总览 · ← 返回运行 20261003-093415-search-t1-r2-D-s0

节点 n14 在终选来历上

composition_trend + α=3 重锚定 + 两阶段逐基因方向位移（按细胞型 EB 收缩，β=1；仅双输入视图激活，X3 A 半 54.51→54.88，proxy10 与父逐字节一致）

方法说明?节点程序自带的 METHOD.md：这个程序做了什么、为什么。

来自 programs.git ba44b165dc:solution/METHOD.md

调研员的计划

代码改动?这个节点的程序和父节点程序的逐行差别：绿色是新增，红色是删除。

对比：父节点版本 da1e39c0bf。改动的文件：solution/METHOD.md +70 −75、solution/run.py +141 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 35aa650..2a3e363 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,83 +1,78 @@-# composition_trend + 型内表达重锚定，α 扫描延伸至拐点：α=3.0 出厂（α=2.5/3.0/3.5 → 节点分 55.78/55.87/55.72），新增按型收缩开关 k（出厂关）--Parent (node 8, 55.48): composition_trend weighted sampling (type-level A_TP=-0.55, cell-level-A_CM=1.2/A_FATE=0.6, K=5 strata) + per-type expression re-anchoring at α=2.0 — each sampled cell-shifted by α·(μ_in(t) − μ_sel(t)) where its value is >0, clipped ≥0, types with <5 sampled cells-skipped. Parent's own next-step suggestion #1: extend the α screen to find the turning point,-watching proxy10 de_score and X3 variogram. This node does exactly that and ships the peak.--## Delta vs parent--1. `RECENTER_ALPHA` 2.0 → **3.0**, chosen by a 3-point paired extension of the parent's 5-point-   monotone screen (same protocol: seed 0, both views, node score = (2·X3 + proxy10)/3, A-half).-2. New **per-type shrinkage** knob `RECENTER_K` (a_t = α·n_sel(t)/(n_sel(t)+k), env-   `VEC_RECENTER_K`; ships k=0 (constant α, parent behaviour). Implements parent suggestion #2,-   but analysis of the screen data rejects k>0: X3's types are small (~65 sampled cells) and X3-   is the ruler that *gains* from larger α, while proxy10's types are large (~284) — shrinkage-   would pull X3 back toward α≈2 (−2 pts) while barely touching proxy. Kept as an off control.-3. Diagnostics only otherwise; sampling, weights, quotas, λ-path (off), write-out unchanged.--## Screen (vec-score A-half, seed 0; parent α=2 baseline 52.40/61.00 → 55.27)--| α | X3 | proxy10 | node score | X3 mmd_u | X3 de_score | p10 de_score |-|---|---|---|---|---|---|---|-| 2.5 | 53.72 | 59.89 | 55.78 | 0.02844 | −0.043 | 0.107 |-| **3.0 (ships)** | **54.51** | **58.59** | **55.87** | 0.02724 | −0.029 | 0.079 |-| 3.5 | 54.99 | 57.18 | 55.72 | 0.02620 | −0.043 | 0.049 |--- **Turning point found at α=3.0**: X3 still rises at 3.5 but only +0.48 while proxy10 drops-  −1.41 → net −0.15. Combined with the parent's 5-point screen (53.69/53.95/54.41/54.82/55.27 at-  α=0/0.5/1/1.5/2) this is an 8-point unimodal curve; the +0.60 A-half gain over α=2 is below-  the ~2-pt single-query noise but the shape (rise–peak–fall) is not a noise artefact pattern.-- Component response α=2→3: X3 mmd_u 0.02889→0.02724 (skill 0.547→0.581), de_score −0.117→−0.029-  (0.463→0.491), de_direction 0.078→0.084, variogram 0.001255→0.001204 — all four improve;-  proxy10 de_score 0.137→0.079 (−1.2 pts), de_direction/mmd_u/variogram mildly down. Same-  trade-off as the parent, weighted 2:1 in X3's favour.-- X3 variogram did NOT turn down (parent's watch item): 0.001255→0.001227→0.001204→0.001200.--## Mechanism evidence / controls--- Mechanism-off: `VEC_RECENTER_ALPHA=0` returns the sampled matrix unchanged (byte-identical to-  grandparent node 4, verified there); `VEC_RECENTER_K` with α=0 is a no-op. Shifts are computed-  from the view's own input cells per type (μ_in over all input cells of the type vs μ_sel over-  sampled cells), applied only at nonzero entries, clipped ≥0 — sparsity and covariation preserved.-- Default run (no env) is **byte-identical** to the `VEC_RECENTER_ALPHA=3.0` run on both views-  (verified via h5 data/indptr/indices comparison).-- View-independence: no manifest identity fields, no absolute times, no file names; identical code-  path for 1- or 2-input views (anchor = latest input's own cells). Deterministic given --seed-  (recenter block has no RNG).+# composition_trend + α=3 重锚定 + 两阶段逐基因方向位移（按细胞型 EB 收缩，β=1；仅双输入视图激活，X3 A 半 54.51→54.88，proxy10 与父逐字节一致）++Parent (node 10, 55.86): composition_trend weighted sampling + per-type re-anchoring α=3.0.+PLAN asked for a two-stage gene-level EB-shrunk direction shift to lift X3 de_recovery+(de_score −0.052, below floor). Implemented as specified, plus a per-type variant found by screening.++## Method (family: composition_program; mechanism = two-stage pseudobulk delta direction shift)++Active only when the view has ≥2 input stages (sorted by time; uses the last two):++1. Δμ_g = pb(later) − pb(earlier), log space. Two modes:+   - `global`: one Δμ per gene; EB shrinkage w_g = n_g/(n_g+K), K=30, n_g = detected-cell count+     over both stages;+   - `type` (ships): per-cell-type Δμ_t using the type's cells in each stage, shrunk by+     min(n1_g,n2_g)/(min+K); genes with <20 detections in either stage of that type fall back to+     the global shrunk Δμ.+2. Sparse gate: keep |Δμ_s| > ZMIN·median(|Δμ_s|), ZMIN=1.5.+3. Apply to each sampled cell after α=3 recentering: x[i,g] += clip(β·Δμ_s, ±MAX_SHIFT), MAX_SHIFT=1.0,+   only at existing nonzeros, clipped ≥0 (sparsity and covariation preserved). β=1.0 ships.+4. Single-input views (proxy10): block never runs → output byte-identical to parent node 10+   (verified via h5 X/data comparison). No RNG in the block; deterministic.++Env knobs: VEC_SHIFT_BETA / VEC_SHIFT_K / VEC_SHIFT_ZMIN / VEC_SHIFT_MAX / VEC_SHIFT_MODE.++## Screen (vec-score A-half, seed 0; X3 parent = 54.51)++| variant | X3 board | de_score | de_dir | mmd_u | variogram |+|---|---|---|---|---|---|+| global β=0.5 | 54.27 | −0.014 | 0.076 | 0.0278 | 0.001215 |+| global β=5 | 47.46 | −0.229 | −0.008 | 0.0360 | 0.001601 |+| global β=20 | 38.53 | −0.157 | −0.061 | 0.0581 | 0.003925 |+| type β=5 | 53.71 | −0.057 | 0.022 | 0.0250 | 0.001412 |+| type β=15 | 48.35 | −0.129 | −0.028 | 0.0296 | 0.002165 |+| **type β=1 (ships)** | **54.88** | −0.043 | 0.075 | 0.0259 | 0.001206 |+| type β=2 | 54.88 | −0.057 | 0.057 | 0.0250 | 0.001232 |+| type β=3 | 54.88 | −0.029 | 0.043 | 0.0246 | 0.001279 |++- **Global mode is actively harmful when amplified** (PLAN risk #3 realized): the whole-embryo+  E8.75→E9.0 Qiu delta, scaled up, moves X3 in the wrong direction on all four metrics. Ships OFF.+- **Typed mode gives a small, flat-topped gain** (β=1/2/3 all 54.88, +0.37 over parent, within the+  ~2-pt noise): the gain is mmd_u (0.0272→0.0259, +0.6 pts), de_score/de_direction ~unchanged at+  β=1, variogram unchanged. **The PLAN's target group (de_recovery) was NOT lifted** — X3 de_score+  stays below floor. The honest read: per-type deltas mostly reposition cell states slightly toward+  the later stage (helps MMD), not a true DE direction recovery.+- β=1 chosen over 2/3 (identical board) because it preserves de_direction best (0.075 vs 0.043).++## Mechanism evidence (off control & diagnostics, X3 seed 0)++- Off control VEC_SHIFT_BETA=0: X3 output byte-identical to parent run (verified).+- Active (typed β=1): 186,400 (type,gene) pairs pass the gate (median gate over |Δμ_s|);+  1.18M nonzero entries touched, all changed; mean |shift| ≈ 0.041 log units; shifts are+  type-specific (not a constant vector): per-type Δμ differ by construction; >0 within-type+  dispersion since gate+fallback differ per type. Cells changed: all 652 sampled X3 cells.+- Proxy10 default run byte-identical to parent (1 input → block inactive), so proxy10 score is+  unchanged by construction (no query spent re-verifying).  ## Verified / not verified -- Verified: vec-check ok on proxy and X3; default ≡ α=3 override; runtime ~85 s proxy / ~40 s X3-  (limit 30 min); memory ~7 GB (limit 28).-- Not verified: B-half scores (A-half only; unimodal 8-point curve argues signal); seeds ≠ 0;-  α between 3.0 and 3.5; parent suggestion #3 (independent direction signal, e.g. CollecTRI--  constrained shifts) — not attempted, remains the open lever for X3's still-sub-floor de_score.-- Queries used: 6 of 20.+- Verified: vec-check ok on proxy and X3; off-control byte-identity; proxy byte-identity;+  default == VEC_SHIFT_BETA=1 VEC_SHIFT_MODE=type run; determinism (no RNG in block).+- Not verified: B-half (A-half gain +0.37 is within noise — may not replicate); seeds ≠0; K/ZMIN+  grids (only β scanned); behaviour on the final view (E8.5+E9.5 official, larger types) — the+  typed delta there is official-data-based, expected better-powered than X3's 2.2k-cell Qiu stages.+- Queries used: 9 of 20. -## Hyper-parameters+## Knowledge used -| Name | Value | Source |-|---|---|---|-| RECENTER_ALPHA | 3.0 | 8-point screen above (peak) |-| RECENTER_K | 0.0 (off) | analysis above; k>0 hurts the gaining ruler |-| RECENTER_MIN_CELLS | 5 | node 8 |-| parent constants (A_TP, A_CM, A_FATE, K_STRAT, LAM=0) | unchanged | nodes 3/4 |--Knowledge used: none beyond the parent (textbook cell-cycle / OXPHOS / glycolysis gene sets).-No external dataset, no held-out stage, no forbidden-window information, no prior files used.--## Honesty notes--- α=3 remains an empirical over-anchoring (shifts past μ_in by 2× the selection bias); the peak-  location is fit to A-half scores of two rulers. If B-half disagrees, α=1 is the principled value.-- proxy10 keeps losing de_score with α (0.268 at α=0 → 0.079 at α=3): the maturity-selection bias-  IS proxy's DE signal and is being removed; X3 (weight 2) says that bias points the wrong way.+None beyond parent (textbook cell-cycle/OXPHOS/glycolysis sets). No external files, no held-out+stage info, no forbidden-window measurements; the two-stage delta is computed live from the view's+own inputs. View-independent: uses only input order by time and time differences implicitly (none+hard-coded), no manifest identity fields, no absolute times, no file names.  ## Next direction -X3 de_score is still below floor (−0.029) and re-anchoring has saturated: add an independent-direction signal (CollecTRI/regulator-activity-constrained per-gene shifts, or two-stage-pseudobulk delta on views that have two inputs with graceful fallback), screened with the same-paired protocol.+Typed deltas help mmd_u only; de_recovery on X3 remains sub-floor for every variant tried. A+direction signal from prior/ regulatory knowledge (CollecTRI TF targets, Reactome cardiac+pathways) constrained to genes with two-stage support is the untested lever; alternatively accept+de_recovery as floor-bound on X3 and push mmd_u (composition) further.diff --git a/solution/run.py b/solution/run.pyindex 856c9d4..bcfe293 100644--- a/solution/run.py+++ b/solution/run.py@@ -62,6 +62,19 @@ RECENTER_K = 0.0       # per-type shrinkage of alpha: a_t = alpha * n_sel(t)/(n_                        # alpha, proxy types large (~284) — shrinkage would pull X3 back toward                        # alpha=2 while barely helping proxy. Kept as an off control. +SHIFT_BETA = 1.0       # per-type two-stage direction shift (only when >=2 input stages; ships MODE=type).+                       # Per gene: dmu_g = pb(last) - pb(prev); EB shrinkage w_g = n_g/(n_g+K) with+                       # n_g = detected-cell count over both stages; sparse gate |dmu_s| > ZMIN*median;+                       # x[i,g] += clip(BETA*dmu_s_g, +-MAX_SHIFT) at nonzero entries, clipped >=0.+                                              # Screened on X3 (A-half, seed 0): global mode negative at every beta+                       # (0.5/5/20 -> 54.27/47.46/38.53 vs parent 54.51); typed mode BETA 1/2/3 ->+                       # 54.88/54.88/54.88 (mmd_u gain, DE ~flat). Ships typed BETA=1, MODE=type.+                       # Single-input views (proxy10): block inactive, byte-identical to node 10.+                       # Env overrides: VEC_SHIFT_BETA / VEC_SHIFT_K / VEC_SHIFT_ZMIN / VEC_SHIFT_MAX.+SHIFT_K = 30.0         # EB shrinkage constant (detection-count based)+SHIFT_ZMIN = 1.5       # sparse gate: multiple of median |dmu_shrunk|+SHIFT_MAX = 1.0        # per-gene shift cap (log space)+ CYCLE = [     "Mki67", "Top2a", "Pcna", "Ccna2", "Ccnb1", "Ccnb2", "Ccnd1", "Ccneg", "Ccne1",     "Cdk1", "Cdk2", "Cdk4", "Cdk6", "Mcm2", "Mcm3", "Mcm4", "Mcm5", "Mcm6", "Mcm7",@@ -342,6 +355,111 @@ def pseudobulk(Xm):     return np.asarray(Xm.tocsr().mean(axis=0), dtype=np.float64).ravel()  +def gene_delta(view, entries, genes, k_shrink):+    """Two-stage per-gene direction signal from the last two input stages (sorted by time).++    dmu_g = pb(later) - pb(earlier) in log space; EB shrinkage w_g = n_g/(n_g+k) where n_g is the+    number of cells the gene is detected in across both stages (low-detection genes pulled to 0).+    Genes not covered by the later stage are filled from earlier-stage means by read_stage -> dmu=0.+    """+    a1 = read_stage(view, entries[-2], genes, missing="fill")+    a2 = read_stage(view, entries[-1], genes, missing="fill")+    X1, X2 = a1.X.tocsr(), a2.X.tocsr()+    dmu = np.asarray(X2.mean(axis=0), dtype=np.float64).ravel() - np.asarray(X1.mean(axis=0), dtype=np.float64).ravel()+    n_g = np.asarray(X1.getnnz(axis=0)).ravel() + np.asarray(X2.getnnz(axis=0)).ravel()+    w = n_g / (n_g + float(k_shrink))+    return dmu * w, dmu, w+++def gene_delta_by_type(view, entries, genes, labels, inv, k_shrink, min_det=20):+    """Per-type two-stage delta: dmu_t = pb_t(later) - pb_t(earlier), EB-shrunk by within-type+    detection count; types whose gene has < min_det detections in either stage fall back to the+    global delta. Returns dmus_by_type (n_types, n_genes)."""+    a1 = read_stage(view, entries[-2], genes, missing="fill")+    a2 = read_stage(view, entries[-1], genes, missing="fill")+    lab1 = labels_of(a1) if "celltype" in a1.obs.columns else np.full(a1.n_obs, "all")+    lab2 = labels_of(a2) if "celltype" in a2.obs.columns else np.full(a2.n_obs, "all")+    X1, X2 = a1.X.tocsr(), a2.X.tocsr()+    gmu1 = np.asarray(X1.mean(axis=0)).ravel()+    gmu2 = np.asarray(X2.mean(axis=0)).ravel()+    gdmu = gmu2 - gmu1+    gn = np.asarray(X1.getnnz(axis=0)).ravel() + np.asarray(X2.getnnz(axis=0)).ravel()+    gdmus = gdmu * (gn / (gn + float(k_shrink)))+    uniq = np.unique(labels)+    out = np.tile(gdmus, (len(uniq), 1))+    pos = {t: i for i, t in enumerate(uniq)}+    for t in np.unique(np.concatenate([lab1, lab2])):+        if t not in pos:+            continue+        r1 = np.where(lab1 == t)[0]+        r2 = np.where(lab2 == t)[0]+        if len(r1) == 0 or len(r2) == 0:+            continue+        m1 = np.asarray(X1[r1].mean(axis=0)).ravel()+        m2 = np.asarray(X2[r2].mean(axis=0)).ravel()+        n1 = np.asarray(X1[r1].getnnz(axis=0)).ravel()+        n2 = np.asarray(X2[r2].getnnz(axis=0)).ravel()+        d = (m2 - m1) * (np.minimum(n1, n2) / (np.minimum(n1, n2) + float(k_shrink)))+        ok = (n1 >= min_det) & (n2 >= min_det)+        out[pos[t]] = np.where(ok, d, gdmus)+    return out+++def apply_gene_shift_typed(Xsel, dmus_t, inv_out, beta, zmin, max_shift):+    from scipy import sparse as sp++    med = float(np.median(np.abs(dmus_t)))+    keep = np.abs(dmus_t) > zmin * med if med > 0 else np.zeros(dmus_t.shape, dtype=bool)+    D = np.where(keep, np.clip(beta * dmus_t, -max_shift, max_shift), 0.0).astype(np.float32)+    Xc = sp.csr_matrix(Xsel).copy()+    row_of = np.repeat(np.asarray(inv_out), np.diff(Xc.indptr))+    dvec = D[row_of, Xc.indices]+    m = dvec != 0+    before = Xc.data[m].copy()+    Xc.data[m] = np.maximum(Xc.data[m] + dvec[m], 0.0)+    changed = Xc.data[m] != before+    Xc.eliminate_zeros()+    diag = {+        "genes_kept_total": int(keep.sum()),+        "entries_touched": int(m.sum()),+        "entries_changed": int(changed.sum()),+        "mean_abs_shift": float(np.abs(dvec[m]).mean()) if m.any() else 0.0,+    }+    return Xc, diag+++def apply_gene_shift(Xsel, dmus, beta, zmin, max_shift):+    """Add clip(beta*dmus, +-max) to each sampled cell at nonzero entries of gated genes, clip >=0.++    Gate: |dmus_g| > zmin * median(|dmus|). Sparsity/covariation preserved (only existing nonzeros+    modified). Returns (csr, diag). beta=0 -> untouched.+    """+    from scipy import sparse as sp++    med = float(np.median(np.abs(dmus)))+    keep = np.abs(dmus) > zmin * med if med > 0 else np.zeros(len(dmus), dtype=bool)+    d = np.where(keep, np.clip(beta * dmus, -max_shift, max_shift), 0.0).astype(np.float32)+    Xc = sp.csr_matrix(Xsel).copy()+    if not keep.any():+        return Xc, {"genes_kept": 0}+    dj = d[Xc.indices]+    m = dj != 0+    before = Xc.data[m].copy()+    Xc.data[m] = np.maximum(Xc.data[m] + dj[m], 0.0)+    changed = Xc.data[m] != before+    Xc.eliminate_zeros()+    diag = {+        "genes_kept": int(keep.sum()),+        "entries_touched": int(m.sum()),+        "entries_changed": int(changed.sum()),+        "genes_up": int((d[keep] > 0).sum()),+        "genes_dn": int((d[keep] < 0).sum()),+        "mean_abs_shift": float(np.abs(d[keep]).mean()),+        "median_abs_dmus": med,+    }+    return Xc, diag++ def main() -> None:     ap = argparse.ArgumentParser()     ap.add_argument("--data", required=True)@@ -388,8 +506,31 @@ def main() -> None:     if lam != 0.0 and delta != 0.0:         beta = type_slopes(X, z_fate, inv, len(uniq))         Xsel = apply_shift(Xsel, inv[idx], beta, lam, delta, act_topq=actq)+    shift_beta = float(os.environ.get("VEC_SHIFT_BETA", SHIFT_BETA))+    shift_k = float(os.environ.get("VEC_SHIFT_K", SHIFT_K))+    shift_zmin = float(os.environ.get("VEC_SHIFT_ZMIN", SHIFT_ZMIN))+    shift_max = float(os.environ.get("VEC_SHIFT_MAX", SHIFT_MAX))     Xpre = Xsel     Xsel, diag = recenter(X, inv, Xsel, inv[idx], alpha, k_adapt=k_adapt)+    n_inputs = len(manifest["inputs"])+    print(f"[inputs] n_input_stages={n_inputs} shift_beta={shift_beta}", file=sys.stderr)+    shift_mode = os.environ.get("VEC_SHIFT_MODE", "type")   # global | type (screening knob)+    if shift_beta != 0.0 and n_inputs >= 2:+        entries = sorted(manifest["inputs"], key=lambda e: float(e["time"]))+        if shift_mode == "type":+            dmus_t = gene_delta_by_type(args.data, entries, genes, labels, inv, shift_k)+            Xsel, sdiag = apply_gene_shift_typed(Xsel, dmus_t, inv[idx], shift_beta, shift_zmin, shift_max)+            sdiag["mode"] = "type"+        else:+            dmus, dmu_raw, w = gene_delta(args.data, entries, genes, shift_k)+            Xsel, sdiag = apply_gene_shift(Xsel, dmus, shift_beta, shift_zmin, shift_max)+            pb_ref = pseudobulk(X)+            dp = pseudobulk(Xsel) - pb_ref+            cos = float(dp @ dmus / max(np.linalg.norm(dp) * np.linalg.norm(dmus), 1e-12))+            sdiag["dp_cosine_with_dmus"] = cos+            sdiag["n_dmus_nonzero"] = int((np.abs(dmu_raw) > 1e-9).sum())+            sdiag["mode"] = "global"+        print(f"[gene_shift] {sdiag}", file=sys.stderr)     if alpha != 0.0:         pb_ref = pseudobulk(X)          # scorer's ref is an unbiased subsample of the last input         dp_pre = pseudobulk(Xpre) - pb_ref

调研来源?调研员查到并用到的知识条目和文献检索结果（只列标题和编号）。

用到的知识库条目

分析结果?分析员写的 ANALYSIS.json：改了什么、各组分数怎么变、假设是否成立、经验和下一步建议。

对话摘要?每个角色和大模型对话的统计：轮数、工具调用、用时、token 数和最后的回答摘录；原始记录只给路径。

只给统计和最后回答的摘录；完整对话请到原始记录位置里列出的文件看。

分析员

ANALYSIS.json written

工程师

Done. Final summary:

**做了什么**：按 PLAN 实现两阶段伪批量差方向位移块（EB 收缩 K=30、稀疏门控 ZMIN=1.5、仅双输入视图激活），并在筛选中新增 per-type 变体（按细胞型算 Δμ，低检出基因回退全局）。出厂配置：typed 模式 β=1.0；off control（VEC_SHIFT_BETA=0）与父节点 X3 输出逐字节一致，单输入的 proxy10 输出与父节点逐字节一致。

**查分结果**（A 半，seed 0，9/20 次）：global 模式全面失败（β=0.5/5/20 → 54.27/47.46/38.53 vs 父 54.51，PLAN 风险 #3 成真）；typed 模式 β=1/2/3 → 均 54.88（+0.37，噪声内），收益全在 mmd_u（0.0272→0.0259），PLAN 的目标组 de_recovery **未被抬升**（X3 de_score 仍在地板下）——METHOD.md 已如实报告。

**下一步最值得试**：X3 de_recovery 对两阶段数据驱动信号无响应，改用 prior/ 里的调控知识（CollecTRI TF 靶基因、Reactome 心脏通路）约束方向位移；或放弃 de_recovery、继续在 mmd_u/组成上挖（typed β=3 的 mmd 最好但 de_direction 掉，可试 β=1 + 更严门控 ZMIN=2.5）。

调研员

```json
{
  "name": "两阶段基因级EB收缩方向位移替代型内常数重锚定",
  "motivation": "父节点10最弱组de_recovery=49.58（X3 de_score=-0.052，skill 0.483，仍在地板下）；ANALYSIS明确指出型内重锚定d(t)=α(μ_in−μ_sel)是每型常数向量，无法产生真正的DE方向恢复；α已扫8点确认单峰，继续调参只有噪声级收益。proxy10 de_score也从0.137降至0.064。需要真正的基因级方向信号。",
  "approach": "在现有composition_trend加权抽样+α=3.0重锚定之后，新增一个可选的基因级方向位移块（仅当视图提供≥2个输入阶段时激活）：\n1. 计算每基因两阶段伪批量均值差 Δμ_g = mean(stage2) − mean(stage1)；\n2. EB收缩：估计基因间先验方差 s0²=median(Δμ²)，逐基因收缩权重 w_g = n_eff/(n_eff + K_SHRINK)，K_SHRINK初值30（搜索10/30/60/100）；Δμ_shrunk_g = Δμ_g · w_g；\n3. 稀疏门控：只保留 |Δμ_shrunk_g| > ZMIN × median(|Δμ_shrunk|) 的基因（ZMIN初值1.5，搜索1.0/1.5/2.0/2.5），其余基因不位移；\n4. 对每个重采样细胞，在保留的基因上施加 x[i,g] += BETA × Δμ_shrunk_g（仅原非零位置，clip≥0），BETA初值0.5（搜索0.3/0.5/0.8/1.0）；\n5. 位移幅度上限 MAX_SHIFT=1.0（防止极端基因主导）；\n6. 单输入视图（如proxy10）：该块不激活，行为与父节点逐字节一致（保留α=3.0重锚定）；\n7. vec-score筛选：先跑BETA=0确认与父节点一致（off control），再跑BETA=0.5看X3 de_score是否转正；若正则扫BETA和K_SHRINK；若X3 de_score无改善则报告机制无效。全程只查A半，最多6次查分。",
  "expected_groups": ["de_recovery", "direction"],
  "risks": "1) X3若只有一个输入阶段则机制不激活，输出与父节点相同——Engineer应在读入视图后立即打印输入阶段数确认；2) EB收缩过强（K过大）导致所有基因被压到接近零，等于没位移——检查实际被位移的基因数，若<100则降低K或ZMIN；3) 两阶段差异方向与真实E9.5→E10.5方向不一致（发育非线性），位移方向错误反而压低de_score——若BETA=0.5时X3 de_score比父节点更负，立即停止该方向；4) 30分钟时限内应无问题（伪批量计算O(n_cells×n_genes)很轻），但若基因面板过大需注意内存。",
  "family_id": "composition_program",
  "mechanism": "用两个输入阶段的伪批量均值差作为逐基因发育方向信号，经经验贝叶斯收缩和稀疏门控后叠加到重采样细胞上，替代原有的每型常数重锚定向量。",
  "vs_constant_shift": "原recenter是每型一个常向量（型内所有细胞同一位移，方向由抽样偏置决定，与真实发育方向无关）；本方案是逐基因、基于两阶段实际观测变化的方向位移，带EB收缩（小样本基因被压向零）和稀疏门控（只有统计支持的基因被位移），且仅在双输入视图激活。",
  "mechanism_evidence": "Engineer应报告：(1)实际被位移的基因数和细胞数；(2)位移向量的型内离散程度（应>0，若=0则退化为常数位移）；(3)X3和proxy10各自的de_score、de_direction、mmd_u、variogram四项变化；(4)被位移基因的Δμ符号与de_score排序的一致性（正相关说明方向正确）。",
  "mechanism_off_control": "设VEC_SHIFT_BETA=0：基因级位移块不执行任何修改，输出应与父节点（α=3.0，无基因级位移）逐字节一致。Engineer用h5数据比对验证。预期差别：关闭时X3 de_score保持-0.052，开启后应上升（若机制有效）。"
}
```

审查员

REVIEW.json written