Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261002T085943Z-deck-restructure-contrast-metrics-zeroshot-len/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/sec08_pairs_needed.py

@@ -0,0 +1,897 @@
+"""Section 8 of the 2026-09-30 paper-summary deck: how many bulk-sc pairs LoRA needs (learning curve).
+
+CPU only. Reads the pooled 5-fold scoring calls of the paired fine-tune plan and the per-fold RUN.json files,
+then draws two slide figures:
+  figures/sec08_pairs_realism.{png,pdf}  FULL: composition L1, coverage, selected step x (ImMicro, sepsis)
+  figures/sec08_pairs_signal.{png,pdf}   TALL: ImMicro NK fraction r, sepsis logFC r, sepsis top-50 DEG
+Every plotted/quoted number is read from a score file here and written with its source to data/sec08_*.csv.
+Model: final f01 6k 1x epoch-16 checkpoint + LoRA (checked against every RUN.json).
+x axis: "0" = no LoRA training. Two zero-shot points sit there, side by side: the pretrained model fed the real bulk
+as measured (score arm zs__raw; uses no pair) and fed gene-wise calibrated bulk (zs__cal). Calibration (back as a
+comparator, user decision 2026-10-01): alpha_g = mean CPM(pseudobulk_g) / mean CPM(bulk_g) over ALL paired training
+donors of the fold (train + inner-val; tools/prepare_dataset.py `fit = fold != f`), applied as bulk CPM x alpha, then
+renormalised. So zs__cal is not pair-free: its alpha uses the same 85-86 / 65-66 donors that LoRA "all" trains and
+early-stops on. LoRA n = training pairs, and every LoRA run also early-stops on 7-8 paired inner-val donors
+(RUN.json inner_val_rows). The LoRA arms are conditioned on the real bulk (*__raw) and share one real-bulk broadcast.
+Cap-limited = ran to max_steps before the patience window closed; its late inner-val loss drop is read from
+history.json.
+When a zero-shot point lies far outside every other value of a panel (BREAK_RATIO), that panel gets a y-axis
+break with a short segment at the same data scale, instead of squeezing the LoRA curve.
+
+Equal-budget calibration (optional, drawn only if present): zero-shot + calibration with alpha fitted on exactly the
+pairs a lora_n{5,10,20,40} run saw (n training + its inner-val pairs), per fold. The scores folder is scanned for
+summary.json files (scores/<dataset>/**/summary.json, outside arms/) whose arm names match CAL_N_RE (contain "cal"
+and an n<digits> token) and carry no fine-tune tag (FT_TAGS). An arm is drawn only from a scoring call over the
+same donors as the pooled call (same reference cells, so metrics are comparable). If found, it is drawn as a second
+line across n, ending at "all" with zs__cal (alpha from the same pairs LoRA "all" saw). If none is found, only the
+two x = 0 points are drawn and a note is printed.
+
+scSemiProfiler line (optional, drawn only if present; added 2026-10-02): scSemiProfiler trained with n pairs as
+representatives (arms scsemi_n{5,10,20,40}__raw in scores/<dataset>/scsemi_budget_20261002/pooled*/) and with all
+training pairs ("all" = arm scsemi__raw of the 5-fold CV run in scores/<dataset>/scsemi_cv_20261002/pooled*/), looked up
+by tools/scsemi_results.py and drawn only for calls over the same donors and reference cells as the pooled call. If
+nothing is found, a note is printed and the figures are unchanged. DECK_FIGDIR / DECK_OUTDATA redirect the outputs and
+DECK_SCSEMI_PLAN_ROOT the scSemiProfiler lookup (synthetic presence test).
+"""
+import os
+import json
+import re
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from scipy.stats import wilcoxon
+
+DECK = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/"
+            "2026-09-30--paper-summary--r00")
+sys.path.insert(0, str(DECK / "tools"))
+import style  # noqa: E402
+import scsemi_results as SR  # noqa: E402
+import fold_sizes as FZ  # noqa: E402  (2026-10-02: the "all" point is labelled with its training-pair count)
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.lines import Line2D  # noqa: E402
+
+PLAN = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/"
+            "paired_finetune_intradataset_20260927")
+FINAL_CKPT = ("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/runs/"
+              "f01_6k_1x_seed20260627/checkpoint_epoch_016.pt")
+FIGDIR = Path(os.environ.get("DECK_FIGDIR", str(DECK / "figures")))
+DATADIR = Path(os.environ.get("DECK_OUTDATA", str(DECK / "data")))
+FIGDIR.mkdir(exist_ok=True)
+DATADIR.mkdir(exist_ok=True)
+
+DSETS = ["immicro", "ye2026_sepsis"]
+DS_LABEL = {"immicro": "ImMicro", "ye2026_sepsis": "Sepsis"}
+LC_N = [5, 10, 20, 40]
+LORA_ARMS = [f"lora_n{n}__raw" for n in LC_N] + ["real_lora__raw"]
+RUN_ARMS = [f"lora_n{n}" for n in LC_N] + ["real_lora"]
+ZS = "zs__raw"                              # zero-shot on the real bulk as measured: 0 pairs
+ZS_CAL = "zs__cal"                          # zero-shot + calibration, alpha from all fold pairs (train + inner-val)
+# 2026-10-02 (user): x = 0 must use no paired data. The second x = 0 point is now the zero-shot model fed the real bulk
+# after exon-length correction (no pairs); zs__cal (alpha from every training pair of the fold) is only drawn where it
+# belongs, as the right end of the equal-budget calibration line (same pairs as the rightmost LoRA point).
+ZS_LEN = "zs__len"                          # zero-shot, exon-length-corrected real bulk (tools/build_tpm_zeroshot.py)
+LEN_SUB = "zeroshot_tpm_20261002/pooled"    # its scoring call (same donors and reference cells as scores/<ds>/pooled)
+X_LAB = ["0", "5", "10", "20", "40", "all"]
+KEY_LAB = ["zero_shot", "5", "10", "20", "40", "all"]  # the same points, as keys in sec08_slide_numbers.csv
+X_POS = [0.0, 1.3, 2.3, 3.3, 4.3, 5.55]   # small gap after the zero-shot points; extra room for the "78–79" tick
+DX0 = 0.2                                   # raw zero-shot left of x = 0, calibrated right of it
+X_RAW, X_CAL = X_POS[0] - DX0, X_POS[0] + DX0      # X_CAL: position of the second zero-shot point
+XLIM = (-0.5, X_POS[-1] + 0.5)              # same x range as the selected-step panel
+COL_RAW, COL_CAL = style.ARM["zs_raw"], style.ARM["zs_cal"]   # archived pair: pale = raw, #0072B2 = calibrated
+RAW_EDGE = "#5E8FB5"                        # edge of the pale raw marker (same as sec05 RAW_EDGE)
+COL_LEN = style.ARM["tpm"]                  # zero-shot, length-corrected bulk (sand)
+# equal-budget calibration arms: "cal" and an n<digits> token, and no fine-tune tag
+CAL_N_RE = re.compile(r"(?:^|[^a-z0-9])n(\d+)(?![0-9])", re.I)
+FT_TAGS = ("lora", "real", "pseudo", "full", "condenc", "multik", "recipe")
+BREAK_RATIO = 1.5  # y-axis break when the zero-shot point sits > 1.5 spans of the other values beyond them
+CELLTYPES = ["T cell", "B cell", "Monocyte", "NK cell"]
+DEG_TYPES = ["T cell", "B cell", "Monocyte", "NK cell", "DC"]
+RAW_ARMS = LORA_ARMS + ["zs__raw", "real_full__raw", "pseudo_full__raw"]  # all conditioned on raw real bulk
+
+REF = {  # reference-line styles; pseudo/broadcast copied from sec09 REFSTYLE so the deck is consistent
+    "pseudo": dict(ls=(0, (4, 2)), color=style.ARM["pseudo"], lw=1.6, label="Zero-shot from sc pseudobulk"),
+    "broadcast": dict(ls=(0, (1, 1.2)), color=style.ARM["broadcast"], lw=2.4, label="Real-bulk broadcast"),
+    "null": dict(ls=(0, (6, 2, 1.5, 2)), color=style.GREY, lw=1.4, label="Cohort-mean null"),
+}
+LEG_ZS = "Zero-shot, real bulk as measured (no pairs)"
+LEG_LEN = "Zero-shot, exon-length-corrected bulk (no pairs)"
+LEG_CAL_N = "Zero-shot + calibration, α from the same pairs"
+COL_SC = style.ARM["scsemi"]
+LEG_SC = "scSemiProfiler, n pairs as representatives"
+
+
+def xlabel(st):
+    """x-axis label; the inner-val range is read from the RUN.json files."""
+    return (f"LoRA training pairs per fold (each run +{rng_txt(st.n_inner_val)} inner-val pairs for early stopping; "
+            f"test fold never trained on)")
+
+
+def xticks(ds):
+    """x tick labels of one cohort: the rightmost point is every training donor of the fold, shown as its count."""
+    return X_LAB[:-1] + [FZ.all_label(ds)]
+
+
+def close(a, b, tol=1e-6):
+    return abs(float(a) - float(b)) <= tol
+
+
+class Call:
+    """One pooled score_ft.py scoring call."""
+
+    def __init__(self, path: Path):
+        self.path = path
+        self.summary = json.loads((path / "summary.json").read_text())
+        self.contrast = pd.read_csv(path / "contrast.csv") if (path / "contrast.csv").exists() else None
+        self.per_donor = pd.read_csv(path / "per_donor.csv")
+        self.src_summary = str(path / "summary.json")
+        self.src_contrast = str(path / "contrast.csv")
+        self.src_per_donor = str(path / "per_donor.csv")
+
+    def head(self, arm, key):
+        return float(self.summary["headline"][arm][key])
+
+    def contrast_r(self, arm, col="r_generated"):
+        c = self.contrast
+        c = c[(c.arm == arm) & (c.statistic == "mean_cp10k") & (c.celltype.isin(CELLTYPES))]
+        assert set(c.celltype) == set(CELLTYPES), (self.path, arm)
+        return float(c[col].mean())
+
+    def deg(self, arm, col="top50_overlap"):
+        d = self.contrast
+        d = d[(d.arm == arm) & (d.statistic == "welch_top50_deg")]
+        assert len(d) == 5, (self.path, arm, len(d))  # T/B/Mono/NK/DC
+        return float(d[col].sum())
+
+    def deg_type(self, arm, ct, col="top50_overlap"):
+        d = self.contrast
+        d = d[(d.arm == arm) & (d.statistic == "welch_top50_deg") & (d.celltype == ct)]
+        assert len(d) == 1, (self.path, arm, ct, len(d))
+        return float(d[col].iloc[0])
+
+    def frac_r(self, arm, ct):
+        d = self.per_donor[self.per_donor.arm == arm]
+        return float(np.corrcoef(d[f"gen_{ct}"].astype(float), d[f"obs_{ct}"].astype(float))[0, 1])
+
+
+def load():
+    calls = {ds: Call(PLAN / "scores" / ds / "pooled") for ds in DSETS}
+    deg_call = Call(PLAN / "scores" / "ye2026_sepsis" / "pooled_deg")
+    for ds, c in calls.items():
+        for a in LORA_ARMS + [ZS, ZS_CAL, "zs__pseudo"]:
+            assert a in c.summary["arms"], (ds, a)
+            assert c.summary["per_arm"][a].get("receipt_verified", True) is not False, (ds, a)
+            # headline composition L1 equals the donor mean of per_donor.csv
+            assert close(c.per_donor.loc[c.per_donor.arm == a, "comp_l1"].mean(), c.head(a, "comp_l1"), 1e-9), a
+    # headline contrast means equal the per-type mean of contrast.csv
+    se = calls["ye2026_sepsis"]
+    for a in LORA_ARMS + [ZS, ZS_CAL, "zs__pseudo"]:
+        assert close(se.contrast_r(a), se.head(a, "contrast_mean_cp10k_r_gen"), 1e-9), a
+        assert close(deg_call.deg(a), deg_call.head(a, "deg_top50_sum"), 1e-9), a
+    # broadcast is the same for every arm conditioned on raw real bulk
+    bc = {a: se.contrast_r(a, "r_broadcast") for a in RAW_ARMS}
+    assert max(bc.values()) - min(bc.values()) < 1e-12, bc
+    bcd = {a: deg_call.deg(a, "top50_overlap_broadcast") for a in RAW_ARMS}
+    assert len(set(bcd.values())) == 1, bcd
+    assert close(bcd["real_lora__raw"], deg_call.head("real_lora__raw", "deg_top50_sum_broadcast"), 1e-9)
+    # non-DEG metrics agree between the two sepsis scoring calls (k-NN ties allow 1e-3)
+    for a in LORA_ARMS + [ZS, ZS_CAL]:
+        for k in ("comp_l1", "coverage"):
+            assert abs(se.head(a, k) - deg_call.head(a, k)) <= 1e-3, (a, k)
+    lens = {ds: Call(PLAN / "scores" / ds / LEN_SUB) for ds in DSETS}
+    for ds, lc in lens.items():
+        c = calls[ds]
+        assert ZS_LEN in lc.summary["arms"] and ZS in lc.summary["arms"], (ds, lc.summary["arms"])
+        assert int(lc.summary["n_donors"]) == int(c.summary["n_donors"]), ds
+        assert lc.summary["reference"]["reference_sha256"] == c.summary["reference"]["reference_sha256"], ds
+        for k in ("comp_l1", "coverage", "precision"):   # the raw arm reproduces in the new call (k-NN ties: 1e-3)
+            assert abs(lc.head(ZS, k) - c.head(ZS, k)) <= 1e-3, (ds, k, lc.head(ZS, k), c.head(ZS, k))
+    assert close(lens["ye2026_sepsis"].contrast_r(ZS), se.contrast_r(ZS), 1e-9)
+    assert close(lens["ye2026_sepsis"].deg(ZS), deg_call.deg(ZS), 1e-9), (lens["ye2026_sepsis"].deg(ZS), deg_call.deg(ZS))
+    return calls, deg_call, lens
+
+
+def find_cal_budget(calls, deg_call):
+    """Scan the scores folder for equal-budget calibration arms (see module docstring).
+
+    Returns (budget, log): budget[ds][n] = dict(arm, calls=[Call, ...]) for drawable arms (pooled call first, then
+    other calls over the same donors; the sepsis DEG call first for DEG); log = one dict per matching arm found."""
+    budget, log = {ds: {} for ds in DSETS}, []
+    cache = {c.path: c for c in list(calls.values()) + [deg_call]}
+    for ds in DSETS:
+        root = PLAN / "scores" / ds
+        pooled = calls[ds]
+        donors = set(pooled.per_donor.donor)
+        cands = {}
+        for sp in sorted(root.rglob("summary.json")):
+            rel = sp.relative_to(root)
+            if "arms" in rel.parts:
+                continue
+            try:
+                arms = json.loads(sp.read_text()).get("arms") or []
+            except (OSError, ValueError):
+                continue
+            for a in arms:
+                m = CAL_N_RE.search(a)
+                if "cal" not in a.lower() or m is None or any(t in a.lower() for t in FT_TAGS):
+                    continue
+                n = int(m.group(1))
+                if sp.parent not in cache:
+                    try:
+                        cache[sp.parent] = Call(sp.parent)
+                    except (OSError, ValueError) as e:
+                        log.append(dict(dataset=ds, call=str(rel.parent), arm=a, n=n, drawn=False,
+                                        note=f"not drawn: cannot read the call ({e.__class__.__name__})"))
+                        continue
+                call = cache[sp.parent]
+                same = set(call.per_donor.loc[call.per_donor.arm == a, "donor"]) == donors
+                ok = same and n in LC_N
+                why = ("drawn" if ok else "not drawn: " + ("n not in " + str(LC_N) if same else
+                                                          "call does not score the pooled donor set"))
+                log.append(dict(dataset=ds, call=str(rel.parent), arm=a, n=n, drawn=ok, note=why))
+                if ok:
+                    cands.setdefault(n, []).append((a, call))
+        for n, lst in cands.items():
+            names = sorted({a for a, _ in lst})
+            if len(names) > 1:
+                pref = sorted({a for a, c in lst if c is pooled})
+                if len(pref) != 1:
+                    sys.exit(f"{ds}: several equal-budget calibration arms for n = {n}: {names}; pick one")
+                names = pref
+            arm = names[0]
+            cl = [c for a, c in lst if a == arm]
+            cl.sort(key=lambda c: (c is not pooled, c is not deg_call, str(c.path)))
+            for c in cl:  # a call other than the pooled one must reproduce shared arms (k-NN ties allow 1e-3)
+                if c is pooled:
+                    continue
+                for shared in (ZS_CAL, "real_lora__raw"):
+                    if shared in c.summary["arms"]:
+                        for k in ("comp_l1", "coverage"):
+                            assert abs(c.head(shared, k) - pooled.head(shared, k)) <= 1e-3, (c.path, shared, k)
+            budget[ds][n] = dict(arm=arm, calls=cl)
+    return budget, log
+
+
+def find_scsemi(calls):
+    """scSemiProfiler at n = 5..40 (equal-budget arms) and "all" (5-fold CV arm), see module docstring.
+
+    Returns (scs, log): scs[ds][pairs label] = dict(arm, calls=[Call, ...]) for accepted calls; log = one dict per
+    arm found (accepted or not)."""
+    scs, log = {ds: {} for ds in DSETS}, []
+    for ds in DSETS:
+        ref = calls[ds].summary
+        found = [(str(n), a, d) for n, lst in sorted(SR.budget_calls(ds).items()) for a, d in lst]
+        found += [("all", SR.CV_ARM, d) for d in SR.cv_calls(ds)]
+        for lab, a, d in found:
+            try:
+                c = Call(d)
+            except (OSError, ValueError) as e:
+                log.append(dict(dataset=ds, call=str(d), arm=a, pairs=lab, drawn=False,
+                                note=f"not drawn: cannot read the call ({e.__class__.__name__})"))
+                continue
+            why = SR.accept(c.summary, ref, a)
+            if lab not in X_LAB[1:]:
+                why.append(f"n {lab} not in {X_LAB[1:]}")
+            ent = scs[ds].get(lab)
+            if ent is not None and ent["arm"] != a:
+                why.append(f"a different arm ({ent['arm']}) already gives n = {lab}")
+            log.append(dict(dataset=ds, call=str(d), arm=a, pairs=lab, drawn=not why,
+                            note="drawn" if not why else "not drawn: " + "; ".join(why)))
+            if not why:
+                scs[ds].setdefault(lab, dict(arm=a, calls=[]))["calls"].append(c)
+    return scs, log
+
+
+def scsemi_series(scs, ds, getter):
+    """[(pairs label, value, source)] of the scSemiProfiler line for one panel ([] when nothing is found)."""
+    out = []
+    for lab in X_LAB[1:]:
+        if lab in scs[ds]:
+            v, c = budget_value(scs[ds][lab], getter)
+            if v is not None:
+                out.append((lab, v, f"{c.path} arm {scs[ds][lab]['arm']}"))
+    return out
+
+
+def budget_value(entry, getter):
+    """First call of a budget entry that provides the metric (None if no call does)."""
+    for c in entry["calls"]:
+        try:
+            return getter(c, entry["arm"]), c
+        except (AssertionError, KeyError, TypeError, AttributeError):
+            continue
+    return None, None
+
+
+def load_steps():
+    rows = []
+    for ds in DSETS:
+        for k in range(5):
+            for arm in RUN_ARMS:
+                p = PLAN / "runs" / ds / f"fold{k}" / arm / "RUN.json"
+                r = json.loads(p.read_text())
+                assert r["source_checkpoint"] == FINAL_CKPT, (p, r["source_checkpoint"])
+                assert r["mode"] == "lora", p
+                assert r["inner_val_rows"], p  # every LoRA run early-stops on paired inner-val donors
+                hist = json.loads((p.parent / "history.json").read_text())
+                h = {int(e["step"]): float(e["inner_val_loss"]) for e in hist}
+                # cap-limited = ran to max_steps without the patience window closing
+                cap_limited = (r["steps_run"] >= r["max_steps"]
+                               and r["steps_run"] - r["selected_step"] < r["patience"] * r["eval_every"])
+                last = max(h)
+                win = r["patience"] * r["eval_every"]  # last patience window (2,000 steps)
+                drop_total = h[0] - h[last]
+                drop_last = h[last - win] - h[last]
+                rows.append({"dataset": ds, "fold": k, "arm": arm, "n_train": len(r["train_rows"]),
+                             "n_inner_val": len(r["inner_val_rows"]), "selected_step": r["selected_step"],
+                             "max_steps": r["max_steps"], "steps_run": r["steps_run"],
+                             "patience_steps": win, "cap_limited": bool(cap_limited),
+                             "val_loss_step0": h[0], "val_loss_last": h[last],
+                             "val_loss_last_window_drop_frac": drop_last / drop_total,
+                             "wall_seconds": r.get("wall_seconds"), "source": str(p),
+                             "history": str(p.parent / "history.json")})
+    st = pd.DataFrame(rows)
+    # nested subsets: n_train equals the arm size for lora_n*
+    for n in LC_N:
+        assert (st[st.arm == f"lora_n{n}"].n_train == n).all(), n
+    return st
+
+
+def load_fold_sizes():
+    """Test donors per fold (PREPARE.json); their sum must equal the donors in the pooled scoring call.
+    alpha = paired donors the zs__cal alpha is fitted on per fold (train + inner-val)."""
+    out = {}
+    for ds in DSETS:
+        p = PLAN / "data" / ds / "PREPARE.json"
+        fs = json.loads(p.read_text())["fold_sizes"]
+        out[ds] = dict(test=[int(v["test"]) for v in fs.values()],
+                       alpha=[int(v["train"]) + int(v["inner_val"]) for v in fs.values()], source=str(p))
+    return out
+
+
+def load_fewshot():
+    """Few-pair call (5 draws): zero-shot + calibration with alpha from the same 5 pairs vs LoRA on those 5 pairs
+    (fixed 3k steps, no inner-val). Its own donor set, so it is only quoted in the notes, never drawn."""
+    out = {}
+    for ds in DSETS:
+        p = PLAN / "scores" / ds / "fewshot_n5" / "aggregate.json"
+        a = json.loads(p.read_text())["arms"]
+        out[ds] = dict(zs_cal=float(a["zs_cal"]["coverage"][0]), lora5=float(a["lora5"]["coverage"][0]),
+                       source=str(p))
+    return out
+
+
+def rng_txt(v):
+    v = sorted(int(x) for x in v)
+    return f"{v[0]}" if v[0] == v[-1] else f"{v[0]}–{v[-1]}"
+
+
+def break_axis(ax, v0, lo, hi, x_txt=X_RAW, extra=()):
+    """Shrink ax to lo..hi and add a short segment above it that shows v0 at the same data scale.
+    extra: further break values (2026-10-02, two zero-shot points at x = 0). When they do not fit at the main scale,
+    the segment gets its own scale and a tick at each value."""
+    p = ax.get_position()
+    h_top, gap = 0.22 * p.height, 0.08 * p.height
+    h_main = p.height - h_top - gap
+    ax.set_position([p.x0, p.y0, p.width, h_main])
+    ax.set_ylim(lo, hi)
+    half = (hi - lo) * h_top / h_main / 2
+    top = ax.figure.add_axes([p.x0, p.y0 + h_main + gap, p.width, h_top], sharex=ax)
+    vals = [v0] + list(extra)
+    if max(vals) - min(vals) > 1.6 * half:      # own scale: ticks at the values
+        spread = max(vals) - min(vals)
+        top.set_ylim(min(vals) - 0.35 * spread, max(vals) + 0.35 * spread)
+        top.set_yticks(sorted(round(v, 2) for v in vals))
+        top.spines["bottom"].set_visible(False)
+        top.tick_params(axis="x", bottom=False, labelbottom=False)
+        mk = dict(marker=[(-1, -0.5), (1, 0.5)], markersize=10, ls="none", color=style.INK, mec=style.INK, mew=1.2,
+                  clip_on=False)
+        top.plot([0], [0], transform=top.transAxes, **mk)
+        ax.plot([0], [1], transform=ax.transAxes, **mk)
+        return top
+    v0 = (max(vals) + min(vals)) / 2
+    top.set_ylim(v0 - half, v0 + half)
+    # one tick in the short segment: the nearest value on the main axis' tick grid that fits inside it
+    step = float(np.diff(ax.get_yticks()[:2])[0]) / 2
+    tick = round(round(v0 / step) * step, 10)
+    top.set_yticks([tick] if abs(tick - v0) < half else [])
+    if not len(top.get_yticks()):  # no grid value fits: print the value next to the point instead
+        top.text(x_txt + 0.3, v0, f"{v0:.2f}", va="center", ha="left", fontsize=12)
+    top.spines["bottom"].set_visible(False)
+    top.tick_params(axis="x", bottom=False, labelbottom=False)
+    mk = dict(marker=[(-1, -0.5), (1, 0.5)], markersize=10, ls="none", color=style.INK, mec=style.INK, mew=1.2,
+              clip_on=False)
+    top.plot([0], [0], transform=top.transAxes, **mk)
+    ax.plot([0], [1], transform=ax.transAxes, **mk)
+    return top
+
+
+def curve(ax, z0, vl, refs, rows, panel, ds, srcl, ylim0=None, vb=None, vs=None):
+    """Zero-shot points at x = 0 (z0: [(series, value, source)] for raw and calibrated), LoRA curve at 5..all,
+    optional equal-budget calibration line (vb: [(pairs label, value, source)]), reference lines.
+    Returns the axis that carries the panel title."""
+    vb = vb or []
+    vs = vs or []   # scSemiProfiler line (n = 5..all)
+    lo, hi = min(vl), max(vl)
+    for _, val, _ in list(refs) + [(None, v, None) for _, v, _ in vb + vs]:
+        lo, hi = min(lo, val), max(hi, val)
+    span = hi - lo
+    # a zero-shot point goes to the break segment when it sits > BREAK_RATIO spans beyond everything else
+    out = {ser for ser, v, _ in z0
+           if ylim0 is None and span > 0 and (v - hi > BREAK_RATIO * span or lo - v > BREAK_RATIO * span)}
+    for ser, v, _ in z0:
+        if ser in out:
+            assert v > hi, (panel, ds, ser, "axis break below the curve is not implemented")
+        else:
+            lo, hi = min(lo, v), max(hi, v)
+    broken = bool(out)
+    pad = (hi - lo) * 0.15
+    axes = [ax]
+    xz = {"zero_shot": X_RAW, "zs_len": X_CAL}
+    if broken:
+        zb = [z for z in z0 if z[0] in out]
+        top = break_axis(ax, zb[0][1], lo - pad, hi + pad, xz[zb[0][0]], extra=[v for _, v, _ in zb[1:]])
+        axes.append(top)
+    else:
+        ax.set_ylim(ylim0 if ylim0 is not None else lo - pad, hi + pad)
+    for ser, v, src in z0:
+        a = axes[-1] if ser in out else ax
+        if ser == "zero_shot":
+            a.scatter([X_RAW], [v], s=70, color=COL_RAW, edgecolor=RAW_EDGE, lw=1.0, zorder=5, clip_on=False)
+        else:
+            a.scatter([X_CAL], [v], s=70, color=COL_LEN, edgecolor=style.INK, lw=0.6, zorder=5, clip_on=False)
+        rows.append(dict(panel=panel, dataset=ds, series=ser, pairs=X_LAB[0], value=v, source=src,
+                         y_axis_break=bool(broken)))
+    if vb:
+        xb = [X_POS[X_LAB.index(lab)] for lab, _, _ in vb]
+        ax.plot(xb, [v for _, v, _ in vb], color=COL_CAL, lw=1.8, zorder=3)
+        ax.scatter(xb, [v for _, v, _ in vb], s=60, marker="s", color=COL_CAL, edgecolor=style.INK, lw=0.6,
+                   zorder=5)
+        for lab, v, src in vb:
+            rows.append(dict(panel=panel, dataset=ds, series="zs_cal_budget", pairs=lab, value=v, source=src,
+                             y_axis_break=bool(broken)))
+    if vs:
+        xs = [X_POS[X_LAB.index(lab)] for lab, _, _ in vs]
+        ax.plot(xs, [v for _, v, _ in vs], color=COL_SC, lw=1.8, zorder=3)
+        ax.scatter(xs, [v for _, v, _ in vs], s=55, marker="D", color=COL_SC, edgecolor=style.INK, lw=0.6, zorder=5)
+        for lab, v, src in vs:
+            rows.append(dict(panel=panel, dataset=ds, series="scsemi", pairs=lab, value=v, source=src,
+                             y_axis_break=bool(broken)))
+    ax.plot(X_POS[1:], vl, color=style.ARM["lora"], lw=2.0, zorder=3)
+    ax.scatter(X_POS[1:], vl, s=70, color=style.ARM["lora"], edgecolor=style.INK, lw=0.6, zorder=5)
+    for lab, v, s in zip(X_LAB[1:], vl, srcl):
+        rows.append(dict(panel=panel, dataset=ds, series="lora", pairs=lab, value=v, source=s,
+                         y_axis_break=bool(broken)))
+    for key, val, src in refs:
+        r = REF[key]
+        for a in axes:
+            a.axhline(val, ls=r["ls"], color=r["color"], lw=r["lw"], zorder=2)
+        rows.append(dict(panel=panel, dataset=ds, series=f"ref_{key}", pairs="", value=val, source=src,
+                         y_axis_break=bool(broken)))
+    ax.set_xlim(*XLIM)
+    ax.set_xticks(X_POS)
+    ax.set_xticklabels(xticks(ds))
+    # nothing is clipped: every plotted value is inside the y range of the axis that draws it
+    for v in list(vl) + [val for _, val, _ in refs] + [v for _, v, _ in vb + vs] + [v for ser, v, _ in z0
+                                                                                if ser not in out]:
+        assert ax.get_ylim()[0] <= v <= ax.get_ylim()[1], (panel, ds, v, ax.get_ylim())
+    for ser, v, _ in z0:
+        if ser in out:
+            assert axes[-1].get_ylim()[0] <= v <= axes[-1].get_ylim()[1], (panel, ds, ser, v)
+    return axes[-1]
+
+
+def budget_series(budget, ds, getter, cal_all, src_all):
+    """Equal-budget calibration line for one panel: [(pairs label, value, source)] at n = 5..40 (where present)
+    plus "all" = zs__cal (alpha from the same train + inner-val pairs LoRA "all" saw). [] when nothing is found."""
+    out = []
+    for n in LC_N:
+        if n not in budget[ds]:
+            continue
+        v, c = budget_value(budget[ds][n], getter)
+        if v is not None:
+            out.append((str(n), v, f"{c.path} arm {budget[ds][n]['arm']}"))
+    if out:
+        out.append(("all", cal_all, src_all + " arm zs__cal (alpha from every training pair of the fold = the rightmost LoRA point's pairs)"))
+    return out
+
+
+def fig_realism(calls, lens, st, fold_sizes, rows, steprows, budget, scs):
+    has_budget = any(budget[ds] for ds in DSETS)
+    has_sc = any(scs[ds] for ds in DSETS)
+    n_leg = 6 + has_budget + has_sc
+    fig, axs = plt.subplots(2, 3, figsize=style.FULL,  # a third legend row when an optional line is drawn
+                            gridspec_kw=dict(left=0.168, right=0.99, top=0.79 if n_leg > 6 else 0.83,
+                                             bottom=0.12, hspace=0.36, wspace=0.30))
+    heads = ["Composition L1 ↓", "Coverage ↑", "Selected step (×1,000)"]
+    pos0 = {ri: axs[ri, 2].get_position() for ri in range(len(DSETS))}  # row extents before any y-axis break
+    title_ax = {ci: axs[0, ci] for ci in range(3)}
+    for ri, ds in enumerate(DSETS):
+        c = calls[ds]
+        for ci, key in enumerate(["comp_l1", "coverage"]):
+            ax = axs[ri, ci]
+            refs = [("pseudo", c.head("zs__pseudo", key), c.src_summary)]
+            if key == "comp_l1":
+                refs.append(("null", c.head("real_lora__raw", "comp_l1_obs_loo_null"), c.src_summary))
+            z0 = [("zero_shot", c.head(ZS, key), c.src_summary),
+                  ("zs_len", lens[ds].head(ZS_LEN, key), lens[ds].src_summary)]
+            vb = budget_series(budget, ds, lambda cc, a, k=key: cc.head(a, k), c.head(ZS_CAL, key), c.src_summary)
+            vs = scsemi_series(scs, ds, lambda cc, a, k=key: cc.head(a, k))
+            tax = curve(ax, z0, [c.head(a, key) for a in LORA_ARMS], refs, rows,
+                        f"{key}", ds, [c.src_summary] * 5, vb=vb, vs=vs)
+            if ri == 0:
+                title_ax[ci] = tax
+        # selected early-stopping step per fold
+        ax = axs[ri, 2]
+        s = st[st.dataset == ds]
+        for xi, arm in zip(X_POS[1:], RUN_ARMS):
+            f = s[s.arm == arm].sort_values("fold")
+            v = f.selected_step.astype(float).values / 1000
+            jit = np.linspace(-0.22, 0.22, len(v))
+            ax.scatter(xi + jit, v, s=36, color=style.ARM["lora"], edgecolor=style.INK, lw=0.5, zorder=4)
+            med = float(np.median(v))
+            ax.plot([xi - 0.36, xi + 0.36], [med, med], color=style.INK, lw=2.0, zorder=5)
+            steprows.append(dict(dataset=ds, arm=arm, pairs=X_LAB[X_POS.index(xi)], median_step=med * 1000,
+                                 min_step=v.min() * 1000, max_step=v.max() * 1000,
+                                 n_cap_limited=int(f.cap_limited.sum()),
+                                 max_last_window_drop_frac=(float(f[f.cap_limited].val_loss_last_window_drop_frac.max())
+                                                            if f.cap_limited.any() else np.nan),
+                                 n_train=rng_txt(f.n_train), n_inner_val=rng_txt(f.n_inner_val),
+                                 source=";".join(f.source)))
+        cap = float(s.max_steps.max()) / 1000
+        ax.axhline(cap, color=style.GREY, lw=1.0, ls=(0, (2, 2)), zorder=1)
+        ax.text(X_POS[0] - 0.4, cap + 0.6, f"cap {cap:g}k", color=style.GREY, fontsize=12, va="bottom")
+        ax.text(X_POS[0], cap * 0.45, "no training", color=style.GREY, fontsize=12, ha="center",
+                va="center", rotation=90)
+        ax.set_ylim(0, cap * 1.18)
+        ax.set_xlim(-0.5, X_POS[-1] + 0.5)
+        ax.set_xticks(X_POS)
+        ax.set_xticklabels(xticks(ds))
+        n_all = rng_txt(s[s.arm == "real_lora"].n_train)
+        n_iv = rng_txt(s.n_inner_val)
+        n_scored = int(c.summary["n_donors"])
+        assert n_scored == sum(fold_sizes[ds]["test"]), (ds, n_scored, fold_sizes[ds]["test"])
+        n_alpha = rng_txt(fold_sizes[ds]["alpha"])
+        # zs__cal alpha uses the same donors LoRA "all" trains and early-stops on (train + inner-val)
+        lora_all_seen = sorted(int(a) + int(b) for a, b in zip(s[s.arm == "real_lora"].sort_values("fold").n_train,
+                                                               s[s.arm == "real_lora"].sort_values("fold").n_inner_val))
+        assert lora_all_seen == sorted(fold_sizes[ds]["alpha"]), (ds, lora_all_seen, fold_sizes[ds]["alpha"])
+        ymid = (pos0[ri].y0 + pos0[ri].y1) / 2
+        fig.text(0.005, ymid + 0.075, DS_LABEL[ds], fontsize=15, fontweight="bold", va="bottom", ha="left")
+        fig.text(0.005, ymid + 0.065,
+                 f"{n_scored} donors\ninner-val = {n_iv}\ntest fold\nnever trained on",
+                 fontsize=12.5, va="top", ha="left", linespacing=1.35)
+        rows.append(dict(panel="meta", dataset=ds, series="n_scored_donors", pairs="", value=n_scored,
+                         source=c.src_summary))
+        rows.append(dict(panel="meta", dataset=ds, series="n_pairs_all", pairs="all", value=n_all,
+                         source="RUN.json train_rows: " + ";".join(s[s.arm == "real_lora"].source)))
+        rows.append(dict(panel="meta", dataset=ds, series="n_inner_val", pairs="", value=n_iv,
+                         source="RUN.json inner_val_rows (all LoRA arms, 5 folds)"))
+        rows.append(dict(panel="meta", dataset=ds, series="n_test_donors_per_fold", pairs="",
+                         value="/".join(map(str, fold_sizes[ds]["test"])),
+                         source=fold_sizes[ds]["source"] + " fold_sizes test"))
+        rows.append(dict(panel="meta", dataset=ds, series="n_alpha_fit_pairs", pairs="0", value=n_alpha,
+                         source=fold_sizes[ds]["source"] + " fold_sizes train+inner_val; prepare_dataset.py "
+                                "fit = fold != f"))
+    for ci, h in enumerate(heads):
+        title_ax[ci].set_title(h, fontsize=14, pad=8)
+    x0, x1 = axs[1, 0].get_position().x0, axs[1, 2].get_position().x1
+    fig.text((x0 + x1) / 2, 0.005, xlabel(st), ha="center", va="bottom", fontsize=14)
+    handles = [  # column-wise fill: [raw, cal, (equal-budget cal)] | [LoRA, (scSemi), median] | [pseudo, null]
+        h_raw(), h_len()] + ([h_budget()] if has_budget else []) + ([h_scsemi()] if has_sc else []) + [
+        Line2D([0], [0], marker="o", color=style.ARM["lora"], lw=2.0, markeredgecolor=style.INK, markersize=8,
+               label="LoRA fine-tune (dots in step panel = folds)"),
+        Line2D([0], [0], color=style.INK, lw=2.0, label="Median of 5 folds"),
+        Line2D([0], [0], ls=REF["pseudo"]["ls"], color=REF["pseudo"]["color"], lw=REF["pseudo"]["lw"],
+               label=REF["pseudo"]["label"]),
+        Line2D([0], [0], ls=REF["null"]["ls"], color=REF["null"]["color"], lw=REF["null"]["lw"],
+               label=REF["null"]["label"]),
+    ]
+    fig.legend(handles=handles, loc="upper left", bbox_to_anchor=(0.005, 1.0), ncol=3, frameon=False,
+               fontsize=12, handlelength=2.4, columnspacing=1.6, labelspacing=0.4, borderaxespad=0.0)
+    style.save(fig, "sec08_pairs_realism", FIGDIR)
+
+
+def h_raw():
+    return Line2D([0], [0], marker="o", color=COL_RAW, lw=0, markeredgecolor=RAW_EDGE, markersize=8, label=LEG_ZS)
+
+
+def h_len():
+    return Line2D([0], [0], marker="o", color=COL_LEN, lw=0, markeredgecolor=style.INK, markersize=8, label=LEG_LEN)
+
+
+def h_scsemi():
+    return Line2D([0], [0], marker="D", color=COL_SC, lw=1.8, markeredgecolor=style.INK, markersize=7, label=LEG_SC)
+
+
+def h_budget():
+    return Line2D([0], [0], marker="s", color=COL_CAL, lw=1.8, markeredgecolor=style.INK, markersize=7,
+                  label=LEG_CAL_N)
+
+
+def fig_signal(calls, deg_call, lens, st, rows, budget, scs):
+    has_budget = any(budget[ds] for ds in DSETS)
+    has_sc = any(scs[ds] for ds in DSETS)
+    leg_rows = -(-(5 + has_budget + has_sc) // 2)
+    fig, axs = plt.subplots(1, 3, figsize=style.TALL,
+                            gridspec_kw=dict(left=0.085, right=0.99, top=0.755 - 0.055 * (leg_rows - 3), bottom=0.115,
+                                             wspace=0.42))  # 3 legend rows (4 with both optional lines)
+    im, se = calls["immicro"], calls["ye2026_sepsis"]
+    iml, sel = lens["immicro"], lens["ye2026_sepsis"]
+    tax = [None] * 3
+    # a: ImMicro across-donor NK fraction r
+    tax[0] = curve(axs[0], [("zero_shot", im.frac_r(ZS, "NK cell"), im.src_per_donor),
+                            ("zs_len", iml.frac_r(ZS_LEN, "NK cell"), iml.src_per_donor)],
+                   [im.frac_r(a, "NK cell") for a in LORA_ARMS],
+                   [("pseudo", im.frac_r("zs__pseudo", "NK cell"), im.src_per_donor)], rows, "nk_fraction_r",
+                   "immicro", [im.src_per_donor] * 5,
+                   vb=budget_series(budget, "immicro", lambda cc, a: cc.frac_r(a, "NK cell"),
+                                    im.frac_r(ZS_CAL, "NK cell"), im.src_per_donor),
+                   vs=scsemi_series(scs, "immicro", lambda cc, a: cc.frac_r(a, "NK cell")))
+    # other cell types, recorded for the footnote ("0+cal" = zero-shot + calibration)
+    for ct in ["T cell", "B cell", "Monocyte"]:
+        for a, lab in zip([ZS, ZS_CAL] + LORA_ARMS, ["0", "0+cal"] + X_LAB[1:]):
+            rows.append(dict(panel="other_fraction_r", dataset="immicro", series=ct, pairs=lab,
+                             value=im.frac_r(a, ct), source=im.src_per_donor))
+    # b: sepsis disease logFC r (mean of T/B/Mono/NK, mean_cp10k)
+    tax[1] = curve(axs[1], [("zero_shot", se.contrast_r(ZS), se.src_contrast),
+                            ("zs_len", sel.contrast_r(ZS_LEN), sel.src_contrast)],
+                   [se.contrast_r(a) for a in LORA_ARMS],
+                   [("pseudo", se.contrast_r("zs__pseudo"), se.src_contrast),
+                    ("broadcast", se.contrast_r("real_lora__raw", "r_broadcast"), se.src_contrast)],
+                   rows, "logfc_r", "ye2026_sepsis", [se.src_contrast] * 5,
+                   vb=budget_series(budget, "ye2026_sepsis", lambda cc, a: cc.contrast_r(a),
+                                    se.contrast_r(ZS_CAL), se.src_contrast),
+                   vs=scsemi_series(scs, "ye2026_sepsis", lambda cc, a: cc.contrast_r(a)))
+    # c: sepsis top-50 DEG (Welch, sum over T/B/Mono/NK/DC)
+    tax[2] = curve(axs[2], [("zero_shot", deg_call.deg(ZS), deg_call.src_contrast),
+                            ("zs_len", sel.deg(ZS_LEN), sel.src_contrast)],
+                   [deg_call.deg(a) for a in LORA_ARMS],
+                   [("pseudo", deg_call.deg("zs__pseudo"), deg_call.src_contrast),
+                    ("broadcast", deg_call.deg("real_lora__raw", "top50_overlap_broadcast"), deg_call.src_contrast)],
+                   rows, "top50_deg", "ye2026_sepsis", [deg_call.src_contrast] * 5, ylim0=0,
+                   vb=budget_series(budget, "ye2026_sepsis", lambda cc, a: cc.deg(a),
+                                    deg_call.deg(ZS_CAL), deg_call.src_contrast),
+                   vs=scsemi_series(scs, "ye2026_sepsis", lambda cc, a: cc.deg(a)))
+    # top-50 DEG overlap per cell type (footnote: which types carry the gain), incl. broadcast
+    for ct in DEG_TYPES:
+        for a, lab in zip([ZS, ZS_CAL] + LORA_ARMS + ["zs__pseudo"], ["0", "0+cal"] + X_LAB[1:] + ["pseudo"]):
+            rows.append(dict(panel="top50_deg_by_type", dataset="ye2026_sepsis", series=ct, pairs=lab,
+                             value=deg_call.deg_type(a, ct), source=deg_call.src_contrast))
+        rows.append(dict(panel="top50_deg_by_type", dataset="ye2026_sepsis", series=ct, pairs="broadcast",
+                         value=deg_call.deg_type("real_lora__raw", ct, "top50_overlap_broadcast"),
+                         source=deg_call.src_contrast))
+    cd = se.summary["contrast_definition"]
+    grp = f"Sepsis {cd['n_case_donors']} vs {cd['n_control_donors']}"
+    heads = [f"ImMicro, {int(im.summary['n_donors'])} donors\ndonor NK fraction r",
+             f"{grp}\ndisease logFC r", f"{grp}\ntop-50 DEG overlap"]
+    for ax, h in zip(tax, heads):
+        ax.set_title(h, fontsize=14, pad=8, linespacing=1.3)
+    x0, x1 = axs[0].get_position().x0, axs[2].get_position().x1
+    fig.text((x0 + x1) / 2, 0.005, xlabel(st), ha="center", va="bottom", fontsize=13)
+    rows.append(dict(panel="meta", dataset="ye2026_sepsis", series="n_case_vs_control", pairs="",
+                     value=f"{se.summary['contrast_definition']['n_case_donors']} vs "
+                           f"{se.summary['contrast_definition']['n_control_donors']}", source=se.src_summary))
+    handles = [  # column-wise fill: [raw, cal, pseudo, ...] | [LoRA, (equal-budget cal), (scSemi), broadcast]
+        h_raw(), h_len(),
+        Line2D([0], [0], ls=REF["pseudo"]["ls"], color=REF["pseudo"]["color"], lw=REF["pseudo"]["lw"],
+               label=REF["pseudo"]["label"]),
+        Line2D([0], [0], marker="o", color=style.ARM["lora"], lw=2.0, markeredgecolor=style.INK, markersize=8,
+               label="LoRA fine-tune"),
+    ] + ([h_budget()] if has_budget else []) + ([h_scsemi()] if has_sc else []) + [
+        Line2D([0], [0], ls=REF["broadcast"]["ls"], color=REF["broadcast"]["color"], lw=REF["broadcast"]["lw"],
+               label=REF["broadcast"]["label"]),
+    ]
+    fig.legend(handles=handles, loc="upper left", bbox_to_anchor=(0.005, 1.0), ncol=2, frameon=False,
+               fontsize=12, handlelength=2.4, columnspacing=1.2, labelspacing=0.4, borderaxespad=0.0)
+    style.save(fig, "sec08_pairs_signal", FIGDIR)
+
+
+def slide_numbers(calls, deg_call, st, fold_sizes, fewshot, budget, budget_log, scs=None, sc_log=(), lens=None):
+    """Numbers quoted in the slide keywords/footnote, each with its source file."""
+    im, se = calls["immicro"], calls["ye2026_sepsis"]
+    out = []
+
+    def add(key, value, src):
+        out.append(dict(key=key, value=value, source=src))
+
+    for ds, c in calls.items():
+        for key in ("coverage", "comp_l1"):
+            add(f"{ds}.{key}.zero_shot", round(c.head(ZS, key), 3), c.src_summary)
+            add(f"{ds}.{key}.zs_cal", round(c.head(ZS_CAL, key), 3), c.src_summary)
+            if lens:
+                add(f"{ds}.{key}.zs_len", round(lens[ds].head(ZS_LEN, key), 3), lens[ds].src_summary)
+            add(f"{ds}.{key}.lora_n5", round(c.head("lora_n5__raw", key), 3), c.src_summary)
+            add(f"{ds}.{key}.lora_all", round(c.head("real_lora__raw", key), 3), c.src_summary)
+            add(f"{ds}.{key}.zs_pseudo", round(c.head("zs__pseudo", key), 3), c.src_summary)
+        add(f"{ds}.comp_l1.null", round(c.head("real_lora__raw", "comp_l1_obs_loo_null"), 3), c.src_summary)
+        # donor-level composition L1, each LoRA size vs zero-shot (paired Wilcoxon, uncorrected)
+        piv = c.per_donor.pivot(index="donor", columns="arm", values="comp_l1")
+        for a, lab in zip(LORA_ARMS, X_LAB[1:]):
+            x = (piv[a] - piv[ZS]).dropna()
+            add(f"{ds}.comp_l1.lora_{lab}_minus_zero_shot.mean", round(float(x.mean()), 3), c.src_per_donor)
+            add(f"{ds}.comp_l1.lora_{lab}_minus_zero_shot.frac_donors_lower", round(float((x < 0).mean()), 2),
+                c.src_per_donor)
+            add(f"{ds}.comp_l1.lora_{lab}_vs_zero_shot.wilcoxon_p", float(f"{wilcoxon(x).pvalue:.2g}"),
+                c.src_per_donor)
+        # each LoRA size vs zero-shot + calibration (donor L1 and donor coverage; paired Wilcoxon, uncorrected)
+        for met in ("comp_l1", "coverage"):
+            pv = c.per_donor.pivot(index="donor", columns="arm", values=met)
+            for a, lab in zip(LORA_ARMS, X_LAB[1:]):
+                x = (pv[a] - pv[ZS_CAL]).dropna()
+                add(f"{ds}.{met}.lora_{lab}_minus_zs_cal.mean", round(float(x.mean()), 3), c.src_per_donor)
+                add(f"{ds}.{met}.lora_{lab}_minus_zs_cal.frac_donors_lower", round(float((x < 0).mean()), 2),
+                    c.src_per_donor)
+                add(f"{ds}.{met}.lora_{lab}_vs_zs_cal.wilcoxon_p", float(f"{wilcoxon(x).pvalue:.2g}"),
+                    c.src_per_donor)
+            # the raw -> calibrated zero-shot gap
+            x = (pv[ZS_CAL] - pv[ZS]).dropna()
+            add(f"{ds}.{met}.zs_cal_minus_zero_shot.mean", round(float(x.mean()), 3), c.src_per_donor)
+            add(f"{ds}.{met}.zs_cal_vs_zero_shot.wilcoxon_p", float(f"{wilcoxon(x).pvalue:.2g}"), c.src_per_donor)
+        add(f"{ds}.cal_alpha_fit_pairs", rng_txt(fold_sizes[ds]["alpha"]),
+            fold_sizes[ds]["source"] + " fold_sizes train+inner_val; tools/prepare_dataset.py fit = fold != f")
+        # few-pair call (its own donors): alpha from the same 5 pairs vs LoRA on them (fixed 3k steps, no inner-val)
+        add(f"{ds}.fewshot_n5.coverage.zs_cal_same5", round(fewshot[ds]["zs_cal"], 3), fewshot[ds]["source"])
+        add(f"{ds}.fewshot_n5.coverage.lora5_fixed3k", round(fewshot[ds]["lora5"], 3), fewshot[ds]["source"])
+        add(f"{ds}.n_scored_donors", int(c.summary["n_donors"]), c.src_summary)
+        s = st[st.dataset == ds]
+        add(f"{ds}.selected_step.lora_n5.range", rng_txt(s[s.arm == "lora_n5"].selected_step), "RUN.json: "
+            + ";".join(s[s.arm == "lora_n5"].source))
+        add(f"{ds}.n_train.all", rng_txt(s[s.arm == "real_lora"].n_train), "RUN.json: "
+            + ";".join(s[s.arm == "real_lora"].source))
+        add(f"{ds}.n_inner_val", rng_txt(s.n_inner_val), "RUN.json inner_val_rows (all LoRA arms, 5 folds)")
+        for arm in RUN_ARMS:
+            f = s[s.arm == arm]
+            add(f"{ds}.folds_cap_limited.{arm}", f"{int(f.cap_limited.sum())}/{len(f)}",
+                "RUN.json steps_run == max_steps and steps_run - selected_step < patience*eval_every: "
+                + ";".join(f.source))
+        capped = s[s.cap_limited]
+        if len(capped):
+            add(f"{ds}.cap_limited.max_val_loss_drop_frac_last_2k_steps",
+                round(float(capped.val_loss_last_window_drop_frac.max()), 4),
+                "history.json inner_val_loss (drop over last patience window / drop from step 0): "
+                + ";".join(capped.history))
+    for a, lab in zip([ZS, ZS_CAL] + LORA_ARMS + ["zs__pseudo"], KEY_LAB[:1] + ["zs_cal"] + KEY_LAB[1:] + ["pseudo"]):
+        add(f"immicro.nk_fraction_r.{lab}", round(im.frac_r(a, "NK cell"), 2), im.src_per_donor)
+        add(f"sepsis.logfc_r.{lab}", round(se.contrast_r(a), 2), se.src_contrast)
+        add(f"sepsis.top50_deg.{lab}", int(deg_call.deg(a)), deg_call.src_contrast)
+    add("sepsis.logfc_r.broadcast", round(se.contrast_r("real_lora__raw", "r_broadcast"), 2), se.src_contrast)
+    add("sepsis.top50_deg.broadcast", int(deg_call.deg("real_lora__raw", "top50_overlap_broadcast")),
+        deg_call.src_contrast)
+
+    def first_sustained(v, thr, strict=True):
+        """First LoRA n from which every later n is above (or >=) thr."""
+        for i in range(len(v)):
+            if all((x > thr) if strict else (x >= thr) for x in v[i:]):
+                return X_LAB[1 + i]
+        return "never"
+    bc_r = se.contrast_r("real_lora__raw", "r_broadcast")
+    bc_d = deg_call.deg("real_lora__raw", "top50_overlap_broadcast")
+    vr = [se.contrast_r(a) for a in LORA_ARMS]
+    vd = [deg_call.deg(a) for a in LORA_ARMS]
+    add("sepsis.logfc_r.first_n_above_broadcast", first_sustained(vr, bc_r), se.src_contrast)
+    add("sepsis.logfc_r.first_n_above_zero_shot", first_sustained(vr, se.contrast_r(ZS)), se.src_contrast)
+    add("sepsis.top50_deg.first_n_at_or_above_broadcast", first_sustained(vd, bc_d, strict=False),
+        deg_call.src_contrast)
+    add("sepsis.top50_deg.first_n_above_zero_shot", first_sustained(vd, deg_call.deg(ZS)), deg_call.src_contrast)
+    add("sepsis.logfc_r.first_n_above_zs_cal", first_sustained(vr, se.contrast_r(ZS_CAL)), se.src_contrast)
+    add("sepsis.top50_deg.first_n_above_zs_cal", first_sustained(vd, deg_call.deg(ZS_CAL)), deg_call.src_contrast)
+    vn = [im.frac_r(a, "NK cell") for a in LORA_ARMS]
+    add("immicro.nk_fraction_r.first_n_above_zs_cal", first_sustained(vn, im.frac_r(ZS_CAL, "NK cell")),
+        im.src_per_donor)
+    for ds, c in calls.items():
+        for key in ("coverage", "comp_l1"):
+            v = [c.head(a, key) for a in LORA_ARMS]
+            thr = c.head(ZS_CAL, key)
+            first = first_sustained([-x for x in v], -thr) if key == "comp_l1" else first_sustained(v, thr)
+            add(f"{ds}.{key}.first_n_better_than_zs_cal", first, c.src_summary)
+    # across-donor fraction r per cell type: first n from which LoRA stays at/above zero-shot (footnote)
+    for ct, key in [("T cell", "T"), ("B cell", "B"), ("Monocyte", "Mono"), ("NK cell", "NK")]:
+        v = [im.frac_r(a, ct) for a in LORA_ARMS]
+        add(f"immicro.fraction_r.{key}.first_n_at_or_above_zero_shot",
+            first_sustained(v, im.frac_r(ZS, ct), strict=False), im.src_per_donor)
+        add(f"immicro.fraction_r.{key}.zero_shot", round(im.frac_r(ZS, ct), 2), im.src_per_donor)
+        add(f"immicro.fraction_r.{key}.zs_cal", round(im.frac_r(ZS_CAL, ct), 2), im.src_per_donor)
+        add(f"immicro.fraction_r.{key}.first_n_at_or_above_zs_cal",
+            first_sustained(v, im.frac_r(ZS_CAL, ct), strict=False), im.src_per_donor)
+        add(f"immicro.fraction_r.{key}.lora_5_10_20_40_all", "/".join(f"{x:.2f}" for x in v), im.src_per_donor)
+    # top-50 DEG overlap by cell type (which types carry the gain)
+    for ct, key in [("T cell", "T"), ("B cell", "B"), ("Monocyte", "Mono"), ("NK cell", "NK"), ("DC", "DC")]:
+        v = [deg_call.deg_type(a, ct) for a in [ZS] + LORA_ARMS]
+        add(f"sepsis.top50_deg.{key}.zero_shot_5_10_20_40_all", "/".join(f"{x:g}" for x in v),
+            deg_call.src_contrast)
+        add(f"sepsis.top50_deg.{key}.zs_cal", int(deg_call.deg_type(ZS_CAL, ct)), deg_call.src_contrast)
+        add(f"sepsis.top50_deg.{key}.broadcast", int(deg_call.deg_type("real_lora__raw", ct,
+                                                                       "top50_overlap_broadcast")),
+            deg_call.src_contrast)
+    add("sepsis.n_case", int(se.summary["contrast_definition"]["n_case_donors"]), se.src_summary)
+    add("sepsis.n_control", int(se.summary["contrast_definition"]["n_control_donors"]), se.src_summary)
+    add("final_checkpoint", FINAL_CKPT, "RUN.json source_checkpoint (all 50 LoRA runs)")
+    # equal-budget calibration scan (zero-shot + calibration with alpha from the n + inner-val pairs of lora_n*)
+    add("cal_budget.pattern", f"scores/<dataset>/**/summary.json arm names: 'cal' and regex {CAL_N_RE.pattern}, "
+        f"none of {list(FT_TAGS)}", str(PLAN / "scores"))
+    add("cal_budget.found", ";".join(f"{r['dataset']}:{r['call']}:{r['arm']}" for r in budget_log) or "none",
+        str(PLAN / "scores"))
+    for ds in DSETS:
+        add(f"{ds}.cal_budget.n_drawn", ",".join(str(n) for n in sorted(budget[ds])) or "none", str(PLAN / "scores"))
+    # scSemiProfiler line (equal-budget n + 5-fold CV "all")
+    add("scsemi.found", ";".join(f"{r['dataset']}:{r['call']}:{r['arm']}:{r['note']}" for r in sc_log) or "none",
+        f"{SR.plan_root()}/scores/<dataset>/{{{SR.BUDGET_EXP},{SR.CV_EXP}}}/pooled*")
+    for ds in DSETS:
+        sc = (scs or {}).get(ds, {})
+        add(f"{ds}.scsemi.pairs_drawn", ",".join(lab for lab in X_LAB[1:] if lab in sc) or "none",
+            f"{SR.plan_root()}/scores/{ds}")
+        for lab in X_LAB[1:]:
+            if lab in sc:
+                for key in ("coverage", "comp_l1"):
+                    v, c = budget_value(sc[lab], lambda cc, a, k=key: cc.head(a, k))
+                    if v is not None:
+                        add(f"{ds}.{key}.scsemi_{lab}", round(v, 3), str(c.path / "summary.json"))
+    return pd.DataFrame(out)
+
+
+def check_against_fig3(rows):
+    """Cross-check the LoRA curve, the zero-shot + calibration point and the reference lines with the plan's own
+    fig3 source data (same numbers, independent code path; the plan's fig3 n=0 point is zs__cal). The zero-shot
+    real-bulk point is not in fig3; it is checked in load() instead (headline vs per_donor.csv / contrast.csv)."""
+    ref = pd.read_csv(PLAN / "figures" / "source_data" / "fig3_learning_curve.csv")
+    pmap = {("comp_l1", "immicro"): "3a", ("comp_l1", "ye2026_sepsis"): "3b",
+            ("coverage", "immicro"): "3c", ("coverage", "ye2026_sepsis"): "3d",
+            ("nk_fraction_r", "immicro"): "3e", ("logfc_r", "ye2026_sepsis"): "3f",
+            ("top50_deg", "ye2026_sepsis"): "3h"}
+    n = 0
+    for r in rows:
+        p = pmap.get((r["panel"], r["dataset"]))
+        if p is None or r["series"] in ("zero_shot", "zs_cal_budget", "scsemi"):
+            continue
+        if r["series"] == "lora":
+            m = ref[(ref.panel == p) & (ref.series == r["series"]) & (ref.draw == f"n={r['pairs']}")]
+        elif r["series"] == "zs_cal":
+            m = ref[(ref.panel == p) & (ref.series == "zs_cal") & (ref.draw == "n=0")]
+        else:
+            m = ref[(ref.panel == p) & (ref.series == r["series"])]
+        assert len(m) == 1, (r, len(m))
+        assert abs(float(m.value.iloc[0]) - float(r["value"])) < 1e-5, (r, float(m.value.iloc[0]))
+        n += 1
+    return n
+
+
+def main():
+    calls, deg_call, lens = load()
+    st = load_steps()
+    fold_sizes = load_fold_sizes()
+    fewshot = load_fewshot()
+    budget, budget_log = find_cal_budget(calls, deg_call)
+    scs, sc_log = find_scsemi(calls)
+    rows, steprows = [], []
+    fig_realism(calls, lens, st, fold_sizes, rows, steprows, budget, scs)
+    rows_sig = []
+    fig_signal(calls, deg_call, lens, st, rows_sig, budget, scs)
+    n_checked = check_against_fig3(rows + rows_sig)
+    pd.DataFrame(rows).to_csv(DATADIR / "sec08_pairs_realism.csv", index=False, float_format="%.6g")
+    pd.DataFrame(steprows).to_csv(DATADIR / "sec08_selected_steps.csv", index=False, float_format="%.6g")
+    st.to_csv(DATADIR / "sec08_selected_steps_per_fold.csv", index=False)
+    pd.DataFrame(rows_sig).to_csv(DATADIR / "sec08_pairs_signal.csv", index=False, float_format="%.6g")
+    nums = slide_numbers(calls, deg_call, st, fold_sizes, fewshot, budget, budget_log, scs, sc_log, lens)
+    nums.to_csv(DATADIR / "sec08_slide_numbers.csv", index=False)
+    print(nums.to_string(index=False, max_colwidth=70))
+    print(f"cross-checked {n_checked} plotted values against fig3_learning_curve.csv")
+    brk = [(r["panel"], r["dataset"]) for r in rows + rows_sig if r.get("series") == "zero_shot" and r["y_axis_break"]]
+    print(f"y-axis breaks: {brk}")
+    print(f"equal-budget calibration scan: arm names with 'cal' and regex {CAL_N_RE.pattern!r}, none of {FT_TAGS}, "
+          f"in {PLAN / 'scores'}/<dataset>/**/summary.json")
+    for r in budget_log:
+        print(f"  found {r['dataset']} {r['call']} {r['arm']} (n {r['n']}): {r['note']}")
+    if not any(budget[ds] for ds in DSETS):
+        print("  NOTE: no equal-budget calibration arms found; only the two x = 0 zero-shot points are drawn.")
+    print(f"scSemiProfiler scan: {SR.plan_root()}/scores/<dataset>/{SR.BUDGET_EXP}/pooled* (scsemi_n<n>__raw) and "
+          f"{SR.CV_EXP}/pooled* ({SR.CV_ARM}, drawn at 'all')")
+    for r in sc_log:
+        print(f"  found {r['dataset']} {r['call']} {r['arm']} (pairs {r['pairs']}): {r['note']}")
+    if not any(scs[ds] for ds in DSETS):
+        print("  NOTE: no scSemiProfiler equal-budget or 5-fold calls found; no scSemiProfiler line is drawn.")
+
+
+if __name__ == "__main__":
+    main()

paper-releases/pbisc-diffusion/updates/20261002T085943Z-deck-restructure-contrast-metrics-zeroshot-len/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/sec09_cat_curves.py

@@ -0,0 +1,161 @@
+#!/usr/bin/env python3
+"""Section 9 of the 2026-09-30 paper-summary deck: CAT curves (correspondence at the top) of the disease genes
+(added 2026-10-02, user request: a threshold-free DEG comparison with the random expectation drawn).
+
+CPU only. Reads plans/paired_finetune_intradataset_20260927/scores_contrast/{fewshot,pooled}_cat.csv
+(tools/contrast_extra.py): per cell type, all tested genes ranked by |limma moderated t| on the observed side and on the
+predicted side; CAT(k) = fraction of the observed top-k genes that are also in the predicted top-k with the same logFC
+sign (Irizarry et al. 2005); chance = k / (2 G) for G genes tested on both sides. Plotted: mean over T / B / Mono / NK;
+5 donors = mean over the 5 draws of that per-draw mean; 5-fold = the pooled call. The broadcast is the raw real bulk of
+the LoRA arm's condition (same ranking statistic on the input bulk).
+
+Outputs (deck folder; DECK_FIGDIR / DECK_OUTDATA override the folders):
+  figures/sec09_cat_curves.{png,pdf}
+  data/sec09_cat_curves.csv                    every plotted value with its source
+  data/sec09_cat_curves_slide_text.json        title, keywords, footnotes (synced through text_source)
+"""
+from __future__ import annotations
+
+import os
+
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+DECK = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/"
+            "2026-09-30--paper-summary--r00")
+sys.path.insert(0, str(DECK / "tools"))
+import sec09_paired_validation as S9  # noqa: E402  (names, colours, labels; sets the deck rcParams)
+import style  # noqa: E402
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.lines import Line2D  # noqa: E402
+
+CX = S9.T / "scores_contrast"
+FIGDIR = Path(os.environ.get("DECK_FIGDIR", str(DECK / "figures")))
+OUTDATA = Path(os.environ.get("DECK_OUTDATA", str(DECK / "data")))
+DIS = S9.DISEASE
+NAME = S9.NAME
+TYPES = ["T cell", "B cell", "Monocyte", "NK cell"]
+FS = style.FS
+FEW = [("zs_raw", "zs_raw"), ("zs_cal", "zs_cal"), ("scsemi5", "scsemi5"), ("lora", "lora5")]
+POOL = [("zs_raw", "zs__raw"), ("zs_cal", "zs__cal"), ("sconly_cal", "pseudo_full__cal"), ("lora", "real_lora__raw")]
+LORA_ARM = {"fewshot": "lora5", "pooled": "real_lora__raw"}
+
+
+def curves(cat: pd.DataFrame, design: str):
+    """{(dataset, key): Series over k} for generated arms, 'broadcast' and 'chance' (chance of the LoRA arm's lists)."""
+    c = cat[cat.celltype.isin(TYPES)].copy()
+    c["kind"] = np.where(c.predictor.str.startswith("generated:"), "gen", "bc")
+    out = {}
+    arms = FEW if design == "fewshot" else POOL
+    for ds in DIS:
+        d = c[c.dataset == ds]
+        for key, arm in arms:
+            g = d[(d.arm == arm) & (d.kind == "gen")]
+            if g.empty:
+                continue
+            out[(ds, key)] = g.groupby(["unit", "k"]).cat_same_sign.mean().groupby("k").mean()
+        lo = d[d.arm == LORA_ARM[design]]
+        out[(ds, "broadcast")] = lo[lo.kind == "bc"].groupby(["unit", "k"]).cat_same_sign.mean().groupby("k").mean()
+        out[(ds, "chance")] = lo[lo.kind == "gen"].groupby(["unit", "k"]).chance_same_sign.mean().groupby("k").mean()
+    return out
+
+
+def main() -> int:
+    few = curves(pd.read_csv(CX / "fewshot_cat.csv"), "fewshot")
+    poo = curves(pd.read_csv(CX / "pooled_cat.csv"), "pooled")
+    fig, axs = plt.subplots(2, len(DIS), figsize=(12.3, 5.2), sharex=True,
+                            gridspec_kw=dict(left=0.105, right=0.995, top=0.90, bottom=0.215, hspace=0.32, wspace=0.28))
+    rows = []
+    for ri, (lab, cv, keys, src) in enumerate([("5 donors", few, FEW, "fewshot_cat.csv"),
+                                               ("5-fold CV", poo, POOL, "pooled_cat.csv")]):
+        for ci, ds in enumerate(DIS):
+            ax = axs[ri, ci]
+            hi = 0.0
+            for key, _ in keys:
+                if (ds, key) not in cv:
+                    continue
+                s = cv[(ds, key)]
+                ax.plot(s.index, s.values, color=S9.COL[key], lw=2.2 if key == "lora" else 1.6, marker="o", ms=4,
+                        mec=S9.EDGE, mew=0.5, zorder=4 if key == "lora" else 3)
+                hi = max(hi, s.max())
+                rows += [dict(design=lab, dataset=ds, series=key, k=int(k), cat=float(v), source=str(CX / src))
+                         for k, v in s.items()]
+            b, ch = cv[(ds, "broadcast")], cv[(ds, "chance")]
+            st = S9.REFSTYLE["broadcast"]
+            ax.plot(b.index, b.values, ls=st["ls"], color=st["color"], lw=st["lw"], zorder=2)
+            ax.plot(ch.index, ch.values, ls=(0, (3, 2)), color=style.INK, lw=1.0, zorder=2)
+            hi = max(hi, b.max())
+            rows += [dict(design=lab, dataset=ds, series=s_, k=int(k), cat=float(v), source=str(CX / src))
+                     for s_, ser in (("broadcast", b), ("chance", ch)) for k, v in ser.items()]
+            ax.set_xscale("log")
+            ax.set_xticks([10, 50, 100, 500], ["10", "50", "100", "500"])
+            ax.set_ylim(0, max(0.12, hi * 1.15))
+            ax.yaxis.set_major_locator(plt.MaxNLocator(nbins=4))
+            ax.tick_params(labelsize=FS - 3)
+            if ri == 0:
+                ax.set_title(S9.ROWNAME[ds], fontsize=FS - 1, pad=6)
+            if ri == 1:
+                ax.set_xlabel("top-k genes (k)", fontsize=FS - 2)
+        axs[ri, 0].annotate(lab, xy=(0, 0.5), xycoords="axes fraction", xytext=(-58, 0), textcoords="offset points",
+                            ha="right", va="center", fontsize=FS, fontweight="bold")
+        axs[ri, 0].set_ylabel("CAT, same sign", fontsize=FS - 3)
+    hk = [("zs_raw", S9.LAB["zs_raw"]), ("zs_cal", S9.LAB["zs_cal"]), ("scsemi5", "scSemiProfiler (5 donors)"),
+          ("sconly_cal", "sc-only fine-tune (5-fold)"), ("lora", "LoRA")]
+    handles = [Line2D([0], [0], color=S9.COL[k], lw=2.0, marker="o", ms=5, mec=S9.EDGE, mew=0.5, label=l) for k, l in hk]
+    handles += [S9.refh("broadcast"), Line2D([0], [0], ls=(0, (3, 2)), color=style.INK, lw=1.0, label="Chance (k / 2G)")]
+    fig.legend(handles=handles, loc="lower center", ncol=4, bbox_to_anchor=(0.53, -0.01), fontsize=FS - 3,
+               frameon=False, columnspacing=1.4, handlelength=2.2)
+    style.save(fig, "sec09_cat_curves", FIGDIR)
+    d = pd.DataFrame(rows)
+    OUTDATA.mkdir(exist_ok=True)
+    d.to_csv(OUTDATA / "sec09_cat_curves.csv", index=False, float_format="%.6g")
+
+    # slide text, from the plotted values at k = 100
+    def at(cv, ds, key, k=100):
+        s = cv.get((ds, key))
+        return np.nan if s is None or k not in s.index else float(s.loc[k])
+
+    top_few = [ds for ds in DIS if at(few, ds, "lora") >= max(at(few, ds, k) for k, _ in FEW if k != "lora")]
+    top_poo = [ds for ds in DIS if at(poo, ds, "lora") >= max(at(poo, ds, k) for k, _ in POOL if k != "lora")]
+    bc_few = [ds for ds in DIS if at(few, ds, "broadcast") > at(few, ds, "lora")]
+    kw = [f"LoRA top model: {len(top_few)}/4 (5 donors), {len(top_poo)}/4 (5-fold)",
+          (f"Broadcast above LoRA: {S9.names(bc_few)}" if bc_few else "LoRA above broadcast in every cohort"),
+          f"Chance at k = 100: {min(at(few, ds, 'chance') for ds in DIS):.3f}–{max(at(few, ds, 'chance') for ds in DIS):.3f}"]
+    for k in kw:
+        if len(k) > 52:
+            raise ValueError(f"keyword too long: {k!r}")
+    vals = "; ".join(
+        f"{NAME[ds]} CAT100 5 donors LoRA {at(few, ds, 'lora'):.3f} / scSemi {at(few, ds, 'scsemi5'):.3f} / "
+        f"calibration {at(few, ds, 'zs_cal'):.3f} / raw {at(few, ds, 'zs_raw'):.3f} / broadcast {at(few, ds, 'broadcast'):.3f}"
+        f" / chance {at(few, ds, 'chance'):.3f}; 5-fold LoRA {at(poo, ds, 'lora'):.3f} / broadcast {at(poo, ds, 'broadcast'):.3f}"
+        for ds in DIS)
+    full = ("Correspondence at the top (Irizarry et al. 2005): per cell type, every gene tested on both sides is ranked by "
+            "|limma-voom moderated t| (case vs control, same donor x type pseudobulks as the DEG panels); CAT(k) = share of "
+            "the observed top-k genes that are also in the generated top-k with the same logFC sign. Mean over T / B / Mono "
+            "/ NK; 5 donors = mean over 5 draws (scSemiProfiler: its 25 draws); 5-fold = pooled call. Broadcast = the "
+            "input real bulk ranked the same way. Chance = k / (2G), G = genes tested on both sides ("
+            f"{min(at(few, ds, 'chance') for ds in DIS):.3f}–{max(at(few, ds, 'chance') for ds in DIS):.3f} at k = 100). "
+            "No significance threshold, so cohorts where few genes pass FDR (Giroux, Ji) still get a curve. "
+            "Rank-biased overlap (Webber et al. 2010, p = 0.98) gives the same order: scores_contrast/{fewshot,pooled}_rbo.csv. "
+            "Giroux: sex-linked genes excluded. " + vals + ".")
+    txt = {"title": "Disease genes at every list length (CAT)", "layout": "full",
+           "figures": [str(FIGDIR / "sec09_cat_curves.png")], "keywords": kw,
+           "footnote_short": "Same-sign overlap of observed and generated top-k genes per cell type; no FDR cut-off.",
+           "footnote": full,
+           "numbers": [{"value": vals, "source": str(CX / "fewshot_cat.csv") + "; " + str(CX / "pooled_cat.csv")}]}
+    (OUTDATA / "sec09_cat_curves_slide_text.json").write_text(json.dumps(
+        {"section": "09-cat", "scripts": [str(Path(__file__).resolve())],
+         "generated_from": "tools/sec09_cat_curves.py (rerun to refresh; do not edit by hand)", "slides": [txt]},
+        indent=1, ensure_ascii=False))
+    print("\n".join(kw))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

paper-releases/pbisc-diffusion/updates/20261002T085943Z-deck-restructure-contrast-metrics-zeroshot-len/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/sec09_extra_metrics.py

@@ -0,0 +1,738 @@
+#!/usr/bin/env python3
+"""Section 9 extras of the 2026-09-30 paper-summary deck: realism / diversity metrics beyond precision and coverage
+(classifier AUC, MMD^2, scSemiProfiler's E_N, precision/coverage at k = 3 / 5 / 10).
+
+CPU only: reads the per-unit outputs of plans/paired_finetune_intradataset_20260927/tools/extra_metrics.py
+(scores_extra/<ds>/{fewshot_n5/draw<j>, pooled}/{metrics.csv, en.csv, manifest.json}) and draws slide figures.
+No torch, no CUDA, no scoring. The headline definitions are the collector's (tools/collect_extra_metrics.py,
+scores_extra/README.md) and every value is checked against scores_extra/{fewshot,pooled}_headline.csv:
+  classifier AUC (kNN, RF), 1-NN accuracy, MMD^2  real side R2 (held-out real cells), no depth matching;
+                                                  control = real vs real, R2 vs R3
+  E_N                                             all cells, squared L2 (their code); control = within-sample half
+                                                  split; undefined (UB <= LB) -> left out of means, ranks last
+  precision / coverage at k = 3, 5, 10            the scorer's space (real side R1); control = R1 vs R2
+
+Outputs (deck folder; DECK_FIGDIR / DECK_OUTDATA override the folders, used by the synthetic test):
+  figures/sec09_extra_fewpair.{png,pdf}    main slide: 5 donors, kNN AUC / MMD^2 / E_N per cohort
+  figures/sec09_extra_fivefold.{png,pdf}   section 9: the same for the 5-fold pooled design
+  figures/sec09_extra_ksens.{png,pdf}      section 9: precision / coverage at k = 3, 5, 10 (5 donors)
+  data/sec09_extra_{fewpair,fivefold,ksens}.csv   every plotted value with its source file
+  data/sec09_extra_scsemi_status.csv       scSemiProfiler CV / equal-budget extra-metric units found or not
+  data/sec09_extra_quoted_numbers.csv      quoted numbers (flat)
+  data/sec09_extra_slide_text.json         titles, keywords, footnotes, quoted numbers (synced into
+                                           data/slides_spec.json by tools/sync_sec09_text.py)
+
+scSemiProfiler with all training pairs (5-fold CV, arm scsemi__raw): drawn in the 5-fold figure as its own bar when an
+extra-metric unit with that set exists (scores_extra/<ds>/pooled/ itself, or below scores_extra/<ds>/scsemi_cv_20261002/,
+see tools/scsemi_results.py) for the same donors and reference cells; otherwise a note is printed and nothing is drawn.
+Equal-budget units (scsemi_n<n>__raw) are only listed in the status CSV.
+"""
+from __future__ import annotations
+
+import os
+
+os.environ["CUDA_VISIBLE_DEVICES"] = ""  # plotting only
+
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+DECK = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/"
+            "2026-09-30--paper-summary--r00")
+sys.path.insert(0, str(DECK / "tools"))
+import sec09_paired_validation as S9  # noqa: E402  (shared section 9 drawing helpers; sets the deck rcParams)
+import scsemi_results as SR  # noqa: E402
+import style  # noqa: E402
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.lines import Line2D  # noqa: E402
+from matplotlib.patches import Patch  # noqa: E402
+
+T = S9.T
+X = T / "scores_extra"
+FIGDIR = Path(os.environ.get("DECK_FIGDIR", str(DECK / "figures")))
+OUTDATA = Path(os.environ.get("DECK_OUTDATA", str(DECK / "data")))
+SCRIPT = DECK / "tools" / "sec09_extra_metrics.py"
+ORDER, NAME, ROWNAME = S9.ORDER, S9.NAME, S9.ROWNAME
+FS, SUB, EDGE = S9.FS, S9.SUB, S9.EDGE
+
+# arm keys (deck colours / labels from sec09) -> set names in scores_extra
+FEW_SETS = {"zs_raw": "zs_raw", "zs_cal": "zs_cal", "scsemi5": "scsemi5", "lora": "lora5", "pseudo": "zs_pseudo"}
+POOL_SETS = {"zs_raw": "zs__raw", "zs_cal": "zs__cal", "sconly_cal": "pseudo_full__cal", "scsemi_all": SR.CV_ARM,
+             "lora": "real_lora__raw", "pseudo": "zs__pseudo"}
+COL = {**S9.COL, "scsemi_all": style.ARM["scsemi"]}
+LAB = {**S9.LAB, "scsemi_all": SR.LABEL_CV}
+# headline metrics (collector's HEADLINE): column, short name, direction (0: closer to 0.5, -1 lower, +1 higher), space
+TWO = [("c2st_knn_auc", "kNN AUC", 0), ("c2st_rf_auc", "RF AUC", 0), ("nn1_acc", "1-NN accuracy", 0),
+       ("mmd2_bw1", "MMD² (σ = m)", -1), ("mmd2_sum", "MMD² (σ sum)", -1)]
+EN = [("en_all_sq", "E_N (all cells)", -1)]
+PC = [(f"{q}_k{k}", f"{q} k{k}", 1) for k in (3, 5, 10) for q in ("precision", "coverage")]
+HEAD = TWO + EN + PC
+DIRS = {c: s for c, _, s in HEAD}
+SHORT = {c: n for c, n, _ in HEAD}
+# drawn columns of the two realism figures
+PANELS = [("c2st_knn_auc", "Classifier AUC (kNN) ↓"), ("mmd2_bw1", "MMD² ↓"), ("en_all_sq", "scSemiProfiler E_N ↓ (log)")]
+CTRL = dict(ls=(0, (6, 2, 1.5, 2)), color="#7A7A7A", lw=1.8, label="Real vs real (same samples)")
+KS = (3, 5, 10)
+NUMS = {k: [] for k in "abc"}
+
+
+def num(slide, value, source):
+    NUMS[slide].append({"value": str(value), "source": str(source)})
+
+
+# ------------------------------------------------------------------------------------------
+# Loading
+# ------------------------------------------------------------------------------------------
+def unit_values(u: Path) -> dict:
+    """set -> {metric: value} in the headline spaces of one extra-metric unit; 'control' = the real-vs-real rows."""
+    m = pd.read_csv(u / "metrics.csv")
+    e = pd.read_csv(u / "en.csv")
+    raw = m[m["mode"] == "raw"]
+    out = {}
+    for s in sorted(set(raw["set"])):
+        r2 = raw[(raw["set"] == s) & (raw.real_side == "R2")]
+        r1 = raw[(raw["set"] == s) & (raw.real_side == "R1")]
+        if len(r2) > 1 or len(r1) > 1:
+            raise ValueError(f"{u}: several raw rows for set {s}")
+        d = {c: (float(r2[c].iloc[0]) if len(r2) else np.nan) for c, _, _ in TWO}
+        d.update({c: (float(r1[c].iloc[0]) if len(r1) else np.nan) for c, _, _ in PC})
+        en = e[(e["set"] == s) & (e.cellset == "all") & (e.distance == "sq")]
+        if len(en) > 1:
+            raise ValueError(f"{u}: several E_N rows for set {s}")
+        ok = bool(len(en)) and float(en.ub.iloc[0]) > float(en.lb.iloc[0])
+        d["en_all_sq"] = float(en.en.iloc[0]) if ok else np.nan
+        d["en_all_sq_undefined"] = bool(len(en)) and not ok
+        out[s] = d
+    ctrl = {c: out["real_R3"][c] for c, _, _ in TWO}
+    ctrl.update({c: out["real_R2"][c] for c, _, _ in PC + EN})
+    ctrl["en_all_sq_undefined"] = False
+    out["control"] = ctrl
+    return out
+
+
+def manifest(u: Path) -> dict:
+    p = u / "manifest.json"
+    return json.loads(p.read_text()) if p.is_file() else {}
+
+
+def load_fewpair():
+    """Long table: dataset, draw, arm (deck key or 'control'), metric, value, undefined, source."""
+    rows, info = [], {}
+    for ds in ORDER:
+        units = [d for d in sorted((X / ds / "fewshot_n5").glob("draw*")) if (d / "manifest.json").is_file()]
+        few_min = []
+        for u in units:
+            j = int(u.name.removeprefix("draw"))
+            mf_ = manifest(u)
+            sj = T / "scores" / ds / "fewshot_n5" / f"draw{j}" / "summary.json"
+            s = json.loads(sj.read_text())
+            if set(map(str, mf_["donors"])) != set(map(str, s["donors"])):
+                raise ValueError(f"{u}: scored donors differ from {sj}")
+            if mf_["reference_sha256"] != s["reference"]["reference_sha256"]:
+                raise ValueError(f"{u}: reference differs from {sj}")
+            few_min.append({"dataset": ds, "arm": "lora", "draw": j, "n_scored": int(s["n_donors"])})
+            vals = unit_values(u)
+            for key in list(FEW_SETS) + ["control"]:
+                sname = FEW_SETS.get(key, "control")
+                if sname not in vals:
+                    continue
+                for c, _, _ in HEAD:
+                    rows.append({"dataset": ds, "draw": j, "arm": key, "set": sname, "metric": c,
+                                 "value": vals[sname][c],
+                                 "undefined": bool(vals[sname].get(f"{c}_undefined", False)),
+                                 "source": str(u / ("en.csv" if c.startswith("en_") else "metrics.csv"))})
+        info[ds] = S9.draw_info(pd.DataFrame(few_min), ds)
+    return pd.DataFrame(rows), info
+
+
+def load_pooled():
+    """dataset -> {arm key: {metric: value}}, plus the scSemiProfiler lookup log."""
+    out, log = {}, []
+    for ds in ORDER:
+        u = X / ds / "pooled"
+        vals = unit_values(u)
+        ref = manifest(u)
+        d = {}
+        for key, sname in POOL_SETS.items():
+            if sname in vals:
+                d[key] = {**vals[sname], "source": str(u)}
+        d["control"] = {**vals["control"], "source": str(u)}
+        # scSemiProfiler, all training pairs: in the pooled unit itself, else below scores_extra/<ds>/scsemi_cv_20261002/
+        if "scsemi_all" in d:
+            log.append(dict(dataset=ds, design="cv", unit=str(u), set=SR.CV_ARM, status="used", reason="in pooled unit"))
+        for cu in SR.extra_units(ds, SR.CV_EXP):
+            cv = unit_values(cu)
+            if SR.CV_ARM not in cv:
+                log.append(dict(dataset=ds, design="cv", unit=str(cu), set="", status="ignored",
+                                reason=f"no set {SR.CV_ARM}"))
+                continue
+            cm = manifest(cu)
+            why = []
+            if cm.get("donors") is not None and set(map(str, cm["donors"])) != set(map(str, ref["donors"])):
+                why.append("donor set differs from the pooled unit")
+            if cm.get("reference_sha256") is not None and cm["reference_sha256"] != ref["reference_sha256"]:
+                why.append("reference differs from the pooled unit")
+            if "scsemi_all" in d:
+                why.append("already taken from an earlier unit")
+            if why:
+                log.append(dict(dataset=ds, design="cv", unit=str(cu), set=SR.CV_ARM, status="rejected",
+                                reason="; ".join(why)))
+                continue
+            d["scsemi_all"] = {**cv[SR.CV_ARM], "source": str(cu)}
+            log.append(dict(dataset=ds, design="cv", unit=str(cu), set=SR.CV_ARM, status="used",
+                            reason="no manifest: donors not checked" if not cm else ""))
+        for bu in SR.extra_units(ds, SR.BUDGET_EXP):
+            for s in sorted(unit_values(bu)):
+                if SR.BUDGET_RE.match(s):
+                    log.append(dict(dataset=ds, design="budget", unit=str(bu), set=s, status="listed only",
+                                    reason="equal-budget units are not drawn"))
+        out[ds] = d
+    return out, pd.DataFrame(log, columns=["dataset", "design", "unit", "set", "status", "reason"])
+
+
+def check_headline(few: pd.DataFrame, pooled: dict):
+    """Every value equals the collector's headline tables (few-pair: mean and n over draws; pooled: the value)."""
+    n = 0
+    hf = pd.read_csv(X / "fewshot_headline.csv")
+    for r in hf.itertuples():
+        key = "control" if r.set == "control" else {v: k for k, v in FEW_SETS.items()}[r.set]
+        for c, _, _ in HEAD:
+            x = few[(few.dataset == r.dataset) & (few.arm == key) & (few.metric == c) & ~few.undefined]["value"]
+            want = getattr(r, c)
+            if int(getattr(r, f"{c}_n")) != len(x) or abs(float(x.mean()) - want) > 1e-9:
+                raise ValueError(f"few-pair {r.dataset} {r.set} {c}: {x.mean()} (n {len(x)}) != headline {want}")
+            n += 1
+    hp = pd.read_csv(X / "pooled_headline.csv")
+    for r in hp.itertuples():
+        key = "control" if r.set == "control" else {v: k for k, v in POOL_SETS.items()}[r.set]
+        for c, _, _ in HEAD:
+            v, want = pooled[r.dataset][key][c], getattr(r, c)
+            if not (abs(v - want) <= 1e-9 or (np.isnan(v) and np.isnan(want))):
+                raise ValueError(f"pooled {r.dataset} {r.set} {c}: {v} != headline {want}")
+            n += 1
+    return n
+
+
+# ------------------------------------------------------------------------------------------
+# Drawing
+# ------------------------------------------------------------------------------------------
+def hbars_base(ax, arms, vals_of, base, seed=0, dots=True, na_text=None):
+    """S9.hbars with bars starting at `base` (0.5 for AUC, the axis floor for a log axis)."""
+    out = {}
+    for i, arm in enumerate(arms):
+        vals = np.asarray([v for v in vals_of(arm) if not np.isnan(v)], float)
+        if len(vals) == 0:
+            ax.text(0.01, i, (na_text or {}).get(arm, "not run"), transform=ax.get_yaxis_transform(), va="center",
+                    fontsize=FS - 3, color=SUB)
+            out[arm] = (np.nan, np.nan, 0)
+            continue
+        m = float(vals.mean())
+        sd = float(vals.std(ddof=1)) if len(vals) > 1 else 0.0
+        ax.barh(i, m - base, left=base, height=S9.BAR_H, color=COL[arm], edgecolor=EDGE, linewidth=0.6, zorder=2)
+        if dots and len(vals) > 1:
+            lo = max(m - sd, base) if ax.get_xscale() == "log" else m - sd
+            ax.errorbar(m, i, xerr=[[m - lo], [sd]], fmt="none", ecolor="#111111", elinewidth=1.0, capsize=3,
+                        capthick=1.0, zorder=4)
+            ax.scatter(vals, i + S9.jitter(len(vals), seed + i), s=16, facecolor="white", edgecolor="#111111",
+                       linewidth=0.7, zorder=5, clip_on=False)
+        out[arm] = (m, sd, len(vals))
+    ax.set_ylim(len(arms) - 0.5, -0.5)
+    ax.set_yticks([])
+    return out
+
+
+def vline(ax, x, style_, na):
+    if x is None or np.isnan(x):
+        return
+    ax.plot([x, x], [-0.5, na - 0.5], ls=style_["ls"], color=style_["color"], lw=style_["lw"], zorder=3,
+            solid_capstyle="butt")
+
+
+def ctrl_handle():
+    return Line2D([0], [0], ls=CTRL["ls"], color=CTRL["color"], lw=CTRL["lw"], label=CTRL["label"])
+
+
+def patch(key, label=None):
+    return Patch(facecolor=COL[key], edgecolor=EDGE, linewidth=0.6, label=label or LAB[key])
+
+
+def nice_log_limits(vals):
+    v = [x for x in vals if x > 0 and not np.isnan(x)]
+    lo = max([c for c in (0.01, 0.02, 0.03, 0.05, 0.1, 0.2, 0.5) if c <= min(v) / 1.3] or [0.005])
+    hi = min([c for c in (2, 5, 10, 20, 50, 100, 200, 500, 1000) if c >= max(v) * 1.3] or [max(v) * 2])
+    return lo, hi
+
+
+def axis_setup(axs_col, metric, allv, label):
+    """Shared x scale of one metric column; returns the bar base."""
+    if metric.startswith("c2st") or metric.startswith("nn1"):
+        lo = min(0.45, np.floor(min(allv) * 20) / 20)
+        S9.shared_x(axs_col, lo, 1.0, label, [0.5, 0.75, 1.0])
+        for ax in axs_col:  # bars start at 0.5 (chance); the real-vs-real line sits there, so no spine at lo
+            ax.spines["left"].set_visible(False)
+        return 0.5
+    if metric.startswith("mmd"):
+        hi = np.ceil(max(allv) * 1.08 / 0.05) * 0.05
+        lo = min(-0.02 * hi, min(allv) - 0.01 * hi)
+        S9.shared_x(axs_col, lo, hi, label, [round(t, 2) for t in np.arange(0, hi + 1e-9, 0.1)])
+        for ax in axs_col:
+            ax.spines["left"].set_visible(False)
+            ax.axvline(0, color="#555555", lw=0.8, zorder=1)
+        return 0.0
+    lo, hi = nice_log_limits(allv)
+    for ax in axs_col:
+        ax.set_xscale("log")
+    ticks = [t for t in (0.01, 0.1, 1, 10, 100, 1000) if lo <= t <= hi]
+    S9.shared_x(axs_col, lo, hi, label, ticks)
+    for ax in axs_col:
+        ax.set_xticklabels([f"{t:g}" for t in ticks])
+        ax.xaxis.set_minor_locator(plt.NullLocator())
+    return lo
+
+
+def realism_figure(name, arms, vals_of, ref_of, row_subs, legend_handles, dots, na_text=None, header=None):
+    """rows = cohorts, columns = PANELS; vals_of(ds, arm, metric) -> list; ref_of(ds, kind, metric) -> float."""
+    leg_rows = -(-len(legend_handles) // 3)
+    fig, axs = S9.grid(len(ORDER), len(PANELS), style.FULL, left=S9.LEFT_IN, right=0.25, top=0.45,
+                       bottom=0.65 + 0.30 * leg_rows, hspace=0.22, wspace=0.12)
+    rows = []
+    for j, (mk, lab) in enumerate(PANELS):
+        allv = [v for ds in ORDER for a in arms for v in vals_of(ds, a, mk) if not np.isnan(v)]
+        allv += [ref_of(ds, k, mk) for ds in ORDER for k in ("control", "pseudo")]
+        allv = [v for v in allv if not np.isnan(v)]
+        base = axis_setup(axs[:, j], mk, allv, lab)
+        for i, ds in enumerate(ORDER):
+            ax = axs[i, j]
+            hbars_base(ax, arms, lambda a: vals_of(ds, a, mk), base, seed=17 * i + j, dots=dots,
+                       na_text=(na_text or {}).get(ds))
+            vline(ax, ref_of(ds, "control", mk), CTRL, len(arms))
+            vline(ax, ref_of(ds, "pseudo", mk), S9.REFSTYLE["pseudo"], len(arms))
+            for a in arms:
+                for v in vals_of(ds, a, mk):
+                    rows.append({"figure": name, "dataset": ds, "metric": mk, "series": a, "value": v})
+            for k in ("control", "pseudo"):
+                rows.append({"figure": name, "dataset": ds, "metric": mk, "series": f"ref_{k}",
+                             "value": ref_of(ds, k, mk)})
+    for i, ds in enumerate(ORDER):
+        S9.row_label(axs[i, 0], ROWNAME[ds], row_subs(ds))
+    if header:
+        S9.label_header(axs[0, 0], header)
+    S9.legend_bottom(fig, legend_handles, ncol=3)
+    S9.check_inside(fig, name)
+    style.save(fig, name, FIGDIR)
+    return rows
+
+
+# ------------------------------------------------------------------------------------------
+# Figures
+# ------------------------------------------------------------------------------------------
+FEW_ARMS = ["zs_raw", "zs_cal", "scsemi5", "lora"]
+
+
+def few_vals(few, ds, arm, mk):
+    s = few[(few.dataset == ds) & (few.arm == arm) & (few.metric == mk)].sort_values("draw")
+    return [np.nan if u else v for v, u in zip(s.value, s.undefined)]
+
+
+def few_mean(few, ds, arm, mk):
+    v = [x for x in few_vals(few, ds, arm, mk) if not np.isnan(x)]
+    return float(np.mean(v)) if v else np.nan
+
+
+def fig_fewpair(few, info):
+    nd = {info[d]["n_draws"] for d in ORDER}
+    per_row = len(nd) > 1
+    handles = [patch("zs_raw"), patch("zs_cal"), patch("scsemi5"), patch("lora"), ctrl_handle(), S9.refh("pseudo"),
+               S9.dot_handle(None if per_row else next(iter(nd)))]
+    return realism_figure(
+        "sec09_extra_fewpair", FEW_ARMS, lambda ds, a, mk: few_vals(few, ds, a, mk),
+        lambda ds, k, mk: few_mean(few, ds, k, mk), lambda ds: S9.row_subs(info[ds], per_row), handles, dots=True,
+        header="pairs per draw")
+
+
+def fig_fivefold(pooled, arms, tab):
+    def vals(ds, a, mk):
+        d = pooled[ds].get(a)
+        return [np.nan] if d is None else [d[mk]]
+
+    def ref(ds, k, mk):
+        return pooled[ds][k][mk]
+
+    def subs(ds):
+        n, nd = tab[ds]
+        return [f"n={n}" + (f" ({nd} donors)" if nd != n else "")]
+
+    handles = [patch(a, LAB["lora_all"] if a == "lora" else None) for a in arms] + [ctrl_handle(), S9.refh("pseudo")]
+    na = {ds: {"scsemi_all": "pending"} for ds in ORDER}
+    return realism_figure("sec09_extra_fivefold", arms, vals, ref, subs, handles, dots=False, na_text=na)
+
+
+def fig_ksens(few):
+    fig, axs = plt.subplots(2, len(ORDER), figsize=style.FULL, sharey=True,
+                            gridspec_kw=dict(left=0.075, right=0.995, top=0.93, bottom=0.25, hspace=0.18,
+                                             wspace=0.10))
+    xs = np.arange(len(KS))
+    rows = []
+    mk_style = {"zs_raw": "o", "zs_cal": "o", "scsemi5": "D", "lora": "o"}
+    off = {"zs_raw": -0.09, "zs_cal": -0.03, "scsemi5": 0.03, "lora": 0.09}
+    for i, q in enumerate(("coverage", "precision")):
+        for j, ds in enumerate(ORDER):
+            ax = axs[i, j]
+            cv = [few_mean(few, ds, "control", f"{q}_k{k}") for k in KS]
+            ax.plot(xs, cv, ls=CTRL["ls"], color=CTRL["color"], lw=CTRL["lw"], marker="s", ms=4, zorder=2)
+            for k, v in zip(KS, cv):
+                rows.append({"figure": "sec09_extra_ksens", "dataset": ds, "metric": f"{q}_k{k}",
+                             "series": "control", "value": v, "sd": np.nan})
+            for a in FEW_ARMS:
+                m = [few_mean(few, ds, a, f"{q}_k{k}") for k in KS]
+                sd = [float(np.std([x for x in few_vals(few, ds, a, f"{q}_k{k}") if not np.isnan(x)], ddof=1))
+                      for k in KS]
+                ax.errorbar(xs + off[a], m, yerr=sd, color=COL[a], lw=2.0, marker=mk_style[a], ms=6,
+                            mec=EDGE, mew=0.6, capsize=2.5, elinewidth=1.0, zorder=4 if a == "lora" else 3)
+                for k, v, s in zip(KS, m, sd):
+                    rows.append({"figure": "sec09_extra_ksens", "dataset": ds, "metric": f"{q}_k{k}",
+                                 "series": a, "value": v, "sd": s})
+            ax.set_xticks(xs)
+            ax.set_xticklabels([str(k) for k in KS])
+            ax.set_xlim(-0.4, len(KS) - 0.6)
+            ax.set_ylim(0, 1.05)
+            ax.set_yticks([0, 0.5, 1.0])
+            if i == 0:
+                ax.set_title(ROWNAME[ds], fontsize=FS, pad=6)
+                ax.tick_params(axis="x", labelbottom=False)
+            if j == 0:
+                ax.set_ylabel(f"{q.capitalize()} ↑")
+    fig.text(0.535, 0.155, "k (nearest neighbours per ball)", ha="center", va="center", fontsize=FS)
+    handles = [Line2D([0], [0], color=COL[a], lw=2.0, marker=mk_style[a], ms=6, mec=EDGE, mew=0.6, label=LAB[a])
+               for a in FEW_ARMS] + [Line2D([0], [0], ls=CTRL["ls"], color=CTRL["color"], lw=CTRL["lw"], marker="s",
+                                            ms=4, label="Real vs real (held-out cells)")]
+    S9.legend_bottom(fig, handles, ncol=3)
+    style.save(fig, "sec09_extra_ksens", FIGDIR)
+    return rows
+
+
+# ------------------------------------------------------------------------------------------
+# Comparisons for the slide text
+# ------------------------------------------------------------------------------------------
+def better(a, b, sign):
+    """a better than b (NaN = undefined E_N ranks last)."""
+    if np.isnan(a):
+        return False
+    if np.isnan(b):
+        return True
+    return (abs(a - 0.5) < abs(b - 0.5)) if sign == 0 else (a > b if sign > 0 else a < b)
+
+
+def few_wide(few, mk):
+    s = few[few.metric == mk].copy()
+    s.loc[s.undefined, "value"] = np.nan
+    return s.pivot_table(index=["dataset", "draw"], columns="arm", values="value", dropna=False)
+
+
+def few_compare(few, comps=("scsemi5", "zs_raw", "zs_cal")):
+    """Per cohort x comparator x metric: draws where LoRA-5 is better / worse / tied."""
+    out = []
+    for c, _, sign in HEAD:
+        w = few_wide(few, c)
+        for ds in ORDER:
+            x = w.loc[ds]
+            for comp in comps:
+                xx = x[["lora", comp]].dropna(how="all")
+                nb = sum(better(a, b, sign) for a, b in zip(xx.lora, xx[comp]))
+                nw = sum(better(b, a, sign) for a, b in zip(xx.lora, xx[comp]))
+                out.append(dict(dataset=ds, comp=comp, metric=c, n=len(xx), n_better=nb, n_worse=nw,
+                                lora=float(np.nanmean(xx.lora)), other=float(np.nanmean(xx[comp]))))
+    return pd.DataFrame(out)
+
+
+def best_draws(few, mk, arms=FEW_ARMS):
+    """Draws where LoRA-5 is best of the arms on one metric."""
+    sign = DIRS[mk]
+    w = few_wide(few, mk)
+    n = 0
+    for _, r in w.iterrows():
+        n += all(better(r["lora"], r[a], sign) for a in arms if a != "lora")
+    return n, len(w)
+
+
+def fmt_metric(c, v):
+    return f"{v:.4f}" if c.startswith("mmd") else (f"{v:.2f}" if c.startswith("en_") else f"{v:.3f}")
+
+
+def exceptions_text(exc):
+    """'vs zero-shot + calibration: Imperial E_N (all cells) 3/5 (1.28 vs 1.22), ...' grouped by comparator / cohort."""
+    parts = []
+    for comp, g in exc.groupby("comp", sort=True):
+        items = []
+        for ds in ORDER:
+            h = g[g.dataset == ds]
+            if len(h):
+                items.append(f"{NAME[ds]} " + ", ".join(
+                    f"{SHORT[r.metric]} {r.n_worse}/{r.n} ({fmt_metric(r.metric, r.lora)} vs "
+                    f"{fmt_metric(r.metric, r.other)})" for r in h.itertuples()))
+        parts.append(f"vs {S9.LAB[comp].lower()}: " + "; ".join(items))
+    return " | ".join(parts)
+
+
+def rng2(v, nd=2):
+    v = [x for x in v if not np.isnan(x)]
+    a, b = f"{min(v):.{nd}f}", f"{max(v):.{nd}f}"
+    return a if a == b else f"{a}–{b}"
+
+
+# ------------------------------------------------------------------------------------------
+def main():
+    FIGDIR.mkdir(parents=True, exist_ok=True)
+    OUTDATA.mkdir(parents=True, exist_ok=True)
+    few, info = load_fewpair()
+    pooled, sclog = load_pooled()
+    n_checked = check_headline(few, pooled)
+    sclog.to_csv(OUTDATA / "sec09_extra_scsemi_status.csv", index=False)
+    tab = {}
+    for ds in ORDER:
+        dj = json.loads((T / "data" / ds / "fewshot_n5" / "DRAWS.json").read_text())
+        n = int(json.loads((T / "scores" / ds / "pooled" / "summary.json").read_text())["n_donors"])
+        if n != int(dj["n_usable_rows"]):
+            raise ValueError(f"{ds}: pooled n {n} != usable rows")
+        tab[ds] = (n, int(dj["n_usable_donors"]))
+    has_sc = [ds for ds in ORDER if "scsemi_all" in pooled[ds]]
+    arms5 = ["zs_raw", "zs_cal", "sconly_cal"] + (["scsemi_all"] if has_sc else []) + ["lora"]
+
+    rows_a = fig_fewpair(few, info)
+    rows_b = fig_fivefold(pooled, arms5, tab)
+    rows_c = fig_ksens(few)
+    srcf = f"{X}/<cohort>/fewshot_n5/draw<j>/{{metrics,en}}.csv"
+    srcp = f"{X}/<cohort>/pooled/{{metrics,en}}.csv"
+    pd.DataFrame(rows_a).assign(source=srcf).to_csv(OUTDATA / "sec09_extra_fewpair.csv", index=False,
+                                                    float_format="%.6g")
+    pd.DataFrame(rows_b).assign(source=srcp).to_csv(OUTDATA / "sec09_extra_fivefold.csv", index=False,
+                                                    float_format="%.6g")
+    pd.DataFrame(rows_c).assign(source=srcf).to_csv(OUTDATA / "sec09_extra_ksens.csv", index=False,
+                                                    float_format="%.6g")
+    slides = []
+
+    # ---------------- slide a: 5 donors, realism beyond P/C ----------------
+    cmp = few_compare(few)
+    nb_knn, ntot = best_draws(few, "c2st_knn_auc")
+    nb_rf, _ = best_draws(few, "c2st_rf_auc")
+    nb_nn1, _ = best_draws(few, "nn1_acc")
+    nb_mmd, _ = best_draws(few, "mmd2_bw1")
+    nb_en, _ = best_draws(few, "en_all_sq")
+    en_sc = cmp[(cmp.comp == "scsemi5") & (cmp.metric == "en_all_sq")]
+    en_sc_better, en_sc_n = int(en_sc.n_better.sum()), int(en_sc.n.sum())
+    auc_l = [few_mean(few, d, "lora", "c2st_knn_auc") for d in ORDER]
+    auc_draws = few[(few.arm == "lora") & (few.metric == "c2st_knn_auc")].value
+    kw1 = S9.keyword_or_short(f"Classifier AUC: LoRA-5 best in {nb_knn}/{ntot} draws",
+                              f"AUC: LoRA-5 best in {nb_knn}/{ntot} draws")
+    kw2 = f"E_N: LoRA-5 below scSemi-5 in {en_sc_better}/{en_sc_n} draws"
+    kw3 = f"Still separable: LoRA-5 AUC {rng2(auc_l)}"
+    kw_a = S9.check_keywords([kw1, kw2, kw3])
+    # exceptions over every headline metric (LoRA-5 worse in >= 1 draw)
+    exc = cmp[cmp.n_worse > 0].copy()
+    exc_comps = sorted(set(exc.comp))
+    drawn = {m for m, _ in PANELS}
+    exc_drawn = exc[exc.metric.isin(drawn)].sort_values(["n_worse", "dataset"], ascending=[False, True])
+    if exc.empty:
+        fs_a = "E_N: scSemiProfiler's own error metric. LoRA-5 best of the 4 arms on every metric in every draw."
+    elif exc_comps == ["zs_cal"]:
+        e0 = exc_drawn.iloc[0] if len(exc_drawn) else exc.sort_values("n_worse", ascending=False).iloc[0]
+        nd = 2 if e0.metric.startswith("en_") else 3
+        fs_a = (f"E_N: scSemiProfiler's own error metric. Misses only vs calibration, e.g. {NAME[e0.dataset]} "
+                f"{SHORT[e0.metric].split(' (')[0]} {e0.lora:.{nd}f} vs {e0.other:.{nd}f}.")
+    else:
+        fs_a = (f"E_N: scSemiProfiler's own error metric. LoRA-5 behind in {len(exc)} cohort × arm × metric cases; "
+                f"see notes.")
+    fs_a = S9.check_foot_short(fs_a)
+    rf_l = [few_mean(few, d, "lora", "c2st_rf_auc") for d in ORDER]
+    ctrl_auc = [few_mean(few, d, "control", "c2st_knn_auc") for d in ORDER]
+    ctrl_en = [few_mean(few, d, "control", "en_all_sq") for d in ORDER]
+    nn_c = cmp[cmp.metric.isin(["c2st_knn_auc", "c2st_rf_auc", "nn1_acc"] + [f"coverage_k{k}" for k in KS])]
+    wins_all = (nn_c.n_better == nn_c.n).all()
+    exc_txt = exceptions_text(exc)
+    en_means = "; ".join(f"{NAME[d]} {few_mean(few, d, 'scsemi5', 'en_all_sq'):.2f} vs "
+                         f"{few_mean(few, d, 'lora', 'en_all_sq'):.2f}" for d in ORDER)
+    foot_a = (f"Same 5-donor draws as the previous slides; every metric on exactly the scored cells (scores_extra/, "
+              f"tools/extra_metrics.py). Real side for AUC and MMD² = held-out real cells of the same samples (R2), "
+              f"projected like the generated cells; real vs real = R2 vs R3 (AUC {rng2(ctrl_auc)}). Classifier AUC: "
+              f"5-fold CV ROC AUC of a kNN classifier (k = 5) separating real from generated cells, 0.5 = "
+              f"indistinguishable; bars start at 0.5. MMD²: unbiased, Gaussian kernel, bandwidth = median distance "
+              f"among real cells. E_N: scSemiProfiler's normalized error (Wang et al. 2024, eq. 17–18), their code's "
+              f"squared L2 on all cells, per-arm PCA-100; real vs real = within-sample half split "
+              f"({rng2(ctrl_en)}); every generated arm is above 1. LoRA-5 best of the 4 arms in {nb_knn}/{ntot} "
+              f"draws on kNN AUC, {nb_rf}/{ntot} on RF AUC, {nb_nn1}/{ntot} on 1-NN accuracy, {nb_mmd}/{ntot} on MMD², "
+              f"{nb_en}/{ntot} on E_N. "
+              + ("LoRA-5 better than scSemi-5, zero-shot and zero-shot + calibration on both AUCs, 1-NN accuracy and "
+                 "coverage at k = 3, 5, 10 in every draw. " if wins_all else "")
+              + f"E_N scSemi-5 vs LoRA-5 (cohort means): {en_means}; LoRA-5 lower in {en_sc_better}/{en_sc_n} draws. "
+              f"Draws where LoRA-5 is worse (cohort means in brackets): {exc_txt if len(exc) else 'none'}. "
+              f"LoRA-5 AUC per draw "
+              f"{auc_draws.min():.2f}–{auc_draws.max():.2f} (kNN), RF cohort means {rng2(rf_l)}: a classifier still "
+              f"tells generated from real cells. Scores: scores_extra/<cohort>/fewshot_n5/draw<j>/.")
+    S9.check_text(foot_a, "footnote")
+    num("a", f"LoRA-5 best of 4 arms (draws): kNN AUC {nb_knn}/{ntot}, RF AUC {nb_rf}/{ntot}, 1-NN {nb_nn1}/{ntot}, "
+             f"MMD² {nb_mmd}/{ntot}, E_N {nb_en}/{ntot}", srcf)
+    num("a", f"E_N LoRA-5 < scSemi-5 in {en_sc_better}/{en_sc_n} draws; cohort means scSemi-5 vs LoRA-5: {en_means}",
+        f"{X}/<cohort>/fewshot_n5/draw<j>/en.csv (cellset all, distance sq)")
+    num("a", "LoRA-5 kNN AUC cohort means: " + ", ".join(f"{NAME[d]} {v:.3f}" for d, v in zip(ORDER, auc_l)) +
+        f"; per draw {auc_draws.min():.3f}–{auc_draws.max():.3f}", srcf)
+    num("a", "real vs real: kNN AUC " + ", ".join(f"{NAME[d]} {v:.3f}" for d, v in zip(ORDER, ctrl_auc)) +
+        "; E_N " + ", ".join(f"{NAME[d]} {v:.3f}" for d, v in zip(ORDER, ctrl_en)), srcf)
+    num("a", f"LoRA-5 worse than a comparator (draws): {exc_txt if len(exc) else 'none'}", srcf)
+    num("a", f"cross-check: {n_checked} cohort × arm × metric values equal scores_extra/{{fewshot,pooled}}_headline.csv",
+        f"{X}/fewshot_headline.csv; {X}/pooled_headline.csv")
+    slides.append({"title": "5 donors: realism beyond precision/coverage", "layout": "full",
+                   "figures": [str(FIGDIR / "sec09_extra_fewpair.png")], "keywords": kw_a, "footnote": foot_a,
+                   "footnote_short": fs_a, "numbers": NUMS["a"]})
+
+    # ---------------- slide b: 5-fold pooled ----------------
+    comps = [a for a in arms5 if a != "lora"]
+    first = {}
+    for c, _, sign in HEAD:
+        first[c] = [ds for ds in ORDER if all(better(pooled[ds]["lora"][c], pooled[ds][a][c], sign)
+                                              for a in comps if a in pooled[ds])]
+    n_all = sum(len(v) == len(ORDER) for v in first.values())
+    if n_all == len(HEAD):
+        kwb1 = f"Every metric: LoRA first in {len(ORDER)}/{len(ORDER)} cohorts"
+    else:
+        kwb1 = f"LoRA first everywhere on {n_all}/{len(HEAD)} metrics"
+    ji = pooled["ji2025_chili"]
+    oth = [ji[a]["en_all_sq"] for a in comps if a in ji]
+    kwb2 = S9.keyword_or_short(f"Ji whole blood: E_N {ji['lora']['en_all_sq']:.1f} vs {rng2(oth, 1)}",
+                               f"Ji: E_N {ji['lora']['en_all_sq']:.1f} vs {rng2(oth, 1)}")
+    auc_p = [pooled[d]["lora"]["c2st_knn_auc"] for d in ORDER]
+    kwb3 = f"Still separable: LoRA AUC {rng2(auc_p)}"
+    kw_b = S9.check_keywords([kwb1, kwb2, kwb3])
+    fs_b = S9.check_foot_short("Donor-level 5-fold CV, one pooled call per cohort. E_N: scSemiProfiler's own error "
+                               "metric.")
+    lost = {SHORT[c]: [NAME[d] for d in ORDER if d not in v] for c, v in first.items() if len(v) < len(ORDER)}
+    sc_txt = (f"scSemiProfiler, 5-fold CV (every training donor of the fold, inner-val included, as "
+              f"representatives; arm {SR.CV_ARM}) drawn for {', '.join(NAME[d] for d in has_sc)}; sources: "
+              + "; ".join(f"{NAME[d]} {pooled[d]['scsemi_all']['source']}" for d in has_sc) + ". "
+              if has_sc else "scSemiProfiler 5-fold CV not available yet (no extra-metric unit with set "
+                             f"{SR.CV_ARM}). ")
+    foot_b = (f"Donor-level 5-fold CV; each pair scored once by the fold model that never trained on its donor (one "
+              f"pooled call per cohort, all cross-fitted pairs). Arms as on the 5-fold slide. Metric definitions as on "
+              f"the 5-donor realism slide. LoRA first among the {len(comps) + 1} arms in all {len(ORDER)} cohorts on "
+              f"{n_all}/{len(HEAD)} headline metrics (kNN and RF AUC, 1-NN accuracy, MMD² at σ = m and summed, E_N all "
+              f"cells, precision and coverage at k = 3, 5, 10)"
+              + (f"; exceptions: " + "; ".join(f"{k}: {', '.join(v)}" for k, v in lost.items()) if lost else "")
+              + f". LoRA kNN AUC " + ", ".join(f"{NAME[d]} {v:.3f}" for d, v in zip(ORDER, auc_p))
+              + f"; real vs real " + ", ".join(f"{NAME[d]} {pooled[d]['control']['c2st_knn_auc']:.3f}" for d in ORDER)
+              + ". E_N LoRA vs other arms: " + "; ".join(
+                  f"{NAME[d]} {pooled[d]['lora']['en_all_sq']:.2f} vs {rng2([pooled[d][a]['en_all_sq'] for a in comps if a in pooled[d]])}"
+                  for d in ORDER)
+              + f". {sc_txt}Scores: scores_extra/<cohort>/pooled/.")
+    S9.check_text(foot_b, "footnote")
+    num("b", f"LoRA first among {len(comps) + 1} arms in all {len(ORDER)} cohorts on {n_all}/{len(HEAD)} metrics", srcp)
+    num("b", f"Ji E_N: LoRA {ji['lora']['en_all_sq']:.3f}; " +
+        ", ".join(f"{LAB[a]} {ji[a]['en_all_sq']:.3f}" for a in comps if a in ji), srcp)
+    num("b", "LoRA kNN AUC (pooled): " + ", ".join(f"{NAME[d]} {v:.3f}" for d, v in zip(ORDER, auc_p)), srcp)
+    num("b", "scSemiProfiler, 5-fold CV: " + (", ".join(NAME[d] for d in has_sc) if has_sc else "not available"),
+        OUTDATA / "sec09_extra_scsemi_status.csv")
+    slides.append({"title": "5-fold: realism beyond precision/coverage", "layout": "full",
+                   "figures": [str(FIGDIR / "sec09_extra_fivefold.png")], "keywords": kw_b, "footnote": foot_b,
+                   "footnote_short": fs_b, "numbers": NUMS["b"]})
+
+    # ---------------- slide c: k sensitivity ----------------
+    sets_k = FEW_ARMS + ["control"]
+    mono = all(few_mean(few, d, a, f"{q}_k3") <= few_mean(few, d, a, f"{q}_k5") <= few_mean(few, d, a, f"{q}_k10")
+               for d in ORDER for a in sets_k for q in ("coverage", "precision"))
+    ctrl_c3 = [few_mean(few, d, "control", "coverage_k3") for d in ORDER]
+    kwc1 = ("Smaller k lowers every set, real included" if mono else
+            f"Real vs real coverage at k = 3: {rng2(ctrl_c3)}")
+    cov_all_k = 0
+    wc = {k: few_wide(few, f"coverage_k{k}") for k in KS}
+    for idx in wc[10].index:
+        cov_all_k += all(all(better(wc[k].loc[idx, "lora"], wc[k].loc[idx, a], 1) for a in FEW_ARMS if a != "lora")
+                         for k in KS)
+    kwc2 = f"Coverage: LoRA-5 first at all k, {cov_all_k}/{len(wc[10])}"
+    pb = [best_draws(few, f"precision_k{k}")[0] for k in KS]
+    kwc3 = f"Precision: LoRA-5 first in {min(pb)}–{max(pb)}/{ntot} draws" if min(pb) != max(pb) else \
+        f"Precision: LoRA-5 first in {pb[0]}/{ntot} draws"
+    kw_c = S9.check_keywords([kwc1, kwc2, kwc3])
+    fs_c = S9.check_foot_short("5 donors, 5 draws per cohort; error bars = SD over draws. k = 10 is the value on the main "
+                               "slides.")
+    # who beats LoRA-5 on precision, per k
+    who_n, who_arms = {}, set()
+    for k in KS:
+        w = few_wide(few, f"precision_k{k}")
+        for (d, j), r in w.iterrows():
+            b = [a for a in FEW_ARMS if a != "lora" and not better(r["lora"], r[a], 1)]
+            if b:
+                who_n.setdefault(d, {kk: 0 for kk in KS})[k] += 1
+                who_arms.update(b)
+    who = ", ".join(f"{NAME[d]} {'/'.join(str(who_n[d][k]) for k in KS)}" for d in ORDER if d in who_n)
+    who = (f"{who} draws at k = 3/5/10, beaten by {' or '.join(S9.LAB[a].lower() for a in sorted(who_arms))}"
+           if who else "none")
+    exp_cov = {}
+    for d in ORDER:
+        mt = pd.concat([pd.read_csv(u / "metrics.csv") for u in sorted((X / d / "fewshot_n5").glob("draw*"))])
+        mt = mt[(mt["set"] == "lora5") & (mt.real_side == "R1") & (mt["mode"] == "raw")]
+        exp_cov[d] = "/".join(f"{mt[f'expected_coverage_identical_k{k}'].mean():.2f}" for k in KS)
+    exp_cov = (f"{next(iter(exp_cov.values()))} in every cohort" if len(set(exp_cov.values())) == 1
+               else "; ".join(f"{NAME[d]} {v}" for d, v in exp_cov.items()))
+    same_rank, lora_top = [], []
+    for d in ORDER:
+        r = {k: sorted([a for a in arms5 if a in pooled[d]], key=lambda a: -pooled[d][a][f"coverage_k{k}"]) for k in KS}
+        same_rank.append(r[3] == r[5] == r[10])
+        lora_top.append(all(r[k][0] == "lora" for k in KS))
+    pooled_rank = (f"{'unchanged' if all(same_rank) else 'changes in ' + ', '.join(NAME[d] for d, x in zip(ORDER, same_rank) if not x)}"
+                   f"; LoRA first at every k in {sum(lora_top)}/{len(ORDER)} cohorts")
+    foot_c = (f"Precision and coverage recomputed with k = 3, 5, 10 nearest-neighbour balls in the scorer's PCA-50 "
+              f"(k = 10 reproduces the main-slide values exactly). Means over the 5 draws per cohort; error bars = SD; "
+              f"grey = real vs real (R1 vs R2, same samples). Both metrics fall with smaller k for every set "
+              f"{'including' if mono else 'but not always'} the real control (real coverage at k = 3: {rng2(ctrl_c3)}), "
+              f"so compare arms with the real row, not with 1. Expected coverage of a perfect generator (Naeem et al. "
+              f"2020, eq. 9) at k = 3/5/10: {exp_cov}. LoRA-5 first on coverage at every k in "
+              f"{cov_all_k}/{len(wc[10])} draws; on precision in {'/'.join(map(str, pb))} of {ntot} draws "
+              f"(k = 3/5/10); not first: {who}. 5-fold pooled design (not drawn): "
+              f"coverage ranking of the {len(arms5)} arms across k {pooled_rank}. Scores: "
+              f"scores_extra/<cohort>/fewshot_n5/draw<j>/metrics.csv.")
+    S9.check_text(foot_c, "footnote")
+    num("c", f"LoRA-5 first on coverage at k = 3, 5 and 10: {cov_all_k}/{len(wc[10])} draws", srcf)
+    num("c", f"LoRA-5 first on precision (draws) at k = 3/5/10: {'/'.join(map(str, pb))} of {ntot}", srcf)
+    num("c", "real vs real coverage at k = 3: " + ", ".join(f"{NAME[d]} {v:.3f}" for d, v in zip(ORDER, ctrl_c3)), srcf)
+    num("c", f"every set falls with smaller k (cohort means): {mono}", srcf)
+    slides.append({"title": "5 donors: precision/coverage at k = 3, 5, 10", "layout": "full",
+                   "figures": [str(FIGDIR / "sec09_extra_ksens.png")], "keywords": kw_c, "footnote": foot_c,
+                   "footnote_short": fs_c, "numbers": NUMS["c"]})
+
+    for s in slides:
+        for k in ("title", "footnote", "footnote_short"):
+            S9.check_text(s[k], k)
+        for n in s["numbers"]:
+            S9.check_text(n["value"], "quoted number")
+    spec = {"section": "09-extra", "slides": slides, "scripts": [str(SCRIPT)],
+            "generated_from": "tools/sec09_extra_metrics.py (rerun to refresh; do not edit by hand)",
+            "scsemi_cv_cohorts": [NAME[d] for d in has_sc]}
+    (OUTDATA / "sec09_extra_slide_text.json").write_text(json.dumps(spec, indent=1, ensure_ascii=False))
+    pd.DataFrame([{"slide": k, "value": n["value"], "source": n["source"]} for k in "abc" for n in NUMS[k]]).to_csv(
+        OUTDATA / "sec09_extra_quoted_numbers.csv", index=False)
+    for p in [OUTDATA / "sec09_extra_slide_text.json"] + sorted(OUTDATA.glob("sec09_extra_*.csv")):
+        m = S9.OFF_STORY.search(p.read_text(encoding="utf-8"))
+        if m:
+            raise ValueError(f"{p}: off-story text {m.group(0)!r}")
+
+    print(f"cross-checked {n_checked} values against the collector's headline tables")
+    if not has_sc:
+        print(f"NOTE: no scSemiProfiler 5-fold extra-metric unit (set {SR.CV_ARM}) under "
+              f"{SR.plan_root()}/scores_extra/<ds>/{SR.CV_EXP}/ or in the pooled units; the 5-fold figure has no "
+              f"scSemiProfiler row.")
+    else:
+        print(f"scSemiProfiler, 5-fold CV drawn for: {[NAME[d] for d in has_sc]}")
+    for r in sclog.itertuples():
+        print(f"  scSemiProfiler extra unit {r.dataset} {r.design} {r.set}: {r.status} {r.reason} ({r.unit})")
+    for s in slides:
+        print(f"\n== {s['title']}")
+        for k in s["keywords"]:
+            print("  *", k, f"({S9.n_words(k)} words, {len(k)} chars)")
+        print("  footnote_short:", s["footnote_short"], f"({len(s['footnote_short'])} chars)")
+        print("  footnote:", s["footnote"])
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())