaining offset is open.

同时，NGC 5746 的结论在每一步都保持成立：科学值减去它自己的两个零场，MOS \(+0.98 \pm 0.18\)、pn \(+0.87 \pm 0.14\)——超额不是仪器在探测器位置上制造的。Throughout, the NGC 5746 conclusion kept holding: science minus its own two nulls gives MOS \(+0.98 \pm 0.18\), pn \(+0.87 \pm 0.14\) — the excess is not made by the instrument at that detector position.

来源：cognition-os/insights/I006（(c)–(m) 全部更新）；cognition-os/07_REVISION_LOG.md R001–R005；cognition-os/04_DECISION_LOG.md D003–D006；null_pilot/fits/results_claude/mos_offset_table_regress.json（图）。Source: cognition-os/insights/I006 (all updates (c)–(m)); cognition-os/07_REVISION_LOG.md R001–R005; cognition-os/04_DECISION_LOG.md D003–D006; null_pilot/fits/results_claude/mos_offset_table_regress.json (figure).

5. 2026-10-21 的转向：Chandra 与 Vrot5. The 2026-10-21 turn: Chandra and Vrot

2026-10-21 PI 重启心跳并改变范围（D007）："可能我们一下子也定义不出来项目的全部细节，我们可以一点点完善项目细节和调整项目目标"。两个新方向进入：(A) Chandra 档案试点现在就做（不等 XMM 设计定稿）；(B) PI 假设：旋转速度比恒星质量更好地预测热 CGM。同天，普查 I007 用档案表数清了三台仪器各自能覆盖什么，结论是 Chandra 适合 Zhang 式远距离 M* 选样与点源深度检验，Vrot 检验先在 eRASS1 公共数据上做。研究任务 P1–P4、检索 S1–S2、红队 RT-01/RT-02 与本手册同时派出，全部待定。On 2026-10-21 the PI restarted the heartbeat and changed the scope (D007): "We probably cannot define all the project details at once; let us refine the details and adjust the goals step by step." Two new directions entered: (A) the Chandra archival pilot starts now (not waiting for the XMM design); (B) the PI hypothesis: rotation velocity predicts the hot CGM better than stellar mass. The same day the census I007 counted, from archive tables, what each of the three instruments can cover: Chandra suits the Zhang-like distant M* sample and the point-source-depth test; the Vrot test runs first on eRASS1 public data. Research tasks P1–P4, searches S1–S2, red teams RT-01/RT-02 and this handbook were dispatched together — all pending.

来源：cognition-os/04_DECISION_LOG.md D007（PI 原话）；cognition-os/insights/I007；cognition-os/08_RUNNING.md。Source: cognition-os/04_DECISION_LOG.md D007 (the PI's words); cognition-os/insights/I007; cognition-os/08_RUNNING.md.

6. 死胡同清单（教训）6. The dead ends (lessons)

来源：cognition-os/07_REVISION_LOG.md；cognition-os/PITFALLS.md；cognition-os/insights/I002、I006。Source: cognition-os/07_REVISION_LOG.md; cognition-os/PITFALLS.md; cognition-os/insights/I002, I006.