Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261012T130738Z-strictpairs-implementation/diffusion/dataset/server_1/model8/plans/paper_figures_strictpairs_20261012/archive/20261012T130738Z-strictpairs-implementation/snapshot/CAMPAIGN_STATUS.json

@@ -0,0 +1,10 @@
+{
+  "state": "running",
+  "stage": "global",
+  "started_at": "2026-10-12T12:33:47.457996+00:00",
+  "units_selected": 40,
+  "units_completed_this_invocation": 0,
+  "figure6": "waiting_user_data_no_execution",
+  "active": "global-imperial_eskd",
+  "updated_at": "2026-10-12T12:45:13.986209+00:00"
+}

paper-releases/pbisc-diffusion/updates/20261012T130738Z-strictpairs-implementation/diffusion/dataset/server_1/model8/plans/paper_figures_strictpairs_20261012/archive/20261012T130738Z-strictpairs-implementation/snapshot/PROGRESS.md

@@ -0,0 +1,43 @@
+# Implementation ledger — paper_figures_strictpairs_20261012
+
+## State (updated 2026-10-12 12:28 UTC)
+- Implementation and guarded execution are in progress; the full scientific campaign is not complete.
+- Source Model8 Git has no commits and existing staged/untracked research assets. All new output is isolated in this versioned plan root; only opt-in legacy adapters may be changed with byte backups.
+- Canonical manuscript is clean at `fb6380f` as observed on 2026-10-12. Other authorized work changed the Abstract during this task; preserve the latest approved frontmatter when assembling the isolated manuscript.
+
+## Work packages
+1. Strict support selection and opt-in legacy adapters: implemented; 40 support units prepared and tested.
+2. Figure1 concept/corpus: complete with SVG/PDF/PNG and source provenance.
+3. Architecture/panel rescoring: eight arms complete on the shared held-out reference; separate sweep floors retained.
+4. Campaign: ImMicro fold0 n5 LoRA training, LoRA/pairedcal/scSemi generation and two all-ImMicro raw/refcal controls complete. The first full five-arm score is running. Remaining strict units await this integration gate.
+5. Figure3–5 visual/statistical pipeline: implemented; strict final outputs await upstream generation/scoring. Historical rare results are explicitly diagnostic.
+6. Review deck and staged manuscript PDF: produced with pending labels. Final rendering/export integration and archive snapshot preparation are in progress. Canonical manuscript is not modified by this task.
+
+Latest completed verification: 93 focused CPU tests passed (2026-10-12). This does not certify unexecuted full-cohort analyses.
+
+## Interface preflight
+| Producer | Consumer | Binding |
+|---|---|---|
+| Support manifests | LoRA/calibration/scSemi/deconvolution | exact sample IDs, unique donors, no test donor or extra visit access |
+| Rescoring/campaign | Figures/Results | source-bound typed metrics and readiness, no mixing old n+validation with strict-n |
+| Figure1 | Methods/other panels | K_cond=1000; rho=3869.603120208939; N_gen=1500 distinct |
+| Rare metrics | Figure5 | include zeros, patient IDs, mandatory marker availability, descriptive Behcet |
+
+Ruling: use a new versioned workspace within the registered Model8 core because its source Git is unborn and has substantial pre-existing research assets; do not stage/commit those assets or clone a misleading incomplete tree. Manuscript changes will be staged as an isolated copy and reviewed before safe integration.
+
+## Prelaunch review correction
+- Fixed inconsistent Imperial date axis (C137 draw4 now2020-04-30), strict QC-excluded visit validation, complete trainer resume binding, runtime outputSHAverification.
+- Restored oldscSemi wrapper byte-exact; added opt-in strictwrapper.
+- Preserved prelaunch generated support/conditions at backups/preflight_rejected_v1; never used for GPU work.
+- Figure1 produced+visually checked; pretrain P/C8arms rescoredCPU; source tables ready.
+
+## Execution checkpoint — 2026-10-12 13:02 UTC
+- First complete five-arm healthy unit (`immicro/healthy/fold0/n5`) passed generation, scoring, extras and donor bootstrap. This is 1/20 healthy learning-curve units and 1/5 Figure3 folds, not a final pooled claim.
+- Sepsis scSemi all five existing exact-five runs were audited and adopted without new GPU training: same 5 representatives, 77 targets, own-n HVG, equivalent prepared cells, model/shard/output SHA checks, independent shard-level export validation. Embedded legacy metadata remains unchanged. See `reports/deconv_scsemi_reuse_audit.md`.
+- Eight pretraining arms now have common-readout composition scores in addition to P/C. Full NPZ hashes and frozen-annotation concordance were verified; neither 6k nor r16 is described as best on every endpoint.
+- Historical healthy rank/full now uses only real-bulk FT with the same outer-fold train/validation/test IDs; strict size curves contain all five comparator arms. Zero-target-pair baselines are explicit.
+- Disease main statistics use patient-aware contrasts and shared patient bootstrap across draws. Old visit-level sensitivity stays separate. Figure4 historical comparison uses actual real-bulk LoRA and scSemi five-fold sources, including Giroux exclusions.
+- Figure3 has common cell-type-colored UMAP panels, full-cell composition fractions, and separate adaptation/generation times. Figure5 has all-state detection/abundance tables and separate Behcet any-C1Q vs multi-C1Q plots; Behcet calibration uses the same three candidate pairs, not external AIFI.
+- `pbisc-strict-global-controls-20261012.service` is generating disease controls sequentially. `pbisc-strict-full-campaign-20261012.service` is queued behind its controller lock, then runs all 40 units, CPU derivations, gated PDFs/PPTX, and a one-time verified final archive delta. Failures stop for inspection; no OOM/restart loop.
+- Latest full test run: 134 passed. Reviewed primary disease statistics/adoption provenance have no outstanding P1/P2 from the final narrow review. Full disease/rare and healthy multi-fold scientific results remain pending execution.
+- Canonical manuscript remains unchanged by this task. Staging refreshes the latest approved frontmatter. Figure6 has no analysis or experiment launched. No GitHub push/Overleaf sync.

paper-releases/pbisc-diffusion/updates/20261012T130738Z-strictpairs-implementation/diffusion/dataset/server_1/model8/plans/paper_figures_strictpairs_20261012/archive/20261012T130738Z-strictpairs-implementation/snapshot/events.jsonl

@@ -0,0 +1,77 @@
+{"at": "2026-10-12T10:45:14.083923+00:00", "operation": "prepare-external-controls", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T10:45:15.467429+00:00", "operation": "prepare-external-controls", "state": "completed", "seconds": 1.3822942541446537}
+{"at": "2026-10-12T10:45:15.519881+00:00", "operation": "prepare-strict-conditions", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T10:45:19.800949+00:00", "operation": "prepare-strict-conditions", "state": "completed", "seconds": 4.280744560994208}
+{"at": "2026-10-12T10:53:34.411976+00:00", "operation": "prepare-strict-conditions-v2", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T10:53:38.844668+00:00", "operation": "prepare-strict-conditions-v2", "state": "completed", "seconds": 4.432218081085011}
+{"at": "2026-10-12T10:56:16.826391+00:00", "operation": "immicro-healthy-fold0-n5-train", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T10:58:18.438144+00:00", "operation": "immicro-healthy-fold0-n5-train", "state": "completed", "seconds": 121.50741662294604}
+{"at": "2026-10-12T10:58:18.492261+00:00", "operation": "immicro-healthy-fold0-n5-lora", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T10:59:46.264460+00:00", "operation": "immicro-healthy-fold0-n5-lora", "state": "completed", "seconds": 87.37215070892125}
+{"at": "2026-10-12T10:59:46.330009+00:00", "operation": "immicro-healthy-fold0-n5-pairedcal", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T11:01:12.194662+00:00", "operation": "immicro-healthy-fold0-n5-pairedcal", "state": "completed", "seconds": 85.47015835391358}
+{"at": "2026-10-12T11:02:35.126883+00:00", "operation": "immicro-healthy-fold0-n5-train", "state": "binding_upgraded_after_review", "cells_sha256": "e1168468b621183604038b79280f46961c5819fe2386d9da0b454f7db642742a"}
+{"at": "2026-10-12T11:05:12.203135+00:00", "operation": "immicro-healthy-fold0-n5-scinputs", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T11:05:14.052055+00:00", "operation": "immicro-healthy-fold0-n5-scinputs", "state": "completed", "seconds": 1.8483133830595762}
+{"at": "2026-10-12T11:05:14.100941+00:00", "operation": "immicro-healthy-fold0-n5-scprepare", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T11:05:42.081655+00:00", "operation": "immicro-healthy-fold0-n5-scprepare", "state": "completed", "seconds": 27.97998751490377}
+{"at": "2026-10-12T11:05:42.136933+00:00", "operation": "immicro-healthy-fold0-n5-scpretrain-0", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T11:06:34.938737+00:00", "operation": "pilot-score-diffusion-only", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T11:06:53.783449+00:00", "operation": "pilot-score-diffusion-only", "state": "completed", "seconds": 18.844257805030793}
+{"at": "2026-10-12T11:14:24.459568+00:00", "operation": "immicro-healthy-fold0-n5-scpretrain-0", "state": "completed", "seconds": 522.3221146850847}
+{"at": "2026-10-12T11:14:24.514474+00:00", "operation": "immicro-healthy-fold0-n5-scpretrain-2", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T11:17:40.865068+00:00", "operation": "immicro-healthy-fold0-n5-scpretrain-2", "state": "completed", "seconds": 196.35023813904263}
+{"at": "2026-10-12T11:17:40.912011+00:00", "operation": "immicro-healthy-fold0-n5-scpretrain-4", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T11:21:26.371024+00:00", "operation": "immicro-healthy-fold0-n5-scpretrain-4", "state": "completed", "seconds": 225.45824180706404}
+{"at": "2026-10-12T11:21:26.412306+00:00", "operation": "immicro-healthy-fold0-n5-scinfer-0", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T11:31:42.831264+00:00", "operation": "immicro-healthy-fold0-n5-scinfer-0", "state": "completed", "seconds": 616.4185961470939}
+{"at": "2026-10-12T11:31:42.897066+00:00", "operation": "immicro-healthy-fold0-n5-scinfer-8", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T11:40:46.096833+00:00", "operation": "immicro-healthy-fold0-n5-scinfer-8", "state": "completed", "seconds": 543.1994235909078}
+{"at": "2026-10-12T11:40:46.149541+00:00", "operation": "immicro-healthy-fold0-n5-scinfer-16", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T11:46:48.753305+00:00", "operation": "immicro-healthy-fold0-n5-scinfer-16", "state": "completed", "seconds": 362.60319669591263}
+{"at": "2026-10-12T11:46:48.795304+00:00", "operation": "immicro-healthy-fold0-n5-scexport", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T11:46:54.280606+00:00", "operation": "immicro-healthy-fold0-n5-scexport", "state": "completed", "seconds": 5.4691027260851115}
+{"at": "2026-10-12T11:46:54.326996+00:00", "operation": "immicro-healthy-fold0-n5-scvalidate", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T11:46:57.583484+00:00", "operation": "immicro-healthy-fold0-n5-scvalidate", "state": "completed", "seconds": 3.2557712129782885}
+{"at": "2026-10-12T11:52:30.409963+00:00", "operation": "immicro-healthy-fold0-n5-scexport", "state": "entrypoint_binding_verified"}
+{"at": "2026-10-12T11:53:34.784746+00:00", "operation": "pilot-deconv-prepare-sepsis0", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T11:54:08.678003+00:00", "operation": "pilot-deconv-prepare-sepsis0", "state": "completed", "seconds": 33.892768485005945}
+{"at": "2026-10-12T11:56:07.952863+00:00", "operation": "global-immicro-raw", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T12:03:37.604803+00:00", "operation": "global-immicro-raw", "state": "completed", "seconds": 447.602641884936}
+{"at": "2026-10-12T12:03:37.719304+00:00", "operation": "global-immicro-refcal", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T12:10:54.387319+00:00", "operation": "global-immicro-refcal", "state": "completed", "seconds": 434.7524452728685}
+{"at": "2026-10-12T12:24:53.722073+00:00", "operation": "immicro-healthy-fold0-n5-score", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T12:31:35.344883+00:00", "operation": "immicro-healthy-fold0-n5-score", "state": "completed", "seconds": 401.62207144801505}
+{"at": "2026-10-12T12:33:49.577517+00:00", "operation": "global-immicro-raw", "state": "verified_skip"}
+{"at": "2026-10-12T12:33:51.458420+00:00", "operation": "global-immicro-refcal", "state": "verified_skip"}
+{"at": "2026-10-12T12:33:51.547071+00:00", "operation": "global-ye2026_sepsis-raw", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T12:39:55.266132+00:00", "operation": "global-ye2026_sepsis-raw", "state": "completed", "seconds": 362.0762440508697}
+{"at": "2026-10-12T12:39:55.362317+00:00", "operation": "global-ye2026_sepsis-refcal", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T12:45:13.985921+00:00", "operation": "global-ye2026_sepsis-refcal", "state": "completed", "seconds": 317.21102344593965}
+{"at": "2026-10-12T12:45:14.111139+00:00", "operation": "global-imperial_eskd-raw", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T12:47:28.111830+00:00", "operation": "figure2-reviewed-preview-v2", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T12:47:29.868899+00:00", "operation": "figure2-reviewed-preview-v2", "state": "completed", "seconds": 1.7551806829869747}
+{"at": "2026-10-12T12:48:11.505500+00:00", "operation": "figure5-expanded-diagnostic-v2", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T12:48:16.941174+00:00", "operation": "figure5-expanded-diagnostic-v2", "state": "completed", "seconds": 5.422949705040082}
+{"at": "2026-10-12T12:55:13.220965+00:00", "operation": "immicro-healthy-fold0-n5-score", "state": "native_validation_signature_migrated", "previous_signature": "b13e46f90317e3f06fd6fb5c6fed206472c5cf498ba25c862342bf8e8b6fd770"}
+{"at": "2026-10-12T12:59:05.407255+00:00", "operation": "figures234-reviewed-preview-v3", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T12:59:08.798280+00:00", "operation": "figures234-reviewed-preview-v3", "state": "completed", "seconds": 3.389978000894189}
+{"at": "2026-10-12T12:59:08.850592+00:00", "operation": "behcet-rule-label-reviewed-v3", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T12:59:09.641226+00:00", "operation": "behcet-rule-label-reviewed-v3", "state": "completed", "seconds": 0.7895921289455146}
+{"at": "2026-10-12T12:59:09.693352+00:00", "operation": "deck-reviewed-preview-v3", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T12:59:10.541832+00:00", "operation": "deck-reviewed-preview-v3", "state": "completed", "seconds": 0.8453024930786341}
+{"at": "2026-10-12T12:59:36.184000+00:00", "operation": "global-imperial_eskd-raw", "state": "completed", "seconds": 858.9609018329065}
+{"at": "2026-10-12T12:59:36.306078+00:00", "operation": "global-imperial_eskd-refcal", "state": "running", "resource": "gpu"}
+{"at": "2026-10-12T13:01:05.931439+00:00", "operation": "manuscript-current-frontmatter-preview-v3", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T13:01:06.089599+00:00", "operation": "manuscript-current-frontmatter-preview-v3", "state": "completed", "seconds": 0.1570085440762341}
+{"at": "2026-10-12T13:01:06.152842+00:00", "operation": "compile-review-pdfs-v3", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T13:01:15.406023+00:00", "operation": "compile-review-pdfs-v3", "state": "completed", "seconds": 9.24971738201566}
+{"at": "2026-10-12T13:03:25.163537+00:00", "operation": "verify-cpu-final", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T13:04:03.055577+00:00", "operation": "verify-cpu-final", "state": "completed", "seconds": 37.88490461395122}
+{"at": "2026-10-12T13:05:19.299649+00:00", "operation": "verify-cpu-final-with-r", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T13:05:55.368337+00:00", "operation": "verify-cpu-final-with-r", "state": "completed", "seconds": 36.068204927956685}
+{"at": "2026-10-12T13:07:22.571134+00:00", "operation": "deck-readable-ablation-preview-v4", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T13:07:23.313441+00:00", "operation": "deck-readable-ablation-preview-v4", "state": "completed", "seconds": 0.7396434370893985}
+{"at": "2026-10-12T13:07:23.366765+00:00", "operation": "convert-readable-review-deck-v4", "state": "running", "resource": "cpu"}
+{"at": "2026-10-12T13:07:28.757394+00:00", "operation": "convert-readable-review-deck-v4", "state": "completed", "seconds": 5.388529522111639}
+{"at": "2026-10-12T13:07:38.190161+00:00", "operation": "archive-implementation-snapshot", "state": "running", "resource": "cpu"}

paper-releases/pbisc-diffusion/updates/20261012T130738Z-strictpairs-implementation/diffusion/dataset/server_1/model8/plans/paper_figures_strictpairs_20261012/archive/20261012T130738Z-strictpairs-implementation/snapshot/legacy_adapters/plans/paired_finetune_intradataset_20260927/tools/ft_sdd.py

@@ -0,0 +1,496 @@
+#!/usr/bin/env python3
+"""Fine-tune the final PBISC-SDD model inside one paired dataset fold.
+
+Training pairs are (condition of row r, one retained cell of row r).  The condition comes from a condition NPZ
+written by prepare_dataset.py (real bulk or pseudobulk, fixed-depth policy) and goes through the checkpoint's own
+conditioner exactly as the standalone sampler does (sample_sdd_standalone.load_conditions), then
+condition_with_measurement_context(base, bulk_measurement_fraction, output_availability_mask).  The loss is the
+pretraining loss (SparseDataDiffusion.training_loss: measured-only value L2 + support BCE, self-conditioning),
+with bf16 autocast, AdamW and gradient clipping as in pretraining.
+
+Rows: fold k's test rows are never loaded.  Training rows = the other folds minus the inner-validation rows;
+inner-validation rows give a fixed-draw diffusion loss every --eval-every steps, and the EMA state with the lowest
+inner-validation loss is kept (early stopping with --patience evaluations).
+
+Modes (what is trainable):
+  full     every denoiser parameter
+  condenc  only denoiser.condition_encoder (the bulk -> context map)
+  lora     low-rank updates W = W0 + (alpha/r) B A on every nn.Linear of the denoiser (B starts at zero);
+           merged into the weights before saving
+
+Batches are donor-balanced: a row is drawn uniformly, then one of its cells uniformly.
+
+The output checkpoint keeps the input checkpoint's schema and self-contained keys, with model_state and
+ema_state.model_state set to the selected weights, so scripts/sample_sdd_standalone.py samples it unchanged.
+src/pbisc_sdd is sealed and is only imported.
+"""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+import os
+import sys
+import time
+from datetime import datetime, timezone
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+import torch
+from torch import nn
+import torch.nn.utils.parametrize as parametrize
+
+M8 = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion")
+for p in (M8 / "src", M8 / "scripts"):
+    if str(p) not in sys.path:
+        sys.path.insert(0, str(p))
+import sample_sdd_standalone as SS  # noqa: E402
+from pbisc_sdd.data import condition_with_measurement_context  # noqa: E402
+
+SCHEMA = "pbisc_sdd.ft_intradataset.run.v1"
+KEEP_KEYS = ("schema_version", "model_config", "diffusion_config", "value_scaler", "conditioner_state",
+             "gene_positions", "gene_ids", "gene_symbols", "gene_measurement_coverage",
+             "full_gene_measurement_coverage", "corpus_fingerprint", "experiment_config", "proportion_hint",
+             "training_contract")
+
+
+def sha256_file(path: Path) -> str:
+    h = hashlib.sha256()
+    with open(path, "rb") as f:
+        for block in iter(lambda: f.read(1 << 24), b""):
+            h.update(block)
+    return h.hexdigest()
+
+
+def log(msg: str) -> None:
+    print(f"{datetime.now().isoformat(timespec='seconds')} {msg}", flush=True)
+
+
+# ----------------------------------------------------------------------------------------------- LoRA
+class LoRADelta(nn.Module):
+    """Parametrization W -> W + scale * B @ A (B zero-initialised, so the start equals the pretrained W)."""
+
+    def __init__(self, out_features: int, in_features: int, rank: int, alpha: float) -> None:
+        super().__init__()
+        self.A = nn.Parameter(torch.empty(rank, in_features))
+        self.B = nn.Parameter(torch.zeros(out_features, rank))
+        nn.init.kaiming_uniform_(self.A, a=math.sqrt(5))
+        self.scale = float(alpha) / float(rank)
+
+    def forward(self, weight: torch.Tensor) -> torch.Tensor:
+        return weight + self.scale * (self.B @ self.A).to(dtype=weight.dtype)
+
+
+def add_lora(model: nn.Module, rank: int, alpha: float) -> list[str]:
+    names = []
+    for name, module in model.named_modules():
+        if isinstance(module, nn.Linear):
+            delta = LoRADelta(module.out_features, module.in_features, rank, alpha).to(
+                device=module.weight.device, dtype=module.weight.dtype)
+            parametrize.register_parametrization(module, "weight", delta)
+            names.append(name)
+    return names
+
+
+def merge_lora(model: nn.Module) -> None:
+    for module in model.modules():
+        if isinstance(module, nn.Linear) and parametrize.is_parametrized(module, "weight"):
+            parametrize.remove_parametrizations(module, "weight", leave_parametrized=True)
+
+
+def set_trainable(denoiser: nn.Module, mode: str, rank: int, alpha: float) -> dict:
+    for p in denoiser.parameters():
+        p.requires_grad_(False)
+    info: dict = {"mode": mode}
+    if mode == "full":
+        for p in denoiser.parameters():
+            p.requires_grad_(True)
+    elif mode == "condenc":
+        for p in denoiser.condition_encoder.parameters():
+            p.requires_grad_(True)
+    elif mode == "lora":
+        info["lora_layers"] = add_lora(denoiser, rank, alpha)
+        info["lora_rank"], info["lora_alpha"] = rank, alpha
+        for name, p in denoiser.named_parameters():
+            p.requires_grad_(name.endswith(".A") or name.endswith(".B"))
+    else:
+        raise SystemExit(f"unknown mode {mode}")
+    trainable = [n for n, p in denoiser.named_parameters() if p.requires_grad]
+    info["n_trainable"] = int(sum(p.numel() for p in denoiser.parameters() if p.requires_grad))
+    info["n_total_base"] = int(sum(p.numel() for n, p in denoiser.named_parameters()
+                                   if not (n.endswith(".A") or n.endswith(".B"))))
+    info["trainable_tensors"] = len(trainable)
+    return info
+
+
+# ----------------------------------------------------------------------------------------------- data
+class Cells:
+    """Retained cells on the 6k panel (CSR, rows sorted by dataset row)."""
+
+    def __init__(self, path: Path, panel_gene_ids: np.ndarray) -> None:
+        z = np.load(path, allow_pickle=False)
+        self.indptr = z["indptr"].astype(np.int64)
+        self.indices = z["indices"].astype(np.int64)
+        self.data = z["data"].astype(np.float32)
+        self.row_sample = z["row_sample"].astype(np.int64)
+        self.n_genes = int(z["shape"][1])
+        if not np.array_equal(z["panel_gene_ids"].astype(str), panel_gene_ids.astype(str)):
+            raise SystemExit("cells_panel6k columns differ from the checkpoint gene panel")
+        if bool((np.diff(self.row_sample) < 0).any()):
+            raise SystemExit("cells are not sorted by dataset row")
+        n_rows = int(self.row_sample.max()) + 1
+        self.start = np.searchsorted(self.row_sample, np.arange(n_rows), side="left")
+        self.count = np.searchsorted(self.row_sample, np.arange(n_rows), side="right") - self.start
+
+    def dense(self, cells: np.ndarray) -> np.ndarray:
+        s, e = self.indptr[cells], self.indptr[cells + 1]
+        lens = e - s
+        idx = np.concatenate([np.arange(a, b) for a, b in zip(s, e)]) if len(cells) else np.zeros(0, np.int64)
+        out = np.zeros((len(cells), self.n_genes), dtype=np.float32)
+        out[np.repeat(np.arange(len(cells)), lens), self.indices[idx]] = self.data[idx]
+        return out
+
+
+def draw_batch(cells: Cells, rows: np.ndarray, batch: int, rng: np.random.Generator) -> tuple[np.ndarray, np.ndarray]:
+    r = rows[rng.integers(0, len(rows), size=batch)]
+    c = cells.start[r] + np.floor(rng.random(batch) * cells.count[r]).astype(np.int64)
+    return r, c
+
+
+# ----------------------------------------------------------------------------------------------- main
+def main(argv=None) -> int:
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--checkpoint", type=Path, required=True)
+    ap.add_argument("--checkpoint-sha", required=True)
+    ap.add_argument("--dataset-dir", type=Path, required=True)
+    ap.add_argument("--condition-npz", type=Path, required=True, help="all-row condition NPZ used for training")
+    ap.add_argument("--fold", type=int, required=True)
+    ap.add_argument("--mode", choices=("full", "condenc", "lora"), required=True)
+    ap.add_argument("--lora-rank", type=int, default=16)
+    ap.add_argument("--lora-alpha", type=float, default=16.0)
+    ap.add_argument("--lr", type=float, required=True)
+    ap.add_argument("--warmup", type=int, default=200)
+    ap.add_argument("--batch-size", type=int, default=256)
+    ap.add_argument("--max-steps", type=int, default=20000)
+    ap.add_argument("--eval-every", type=int, default=250)
+    ap.add_argument("--patience", type=int, default=8)
+    ap.add_argument("--ema-decay", type=float, default=0.999)
+    ap.add_argument("--val-cells-per-row", type=int, default=256)
+    ap.add_argument("--val-repeats", type=int, default=2)
+    ap.add_argument("--seed", type=int, default=20260927)
+    ap.add_argument("--out-dir", type=Path, required=True)
+    ap.add_argument("--device", default="cuda")
+    ap.add_argument("--max-cuda-gib", type=float, default=8.0)
+    ap.add_argument("--min-system-gib", type=float, default=32.0)
+    ap.add_argument("--smoke", action="store_true", help="tiny run for tests (few steps, CPU allowed)")
+    ap.add_argument("--train-n", type=int, default=0, help="learning curve: number of training donors to keep")
+    ap.add_argument("--train-donors-file", default="", help="few-pair scenario: train on these donor_ids (one per "
+                    "line), no inner validation, every other usable row is test; --fold is ignored")
+    ap.add_argument("--fixed-steps", type=int, default=0, help="train exactly this many steps without inner "
+                    "validation and keep the final EMA state (for when no pairs can be spared for early stopping)")
+    ap.add_argument("--subset-seed", type=int, default=0)
+    ap.add_argument("--strata-regex", default="", help="groups matching this regex form one stratum, the rest another")
+    ap.add_argument("--support-manifest", type=Path, default=None,
+                    help="strict paired support.json: exact sample IDs, no inner validation; requires --fixed-steps 3000")
+    a = ap.parse_args(argv)
+
+    if a.support_manifest is not None:
+        if a.mode != "lora" or a.train_donors_file or a.train_n or a.fixed_steps != 3000 or a.smoke:
+            raise SystemExit("strict support requires LoRA, --fixed-steps 3000, no donor/subset override or smoke")
+        strict_tools = M8 / "plans/paper_figures_strictpairs_20261012/tools"
+        sys.path.insert(0, str(strict_tools))
+        from support_manifest import load_manifest, validate_existing_run
+        support_preflight = load_manifest(a.support_manifest, pd.read_csv(a.dataset_dir / "folds.csv",
+                                                                           dtype={"sample_id": str, "donor_id": str}))
+        if support_preflight["cohort"] != a.dataset_dir.name:
+            raise SystemExit("support cohort differs from dataset directory")
+    a.out_dir.mkdir(parents=True, exist_ok=True)
+    done = a.out_dir / "checkpoint_ft.pt"
+    if done.exists() and (a.out_dir / "RUN.json").exists():
+        if a.support_manifest is not None:
+            stored = json.loads((a.out_dir / "RUN.json").read_text())
+            validate_existing_run(support_preflight, stored, fixed_steps=a.fixed_steps,
+                                  lora_rank=a.lora_rank, source_checkpoint_sha256=a.checkpoint_sha,
+                                  condition_npz_sha256=sha256_file(a.condition_npz),
+                                  lr=a.lr, lora_alpha=a.lora_alpha, warmup=a.warmup,
+                                  batch_size=a.batch_size, ema_decay=a.ema_decay, seed=a.seed)
+            if sha256_file(a.checkpoint) != a.checkpoint_sha:
+                raise SystemExit("source checkpoint differs from requested hash")
+            if stored.get("checkpoint_sha256") != sha256_file(done):
+                raise SystemExit("existing strict checkpoint differs from RUN.json hash")
+        log(f"{done} exists; nothing to do")
+        return 0
+    if a.support_manifest is not None and (done.exists() or (a.out_dir / "RUN.json").exists()):
+        raise SystemExit("strict output is incomplete; inspect rather than overwrite")
+    avail = int(open("/proc/meminfo").read().split("MemAvailable:")[1].split()[0]) / 1048576
+    if avail < a.min_system_gib:
+        raise SystemExit(f"MemAvailable {avail:.1f} GiB < floor {a.min_system_gib}")
+    got = sha256_file(a.checkpoint)
+    if got != a.checkpoint_sha:
+        raise SystemExit(f"checkpoint sha {got[:12]} != {a.checkpoint_sha[:12]}")
+    device = torch.device(a.device)
+    if device.type == "cuda" and device.index is None:
+        device = torch.device("cuda", 0)
+    if device.type == "cuda":
+        total = torch.cuda.get_device_properties(device).total_memory
+        torch.cuda.set_per_process_memory_fraction(min(1.0, a.max_cuda_gib * 2**30 / total), device)
+    torch.manual_seed(a.seed)
+    rng = np.random.default_rng(a.seed)
+
+    bundle = SS.restore_standalone(a.checkpoint, device=device, weights="ema", conditioner_source="embedded")
+    diffusion, scaler = bundle.diffusion, bundle.scaler
+    ext = SS.load_conditions(a.condition_npz, bundle, default_cells_per_bulk=1)
+    folds = pd.read_csv(a.dataset_dir / "folds.csv")
+    if not np.array_equal(folds["sample_id"].astype(str).to_numpy(), ext.sample_ids.astype(str)):
+        raise SystemExit("condition NPZ rows differ from folds.csv")
+    support = None
+    if a.support_manifest is not None:
+        from support_manifest import select_rows
+        support = support_preflight
+        train_rows, test_rows = select_rows(support, folds)
+        inner_rows = np.zeros(0, dtype=np.int64)
+        selected_visits = set(folds.loc[folds["donor_id"].astype(str).isin(support["train_donor_ids"]),
+                                        "sample_id"].astype(str))
+        if selected_visits - set(support["train_sample_ids"]) != set(support["excluded_sample_ids"]):
+            raise SystemExit("support excludes inconsistent selected-donor visits")
+        folds = folds.assign(fold=np.where(folds.index.isin(test_rows), 0,
+                                           np.where(folds.index.isin(train_rows), 1, -1)),
+                             inner_val_custom=False)
+        a.fold = 0
+        inner_col = "inner_val_custom"
+    elif a.train_donors_file:
+        # few-pair deployment scenario: train on the listed donors only, no inner validation, and treat every other
+        # usable row (fold >= 0) as test; --fold is ignored
+        want = {l.strip() for l in Path(a.train_donors_file).read_text().splitlines() if l.strip()}
+        is_train = folds["donor_id"].astype(str).isin(want) & (folds["fold"] >= 0)
+        if int(folds.loc[is_train, "donor_id"].nunique()) != len(want):
+            raise SystemExit("some listed training donors are missing or excluded")
+        folds = folds.assign(fold=np.where(folds["fold"] < 0, -1, np.where(is_train, 1, 0)))
+        folds["inner_val_custom"] = False
+        a.fold = 0
+        inner_col = "inner_val_custom"
+    else:
+        inner_col = f"inner_val_f{a.fold}"
+    test_rows = folds.index[folds["fold"] == a.fold].to_numpy()
+    inner_rows = folds.index[folds[inner_col]].to_numpy()
+    # fold -1 marks rows excluded by the pairing identity checks: never trained on, never tested
+    train_rows = folds.index[(folds["fold"] != a.fold) & (folds["fold"] >= 0)
+                             & ~folds[inner_col]].to_numpy()
+    subset_info = None
+    if a.train_n:
+        # learning curve: keep a stratified, nested subset of the training donors (inner-validation and test
+        # rows unchanged).  Each stratum is permuted once by --subset-seed, so smaller subsets are prefixes of
+        # larger ones; allocation is proportional to stratum size with at least one donor per stratum.
+        grp = folds.loc[train_rows, "group"].astype(str)
+        strata = (grp.str.contains(a.strata_regex, regex=True) if a.strata_regex
+                  else pd.Series(True, index=grp.index)).to_numpy()
+        srng = np.random.default_rng(a.subset_seed)
+        labels = sorted(set(strata.tolist()))
+        perm = {s: train_rows[strata == s][srng.permutation(int((strata == s).sum()))] for s in labels}
+        n = min(int(a.train_n), len(train_rows))
+        sizes = {s: len(perm[s]) for s in labels}
+        alloc = {s: max(1, int(np.floor(n * sizes[s] / len(train_rows)))) for s in labels}
+        while sum(alloc.values()) < n:
+            s = max(labels, key=lambda s: n * sizes[s] / len(train_rows) - alloc[s])
+            alloc[s] += 1
+        while sum(alloc.values()) > n:
+            s = max((s for s in labels if alloc[s] > 1), key=lambda s: alloc[s])
+            alloc[s] -= 1
+        train_rows = np.sort(np.concatenate([perm[s][:alloc[s]] for s in labels]))
+        subset_info = {"train_n": int(n), "subset_seed": a.subset_seed, "strata_regex": a.strata_regex,
+                       "allocation": {str(k): int(v) for k, v in alloc.items()}}
+    if set(test_rows) & (set(inner_rows) | set(train_rows)):
+        raise SystemExit("test rows overlap training rows")
+    if set(folds.loc[test_rows, "donor_id"]) & set(folds.loc[np.r_[train_rows, inner_rows], "donor_id"]):
+        raise SystemExit("a test donor also appears in training")
+
+    cells = Cells(a.dataset_dir / "cells_panel6k.npz", bundle.gene_ids)
+    z = np.load(a.dataset_dir / "cells_panel6k.npz", allow_pickle=False)
+    panel_measured = z["panel_measured"].astype(bool)
+    if not np.array_equal(panel_measured, ext.output_availability_mask.astype(bool)):
+        raise SystemExit("cells' measured panel differs from the condition NPZ output_availability_mask")
+
+    base = torch.from_numpy(ext.base_conditions.astype(np.float32)).to(device)
+    bulk_frac = torch.from_numpy(ext.bulk_measurement_fraction.astype(np.float32)).to(device)
+    avail_rows = torch.from_numpy(ext.output_availability_mask.astype(bool)).to(device)
+    cond_rows = condition_with_measurement_context(base, bulk_frac, avail_rows)   # [rows, condition_dim]
+
+    denoiser = diffusion.denoiser
+    info = set_trainable(denoiser, a.mode, a.lora_rank, a.lora_alpha)
+    params = [p for p in denoiser.parameters() if p.requires_grad]
+    named = [(n, p) for n, p in denoiser.named_parameters() if p.requires_grad]
+    opt = torch.optim.AdamW(params, lr=a.lr, weight_decay=0.0)
+    ema = {n: p.detach().clone() for n, p in named}
+    log(f"mode={a.mode} trainable={info['n_trainable']:,} rows train/inner/test="
+        f"{len(train_rows)}/{len(inner_rows)}/{len(test_rows)} cells(train)={int(cells.count[train_rows].sum()):,}")
+
+    # fixed inner-validation draws
+    vrng = np.random.default_rng(a.seed + 1)
+    val_cells, val_rows = [], []
+    for r in inner_rows:
+        n = int(cells.count[r])
+        take = min(a.val_cells_per_row, n)
+        pick = cells.start[r] + np.sort(vrng.choice(n, size=take, replace=False))
+        val_cells.append(pick)
+        val_rows.append(np.full(take, r))
+    val_cells = np.concatenate(val_cells) if val_cells else np.zeros(0, np.int64)
+    val_rows = np.concatenate(val_rows) if val_rows else np.zeros(0, np.int64)
+
+    def to_batch(r: np.ndarray, c: np.ndarray):
+        x_log = torch.from_numpy(np.log1p(cells.dense(c))).to(device)
+        rr = torch.from_numpy(r).to(device)
+        availability = avail_rows[rr]
+        return scaler.to_unit_tensor(x_log), cond_rows[rr], availability, x_log > 0
+
+    def swap_in_ema() -> dict:
+        backup = {n: p.detach().clone() for n, p in named}
+        with torch.no_grad():
+            for n, p in named:
+                p.copy_(ema[n])
+        return backup
+
+    def restore(backup: dict) -> None:
+        with torch.no_grad():
+            for n, p in named:
+                p.copy_(backup[n])
+
+    @torch.no_grad()
+    def val_loss() -> float:
+        if len(val_cells) == 0:
+            return float("nan")
+        diffusion.eval()
+        tot, n = 0.0, 0
+        for rep in range(a.val_repeats):
+            for i in range(0, len(val_cells), a.batch_size):
+                g = torch.Generator(device=device).manual_seed(a.seed + 1_000_003 * rep + i)
+                v, c, m, act = to_batch(val_rows[i:i + a.batch_size], val_cells[i:i + a.batch_size])
+                with torch.autocast(device_type=device.type, dtype=torch.bfloat16, enabled=device.type == "cuda"):
+                    res = diffusion.training_loss(v, c, availability_mask=m, active_target=act, generator=g)
+                tot += float(res.loss) * v.shape[0]
+                n += v.shape[0]
+        diffusion.train()
+        return tot / max(n, 1)
+
+    history = []
+    best = {"loss": val_loss(), "step": 0}
+    best_state = {n: t.detach().cpu().clone() for n, t in ema.items()}
+    history.append({"step": 0, "inner_val_loss": best["loss"], "train_loss": None})
+    log(f"step 0 inner-val loss {best['loss']:.5f} (pretrained)")
+    gen = torch.Generator(device=device).manual_seed(a.seed + 7)
+    diffusion.train()
+    t0 = time.time()
+    run_loss, run_n, bad = 0.0, 0, 0
+    max_steps = 20 if a.smoke else (a.fixed_steps or a.max_steps)
+    eval_every = 5 if a.smoke else a.eval_every
+    step = 0
+    for step in range(1, max_steps + 1):
+        for gparam in opt.param_groups:
+            gparam["lr"] = a.lr * min(1.0, step / max(1, a.warmup))
+        r, c = draw_batch(cells, train_rows, a.batch_size, rng)
+        v, cond, m, act = to_batch(r, c)
+        opt.zero_grad(set_to_none=True)
+        with torch.autocast(device_type=device.type, dtype=torch.bfloat16, enabled=device.type == "cuda"):
+            res = diffusion.training_loss(v, cond, availability_mask=m, active_target=act, generator=gen)
+        if not bool(torch.isfinite(res.loss)):
+            raise FloatingPointError(f"non-finite loss at step {step}")
+        res.loss.backward()
+        norm = nn.utils.clip_grad_norm_(params, 5.0)
+        if not bool(torch.isfinite(norm)):
+            raise FloatingPointError(f"non-finite grad norm at step {step}")
+        opt.step()
+        with torch.no_grad():
+            for n_, p in named:
+                ema[n_].lerp_(p.detach(), 1.0 - a.ema_decay)
+        run_loss += float(res.loss.detach())
+        run_n += 1
+        if a.fixed_steps and step == max_steps:
+            best = {"loss": float("nan"), "step": step}
+            best_state = {n_: t.detach().cpu().clone() for n_, t in ema.items()}
+        if step % eval_every == 0 and not a.fixed_steps:
+            backup = swap_in_ema()
+            vl = val_loss()
+            restore(backup)
+            rec = {"step": step, "inner_val_loss": vl, "train_loss": run_loss / run_n,
+                   "sec": round(time.time() - t0, 1)}
+            history.append(rec)
+            run_loss, run_n = 0.0, 0
+            if vl < best["loss"] - 1e-6:
+                best = {"loss": vl, "step": step}
+                best_state = {n_: t.detach().cpu().clone() for n_, t in ema.items()}
+                bad = 0
+            else:
+                bad += 1
+            log(f"step {step} train {rec['train_loss']:.5f} inner-val {vl:.5f} best {best['loss']:.5f}@{best['step']} "
+                f"({rec['sec']:.0f}s)")
+            (a.out_dir / "history.json").write_text(json.dumps(history, indent=1))
+            if bad >= a.patience:
+                log(f"early stop: no improvement in {a.patience} evaluations")
+                break
+
+    # load the selected EMA state, merge LoRA, save a self-contained checkpoint
+    with torch.no_grad():
+        for n_, p in named:
+            p.copy_(best_state[n_].to(device))
+    if a.mode == "lora":
+        merge_lora(denoiser)
+    state = {k: v.detach().cpu().clone() for k, v in denoiser.state_dict().items()}
+    payload_in = SS._load_checkpoint_cpu(a.checkpoint)
+    ref_state = payload_in["ema_state"]["model_state"]
+    if set(ref_state.keys()) != set(state.keys()):
+        raise SystemExit("fine-tuned state keys differ from the checkpoint's")
+    changed = {k: float((state[k].float() - ref_state[k].float()).abs().max()) for k in state}
+    out = {k: payload_in[k] for k in KEEP_KEYS if k in payload_in}
+    out["model_state"] = state
+    out["ema_state"] = {"model_state": state}
+    ft_meta = {
+        "schema": SCHEMA, "source_checkpoint": str(a.checkpoint), "source_checkpoint_sha256": got,
+        "dataset_dir": str(a.dataset_dir), "condition_npz": str(a.condition_npz),
+        "condition_npz_sha256": ext.input_sha256, "fold": a.fold, "mode": a.mode, "trainable": info,
+        "lr": a.lr, "warmup": a.warmup, "batch_size": a.batch_size, "ema_decay": a.ema_decay,
+        "max_steps": max_steps, "steps_run": step, "eval_every": eval_every, "patience": a.patience,
+        "train_subset": subset_info, "train_donors_file": a.train_donors_file or None,
+        "support_manifest": str(a.support_manifest) if support else None,
+        "support_sha256": support["sha256"] if support else None,
+        "train_donor_ids": support["train_donor_ids"] if support else None,
+        "excluded_selected_donor_visits": support["excluded_sample_ids"] if support else None,
+        "fixed_steps": a.fixed_steps or None,
+        "selected_step": best["step"], "selected_inner_val_loss": best["loss"],
+        "pretrained_inner_val_loss": history[0]["inner_val_loss"],
+        "train_rows": [str(x) for x in folds.loc[train_rows, "sample_id"]],
+        "inner_val_rows": [str(x) for x in folds.loc[inner_rows, "sample_id"]],
+        "test_rows_never_loaded_for_training": [str(x) for x in folds.loc[test_rows, "sample_id"]],
+        "max_abs_weight_change": max(changed.values()),
+        "n_tensors_changed": int(sum(v > 0 for v in changed.values())),
+        "seed": a.seed, "wall_seconds": round(time.time() - t0, 1),
+        "written_utc": datetime.now(timezone.utc).isoformat(),
+    }
+    out["finetune"] = ft_meta
+    tmp = a.out_dir / "checkpoint_ft.pt.tmp"
+    torch.save(out, tmp)
+    os.replace(tmp, done)
+    ft_meta["checkpoint_sha256"] = sha256_file(done)
+    # reload through the sampler's own restore path and compare one forward pass
+    check = SS.restore_standalone(done, device=torch.device("cpu"), weights="ema", conditioner_source="embedded")
+    ref = SS.restore_standalone(done, device=torch.device("cpu"), weights="raw", conditioner_source="embedded")
+    xs = torch.randn(4, check.diffusion.state_dim, generator=torch.Generator().manual_seed(0))
+    tt = torch.full((4,), 0.5)
+    cc = cond_rows[torch.from_numpy(train_rows[:1]).to(device)].repeat(4, 1).float().cpu()
+    denoiser_cpu = denoiser.to("cpu").float().eval()
+    with torch.no_grad():
+        o1 = check.diffusion.denoiser(xs, tt, cc).value_prediction
+        o2 = denoiser_cpu(xs, tt, cc).value_prediction
+        o3 = ref.diffusion.denoiser(xs, tt, cc).value_prediction
+    ft_meta["reload_forward_max_abs_diff"] = float(max((o1 - o2).abs().max(), (o1 - o3).abs().max()))
+    if ft_meta["reload_forward_max_abs_diff"] > 1e-4:
+        raise SystemExit(f"reloaded checkpoint differs: {ft_meta['reload_forward_max_abs_diff']}")
+    (a.out_dir / "history.json").write_text(json.dumps(history, indent=1))
+    (a.out_dir / "RUN.json").write_text(json.dumps(ft_meta, indent=1))
+    log(f"saved {done} (selected step {best['step']}, inner-val {best['loss']:.5f} vs pretrained "
+        f"{history[0]['inner_val_loss']:.5f}); sha {ft_meta['checkpoint_sha256'][:12]}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())