Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261015T121654Z-x1-nine-core40-deck/diffusion/dataset/server_1/model8/AGENTS.md

@@ -0,0 +1,99 @@
+# Model 8 PBISC-Diffusion paper maintenance
+
+This active Model 8 source and its experiments are part of the
+`pbisc-diffusion` paper. At task start, read
+`../../paper-release-20260928/maintenance/POLICY.md` along with this
+directory's README and applicable parent instructions.
+
+At the end of each related task, follow that policy's versioned update
+workflow for the exact source, experiment, checkpoint, or evidence assets that
+changed. From the release root, use
+`maintenance/update.py prepare --request FILE --id ID`, inspect the prepared
+update, then run `maintenance/update.py publish ID --execute` and
+`maintenance/update.py status` under the policy. Record a no-op or pending
+status when no
+asset changed or publication failed; never describe that as an upload. Do not
+regenerate or alter the frozen 2026-09-28 release, its root manifests, or the
+HF `paper-20260928` revision. Inference assets are location plus receipt by
+default unless the task explicitly needs their payload.
+
+The private dataset must stay private. Include datalake, VAE, or Model 7 assets
+only with exact evidence that they are dependencies of this diffusion paper.
+The parent host GPU safety rules still apply.
+
+## Reuse existing experiments before starting compute
+
+User instruction, 2026-10-13: avoid repeating completed Model 8 experiments.
+Start with `EXPERIMENT_INDEX.md` and query `EXPERIMENT_REUSE_INDEX.jsonl` using
+`scripts/audit_experiment_reuse.py`. A missing output in a new plan directory
+does not establish that an experiment has never been done.
+
+- Before training, generation, or deconvolution, compare existing source
+  receipts: base checkpoint, exact support/test sample and donor IDs, data and
+  condition content, measured genes/masks, validation access, optimization and
+  selection schedule, seed, and implementation/configuration provenance.
+- Registry matches are candidates, not automatic equivalence. Verify payloads
+  and write truthful adoption evidence. Preserve original execution provenance;
+  do not relabel an old artifact as newly executed strict work.
+- Changing a sampling seed or receipt format does not justify retraining an
+  existing model. Reuse verified checkpoints and run only the genuinely changed
+  operation. Missing native receipts call for adoption, not automatic reruns.
+- Record the candidate sources and the exact scientific reason for any new
+  operation before launching it. Unknown provenance requires investigation,
+  not a conclusion that new compute is necessary. Consult newer live receipts
+  when the inventory predates an ongoing campaign.
+- Distinguish existing results, CPU-only statistics/plotting, a new protocol
+  update, and missing data in plans and decks. Never label all four as an empty
+  TODO. Archive/backup copies and tensor-identical models are not independent
+  experimental evidence.
+
+The 2026-10-13 audit and confirmed duplicate-training proof are in
+`plans/paper_figures_strictpairs_20261012/audit_reuse_20261013/`.
+Sepsis deconvolution uses verified historical draw-specific outputs through
+`tools/adopt_deconv_legacy.py` in that plan; its methods must not be rerun as a
+fallback when adoption validation fails. Current expanded evidence master (see `presentations/2026-09-30--paper-summary--r00/CURRENT_MASTER.json` for the live page count; all 156 Figure1-focused parent pages and original76 IDs are preserved):
+`presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261015_figure1_focused.pptx`.
+Use the build and page map named by `CURRENT_MASTER.json`; the frozen156-page parent remains in `data/x1_priority_20261015/base/`. New x1 core and full CPU evaluations are explicitly distinguished.
+Earlier 155/154/142/140/131-slide decks remain preserved source versions, not the current master.
+The 34-slide review is a historical subset, not the master. The approved method palette is
+`presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_color_palette_20261013.md`:
+PBISC LoRA #168FE5, calibrated #59A8DF, raw #C3E2F4; scSemiProfiler #83C6A8;
+BayesPrism #D9B23F and CIBERSORTx #F0D77D. Preserve biological/cohort colours and quantitative heatmap scales.
+Older gold/blue source figures are preserved historical assets; use the final semantic palette derivatives in current decks.
+
+## Current manuscript figure scopes
+
+The user's 2026-10-13 definitions supersede old figure numbers inside the deck and earlier plans:
+Figure 1 overview/corpus statistics; Figure 2 architecture/gene-context ablations on SLE pseudobulk;
+Figure 3 healthy measured bulk–sc paired mismatch, full FT, LoRA ranks and support budgets;
+Figure 4 disease measured-paired quantitative integrity comparisons;
+Figure 5 paired rare-population qualitative recovery; Figure 6 paired adaptation followed by
+bulk-only published-finding reproduction (TODO).
+The authoritative scope and proposed 131-page mapping are in
+`presentations/2026-09-30--paper-summary--r00/data/figure_assignment_20261013/`.
+Read `AUTHOR_SCOPE.md` and `ASSIGNMENT.md` before figure work. Preserve the distinction
+between total-five pairs and five training plus extra validation; rank16 is not universally
+best in existing metrics, and disease comparisons have cohort/endpoint exceptions.
+Existing rare-state quantitative slides do not substitute for paired qualitative panels.
+The scope is now applied in the actual reordered PPTX/PDF, not only in planning documents.
+Use `tools/build_x1_priority_master_20261015.py --if-changed` in the presentation directory for current x1 result refreshes. It consumes the immutable156-page base and verified evaluation receipts. `build_figure1_focused_master_20261015.py` is the preserved historical Figure1 branch; rerunning it would omit the new x1 result additions.
+Figure 2 now includes corpus-validation curves and cost metadata alongside SLE pseudobulk
+benchmarks; these are not paired real-bulk performance panels. Figure 1's recorded PCA
+uses an existing ctx2048 E12 illustration, with endpoint colours revealed by denoising progress,
+not per-step confidence. Display real R2 as a real-cell repeat from the same donors, not new donors.
+Healthy pair-budget panels cover all six types with both Pearson correlation and MAE, including adverse results.
+They use historical n+7 validation access, not total-n. Do not omit NK/DC/Other to make a five-pair claim appear stronger.
+
+The latest strengthening proposal is `data/figure_assignment_20261013/STRENGTHENING_PROPOSAL.md` under the deck.
+Corpus 41,395,184 vs train 33,141,380 is a split difference, not K exclusion; all training cells are covered after top-up.
+Healthy 88.51% is a source-cell share. The remainder includes ACPA+ at-risk and is labelled Disease / at-risk.
+Historical LoRA all means 78–79 training +7 validation, keeping21–22 test/fold. Nested random training subsets and
+outer/inner splits replay exactly. The legacy total-five102-test/draw design differs from strict healthy outer-fold tests.
+Figure5 original-paper targets require author identity: Ye Mono_c06 pediatric pulmonary and Imperial wave2
+steroid-associated monocytes. Ye observed-marker co-detection is a diagnostic, not reproduced cluster identity;
+fine labels were not found in133 audited canonical obs schemas. Existing generic rare-state rules do not substitute
+for those source-defined populations. Giroux/Ji program results are not automatically rare-population evidence.
+
+Figure 1 main corpus panel now uses exact train metadata: 33,141,380 cells, 5,232 samples, 2,572 corrected donor identities; Healthy 88.4241%. The 88.51% healthy share applies to the full source pool only. Training cell-type proportions were recomputed across 432 metadata shards, not scaled from full-source proportions. Use `data/figure1_revision_20261014/train_corpus/result.json`. Figure 1a uses organic vector glyphs; cells/model nodes are schematic, and K is distinct from generated-cell count.
+
+Figure 1a now explicitly shows noisy gene-space state -> iterative PBISC denoising -> sparse expression with exact zeros; those profiles are schematic, not recorded data. Figure 1b main proportions use grouped small categories with source membership in `data/figure1_focus_20261015/corpus/display_aggregations.csv`; Unlabelled and ACPA+ at-risk remain separate. Complete training counts are preserved on support page9, full-source composition page10. Do not revert the denoiser to a generic molecular-network icon or drop the detail panel.

paper-releases/pbisc-diffusion/updates/20261015T121654Z-x1-nine-core40-deck/diffusion/dataset/server_1/model8/plans/paper_figures_strictpairs_20261012/recoveries/x1_priority_completion_20261015/GPU_UNITS.json

@@ -0,0 +1,11 @@
+[
+  "ji2025_chili-disease-draw2",
+  "ji2025_chili-disease-draw3",
+  "ji2025_chili-disease-draw4",
+  "immicro-healthy-fold0-n10",
+  "immicro-healthy-fold0-n20",
+  "immicro-healthy-fold0-n40",
+  "immicro-healthy-fold1-n10",
+  "immicro-healthy-fold1-n20",
+  "immicro-healthy-fold1-n40"
+]

paper-releases/pbisc-diffusion/updates/20261015T121654Z-x1-nine-core40-deck/diffusion/dataset/server_1/model8/plans/paper_figures_strictpairs_20261012/recoveries/x1_priority_completion_20261015/OWNERS.json

@@ -0,0 +1,11 @@
+{
+  "ji2025_chili-disease-draw2": "server_1",
+  "ji2025_chili-disease-draw3": "server_2",
+  "ji2025_chili-disease-draw4": "server_1",
+  "immicro-healthy-fold0-n10": "server_2",
+  "immicro-healthy-fold0-n20": "server_1",
+  "immicro-healthy-fold0-n40": "server_2",
+  "immicro-healthy-fold1-n10": "server_1",
+  "immicro-healthy-fold1-n20": "server_2",
+  "immicro-healthy-fold1-n40": "server_1"
+}

paper-releases/pbisc-diffusion/updates/20261015T121654Z-x1-nine-core40-deck/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/build_x1_priority_master_20261015.py

@@ -0,0 +1,470 @@
+#!/usr/bin/env python3
+"""CPU-only, repeatable, source-verified x1 results expansion of the 156-page master.
+
+Completed core results and completed full CPU statistics have separate admission
+rules. No generated-array loading, fitting, training or inference occurs here.
+The parent deck/spec and Figure 1 assets are immutable inputs. Every successful
+refresh is retained under data/x1_priority_20261015/builds/<UTC-id>/ before the
+versioned x1_priority PPTX/PDF aliases advance.
+"""
+from __future__ import annotations
+import argparse, copy, csv, fcntl, hashlib, json, os, re, shutil, subprocess, sys
+from collections import Counter, defaultdict
+from datetime import datetime, timezone
+from pathlib import Path
+if os.environ.get('CUDA_VISIBLE_DEVICES') != '':
+    raise SystemExit('CPU-only: set CUDA_VISIBLE_DEVICES to the empty string')
+D=Path(__file__).resolve().parents[1]; M=D.parents[1]
+S=M/'plans/paper_figures_strictpairs_20261012'
+A=D/'data/x1_priority_20261015'
+ORIGINAL_PARENT=D/'data/figure1_focus_20261015/spec.json'
+PARENT=A/'base/spec.json'
+STATUS=S/'recoveries/x1_priority_completion_20261015/evaluation/STATUS.json'
+OUT='PBISC-Diffusion_expanded_master_20261015_x1_priority'
+sys.path[:0]=[str(D),str(D/'tools')]
+import numpy as np
+import pandas as pd
+import style as ST
+import matplotlib.pyplot as plt
+from matplotlib.lines import Line2D
+from matplotlib.ticker import MaxNLocator
+from pptx import Presentation
+from pptx.util import Inches
+import fitz
+import generate_deck as G
+from reorder_master_figures_20261013 import draw_card, draw_note
+import verify_expanded_master_20261013 as V
+
+COHORTS=['ye2026_sepsis','imperial_eskd','giroux2022','ji2025_chili']
+NAMES={'ye2026_sepsis':'Sepsis','imperial_eskd':'Imperial ESKD','giroux2022':'Giroux','ji2025_chili':'CHILI','immicro':'Healthy ImMicro'}
+ARMS=['raw','refcal','pairedcal','scsemi','lora']
+COL={'raw':'#C3E2F4','refcal':'#59A8DF','pairedcal':'#59A8DF','scsemi':'#83C6A8','lora':'#168FE5'}
+LAB={'raw':'Raw','refcal':'Reference cal.','pairedcal':'Paired cal.','scsemi':'scSemiProfiler','lora':'PBISC x1 LoRA'}
+MARK={'raw':'D','refcal':'s','pairedcal':'s','scsemi':'^','lora':'o'}
+TYPES=['T cell','B cell','Monocyte','NK cell','DC','Other']
+FULL={'full_completed','canonical_full_reused'}
+
+def sha(p):
+    with Path(p).open('rb') as f:return hashlib.file_digest(f,'sha256').hexdigest()
+
+def js(p):return json.loads(Path(p).read_text())
+def savejs(p,v):Path(p).write_text(json.dumps(v,indent=2,allow_nan=False)+'\n')
+def finite(x):
+    try:return np.isfinite(float(x))
+    except (TypeError,ValueError):return False
+
+def initialize_base():
+    """Freeze the complete parent once; user-facing master aliases may advance later."""
+    base=A/'base';proof=base/'BASE_VERIFIED.json'
+    if proof.exists():
+        v=js(proof)
+        for p,h in v['protected'].items():
+            if sha(p)!=h:raise ValueError(f'Immutable parent changed: {p}')
+        return
+    base.mkdir(exist_ok=True);(base/'figures').mkdir(exist_ok=True)
+    source=js(ORIGINAL_PARENT);original_manifest=js(ORIGINAL_PARENT.parent/'manifest.json')
+    if len(source['slides'])!=156:raise ValueError('Expected original156 parent spec')
+    original_files=[ORIGINAL_PARENT,D/'PBISC-Diffusion_expanded_master_20261015_figure1_focused.pptx',D/'PBISC-Diffusion_expanded_master_20261015_figure1_focused.pdf']
+    if len(Presentation(original_files[1]).slides)!=156:raise ValueError('Create immutablebase before advancing named master aliases')
+    if len(fitz.open(original_files[2]))!=156:raise ValueError('Named original PDF is no longer156pages')
+    protected={};originals={};bindings={}
+    for p in original_files:
+        target=base/('source_spec.json' if p==ORIGINAL_PARENT else p.name)
+        shutil.copy2(p,target);digest=sha(p)
+        if sha(target)!=digest:raise ValueError('Parentcopy hash mismatch')
+        protected[str(target)]=digest;originals[str(p)]=digest
+    effective=copy.deepcopy(source)
+    for slide in effective['slides']:
+        slide['parent_original_figure_paths']=slide.get('figures',[])
+        repl=[]
+        for f in slide.get('figures',[]):
+            p=D/f;h=sha(p)
+            if original_manifest['sources'][str(p)]['sha256']!=h:raise ValueError(f'Original figure changed: {p}')
+            dest=base/'figures'/f'{h[:20]}_{p.name}'
+            if not dest.exists():shutil.copy2(p,dest)
+            if sha(dest)!=h:raise ValueError('Parentfigure copy hash mismatch')
+            repl.append(str(dest.relative_to(D)));bindings[str(dest)]={'sha256':h};protected[str(dest)]=h
+        if 'figures'in slide:slide['figures']=repl
+    savejs(PARENT,effective);protected[str(PARENT)]=sha(PARENT)
+    savejs(base/'manifest.json',{'sources':bindings,'parent_original_sources':originals,'total_slides':156})
+    protected[str(base/'manifest.json')]=sha(base/'manifest.json')
+    savejs(proof,{'schema':'pbisc.x1.parent156.immutable.v1','status':'PASS','protected':protected,'original_named_sources':originals,'source_page_count':156,'original76_preserved':True})
+
+def expected_units():
+    out=[]
+    for ds in COHORTS:
+        for i in range(5):out.append((ds,f'{ds}/disease/draw{i}',f'{ds}-disease-draw{i}'))
+    for f in range(5):
+        for n in [5,10,20,40]:out.append(('immicro',f'immicro/healthy/fold{f}/n{n}',f'immicro-healthy-fold{f}-n{n}'))
+    return out
+
+class Evidence:
+    def __init__(self, build):
+        self.build=build;self.sources={};self.records=[];self.pending=[];self.units=[];self.failures=[];self.optional_gaps=[]
+    def snapshot(self,path,role):
+        p=Path(path).resolve()
+        if str(p) not in self.sources:
+            digest=sha(p);target=self.build/'sources'/digest[:16]/p.name
+            target.parent.mkdir(parents=True,exist_ok=True);shutil.copy2(p,target)
+            if sha(p)!=digest or sha(target)!=digest:raise ValueError(f'Source changed while copying {p}')
+            self.sources[str(p)]={'path':str(p),'sha256':digest,'bytes':p.stat().st_size,'role':role,'snapshot':str(target)}
+        return Path(self.sources[str(p)]['snapshot'])
+    def doc(self,path,role):return js(self.snapshot(path,role))
+    def table(self,path,role):return pd.read_csv(self.snapshot(path,role))
+    def optional_intervals(self,path,role,required,unit,arm):
+        # A one-arm evaluator has no cross-method comparison to estimate.
+        # Empty/header-only optional interval output must not discard valid
+        # core, realism, contrast, or donor-DEG results from the same arm.
+        snap=self.snapshot(path,role)
+        try:table=pd.read_csv(snap)
+        except pd.errors.EmptyDataError:table=pd.DataFrame(columns=sorted(required))
+        if table.empty:
+            self.optional_gaps.append({'unit':unit,'arm':arm,'path':str(path),'reason':'no interval rows; not a missing core/full evaluation'})
+            return pd.DataFrame(columns=sorted(required))
+        if not required <= set(table.columns):
+            self.optional_gaps.append({'unit':unit,'arm':arm,'path':str(path),'reason':'interval columns not recognized; other completed metrics retained'})
+            return pd.DataFrame(columns=sorted(required))
+        return table
+    def receipt(self,path):
+        j=self.doc(path,'completed evaluation receipt')
+        if j.get('state',j.get('status')) not in {'completed','verified','PASS','pass','full_completed','canonical_full_reused'}:
+            raise ValueError('receipt not completed')
+        for x in j.get('outputs',[]):
+            if not isinstance(x,dict) or 'path' not in x or 'sha256' not in x:continue
+            p=Path(x['path'])
+            if not p.is_file() or sha(p)!=x['sha256']:raise ValueError(f'Receipt hash mismatch: {p}')
+            if 'bytes' in x and p.stat().st_size!=x['bytes']:raise ValueError(f'Receipt size mismatch: {p}')
+        return j
+    def load(self, manifest):
+        provided={}
+        if manifest.exists():
+            st=self.doc(manifest,'evaluation availability snapshot')
+            for u in st.get('units',[]):provided[u.get('unit',u.get('id',''))]=u
+        for ds,rel,unit in expected_units():
+            sup=S/'protocol/splits'/rel/'support.json'
+            sp=self.doc(sup,'frozen support/test split')
+            h=hashlib.sha256(json.dumps({k:v for k,v in sp.items() if k!='sha256'},sort_keys=True,separators=(',',':'),allow_nan=False).encode()).hexdigest()
+            if h!=sp['sha256']:raise ValueError(f'Invalid support hash {unit}')
+            n=5 if ds!='immicro' else int(rel.rsplit('n',1)[1])
+            if len(set(sp['train_sample_ids']))!=n or len(set(sp['train_donor_ids']))!=n:raise ValueError(f'Wrong strict support {unit}')
+            canonical=S/'receipts'/f'{unit}-score.json';base=S/'scores'/rel
+            configs={};canonical_ok=False
+            if canonical.exists():
+                try:
+                    self.receipt(canonical);canonical_ok=True
+                    stat=self.doc(base/'statistics/status.json','full statistics status')
+                    if stat['support_sha256']!=h:raise ValueError('Statistics support mismatch')
+                    for arm in ARMS:
+                        configs[arm]={'state':'canonical_full_reused','score_summary':str(base/'score/summary.json'),
+                        'receipt':str(canonical),'statistics_status':str(base/'statistics/status.json'),
+                        'extras_manifest':str(base/'extras'/ds/'strict/manifest.json'),
+                        'contrast_manifest':str(base/'contrast'/ds/'strict/manifest.json'),
+                        'bootstrap_csv':str(base/'statistics/donor_bootstrap_2000.csv'),
+                        'individual_bootstrap_csv':str(base/'statistics/individual_donor_bootstrap_2000.csv'),
+                        'donor_deg_summary':stat.get('disease_deg',{}).get('summary_csv')}
+                except (ValueError,FileNotFoundError,KeyError) as e:
+                    self.failures.append({'unit':unit,'reason':str(e),'excluded_source':str(canonical)})
+                    configs={}
+            pu=provided.get(unit,{})
+            if pu.get('support_sha256') and pu['support_sha256']!=h:raise ValueError(f'Manifest support mismatch {unit}')
+            for arm,c in pu.get('arms',{}).items():
+                if arm in ARMS and (arm not in configs or c.get('state') in FULL):configs[arm]=c
+            accepted={}
+            for arm,c in configs.items():
+                if c.get('state') not in FULL|{'core_only'}:continue
+                try:
+                    if c.get('receipt'):
+                        bound=self.receipt(c['receipt'])
+                        if bound.get('support_sha256',h)!=h:raise ValueError('Receipt support mismatch')
+                    elif not canonical_ok:raise ValueError('Missing validated receipt')
+                    summary=self.doc(c['score_summary'],'verified core metrics')
+                    if ds!=summary['dataset'] or arm not in summary['headline']:raise ValueError('Dataset/arm mismatch')
+                    if set(summary['donors'])!=set(sp['test_sample_ids']):raise ValueError('Test sample mismatch')
+                    if not summary['per_arm'][arm]['receipt_verified']:raise ValueError('Generation receipt not verified')
+                    record={'unit':unit,'cohort':ds,'relative':rel,'arm':arm,'state':c['state'],'support_sha256':h,'n':n,
+                            'fold':int(re.search(r'fold(\d+)',rel).group(1)) if ds=='immicro' else None,
+                            'draw':int(re.search(r'draw(\d+)',rel).group(1)) if ds!='immicro' else None,
+                            'headline':summary['headline'][arm], 'composition':summary['per_arm'][arm]['composition'],
+                            'test_n':len(sp['test_sample_ids']),'source_summary':c['score_summary']}
+                    if c['state'] in FULL:
+                        status=self.doc(c['statistics_status'],'completed full-statistics status')
+                        if status['support_sha256']!=h:raise ValueError('Full statistics support mismatch')
+                        for key in ['extras_manifest','contrast_manifest']:
+                            if c.get(key) and Path(c[key]).exists():
+                                mm=self.doc(c[key],key)
+                                if not mm.get('skipped') and set(mm.get('donors',[]))!=set(sp['test_sample_ids']):raise ValueError(f'{key}: samples differ')
+                                if key=='contrast_manifest' and not mm.get('skipped'):
+                                    hh=self.table(Path(c[key]).parent/'headline.csv','disease contrast and ranking metrics')
+                                    rr=hh.loc[hh.arm==arm]
+                                    if len(rr)!=1 or not bool(rr.iloc[0].repro_pass):raise ValueError('Contrast row not reproducible')
+                                    record['contrast']=rr.iloc[0].to_dict()
+                                if key=='extras_manifest':
+                                    xx=self.table(Path(c[key]).parent/'metrics.csv','full realism metrics')
+                                    rr=xx.loc[(xx['set']==arm)&(xx.real_side=='R1')&(xx['mode']=='raw')]
+                                    if len(rr)==1:record['extras']=rr.iloc[0].to_dict()
+                                    en=self.table(Path(c[key]).parent/'en.csv','sample energy score')
+                                    ee=en.loc[(en['set']==arm)&(en.cellset=='pc')&(en.distance=='sq')]
+                                    if len(ee)==1:record['en']=float(ee.iloc[0].en)
+                        if c.get('bootstrap_csv') and Path(c['bootstrap_csv']).exists():
+                            bb=self.optional_intervals(c['bootstrap_csv'],'paired donor bootstrap intervals',{'arm','baseline','metric','status','estimate','ci_low','ci_high'},unit,arm)
+                            record['bootstrap']=bb.loc[(bb.arm==arm)|(bb.baseline==arm)].to_dict('records')
+                            record['paired_bootstrap_status']='available' if len(record['bootstrap']) else 'not_available_no_cross_method_rows'
+                        if c.get('individual_bootstrap_csv') and Path(c['individual_bootstrap_csv']).exists():
+                            ib=self.optional_intervals(c['individual_bootstrap_csv'],'single-arm unique-donor bootstrap intervals',{'arm','metric','status'},unit,arm)
+                            record['individual_bootstrap']=ib.loc[ib.arm==arm].to_dict('records')
+                            record['individual_bootstrap_status']='available' if len(record['individual_bootstrap']) else 'not_estimable_or_empty'
+                        if c.get('donor_deg_summary') and Path(c['donor_deg_summary']).exists():
+                            dd=self.table(c['donor_deg_summary'],'donor-aware DEG summary')
+                            record['donor_deg']=dd.loc[dd.arm==arm].to_dict('records')
+                    self.records.append(record);accepted[arm]=c['state']
+                except (ValueError,KeyError,FileNotFoundError) as e:
+                    self.failures.append({'unit':unit,'arm':arm,'reason':str(e)})
+            self.units.append({'unit':unit,'cohort':ds,'relative':rel,'n':n,'arms':accepted})
+        return self
+
+class Plots:
+    def __init__(self,e):self.e=e;self.b=e.build;self.slides=[]
+    def add(self,fig,key,title,group,foot,notes):
+        folder=self.b/'figures';folder.mkdir(exist_ok=True)
+        for ext in ['png','pdf','svg']:fig.savefig(folder/f'{key}.{ext}',dpi=250,bbox_inches='tight',pad_inches=.06)
+        plt.close(fig)
+        self.slides.append(dict(kind='content',layout='full',deck_group=group,section=group,section_label=group,
+            title=title,figures=[str((folder/f'{key}.png').relative_to(D))],keywords=[],footnote=foot,
+            footnote_full=notes,sources=str(self.b/'SOURCE_MANIFEST.json'),addition_key='x1_priority:'+key))
+    def status(self):
+        fig,axes=plt.subplots(1,2,figsize=ST.FULL,gridspec_kw={'width_ratios':[1,1]});fig.subplots_adjust(wspace=.45,left=.07,right=.97,top=.9,bottom=.25)
+        categories=['No admitted result','Core metrics','Full CPU statistics']; colors=['#F0F0F0','#A8CAE3','#168FE5']
+        for ax,dslist in zip(axes,[['immicro'],COHORTS]):
+            labels=[f'{n} pairs' for n in [5,10,20,40]] if dslist==['immicro'] else [NAMES[x] for x in dslist]
+            mat=[]
+            for n,ds in enumerate(dslist*4 if dslist==['immicro'] else dslist):
+                units=[x for x in self.e.units if x['cohort']==ds and (ds!='immicro' or x['n']==[5,10,20,40][n])]
+                units=sorted(units,key=lambda x:x['relative']);mat.append([2 if u['arms'].get('lora') in FULL else 1 if u['arms'].get('lora')=='core_only' else 0 for u in units])
+            from matplotlib.colors import ListedColormap
+            ax.imshow(mat,cmap=ListedColormap(colors),vmin=0,vmax=2,aspect='auto');ax.set_yticks(range(4),labels)
+            ax.set_xticks(range(5),[str(n) for n in range(5)]);ax.set_xlabel('Outer fold' if dslist==['immicro'] else 'Support draw')
+            ax.set_title('Healthy strict support' if dslist==['immicro'] else 'Disease strict five-pair')
+            ax.spines[['top','right','left','bottom']].set_visible(False)
+            for i,row in enumerate(mat):
+                for j,v in enumerate(row):ax.text(j,i,['Pending','Core','Full'][v],ha='center',va='center',fontsize=12,color='white' if v==2 else '#4D4D4D')
+        self.add(fig,'availability','PBISC x1: completed evidence and remaining units','Supp S07',
+            'Frozen snapshot; each square is one strict support/test unit. Core scores and full CPU statistics remain distinct.',
+            'Only validated receipt-bound scores are included. Missing scores do not imply missing training or generation. Healthy total-n pairs, no extra validation; disease exact-five support draws. Historical panels remain unchanged.')
+    def healthy(self):
+        rows=[r for r in self.e.records if r['cohort']=='immicro'];fig,axes=plt.subplots(1,3,figsize=ST.FULL)
+        fig.subplots_adjust(left=.07,right=.98,top=.84,bottom=.18,wspace=.38)
+        for ax,(key,label) in zip(axes,[('comp_l1','Composition L1 ↓'),('pb_r_median','Sample expression r ↑'),('coverage_dm','Depth-matched coverage ↑')]):
+            foldsets=[{r['fold'] for r in rows if r['arm']==arm and r['n']==n} for arm in ['lora','scsemi','pairedcal'] for n in [5,10,20,40]]
+            common=set.intersection(*foldsets) if all(foldsets) else set()
+            for arm in ['lora','scsemi','pairedcal']:
+                rr=[r for r in rows if r['arm']==arm]
+                for ni,n in enumerate([5,10,20,40]):
+                    pts=[r for r in rr if r['n']==n and finite(r['headline'].get(key))]
+                    for r in pts:ax.scatter(ni+(.035*(r['fold']-2)),r['headline'][key],color=COL[arm],marker=MARK[arm],s=35,alpha=.6,edgecolor='#4D4D4D',linewidth=.35)
+                if common:
+                    vals=[np.mean([r['headline'][key] for r in rr if r['n']==n and r['fold'] in common]) for n in [5,10,20,40]]
+                    ax.plot(range(4),vals,c=COL[arm],marker=MARK[arm],lw=2,label=LAB[arm]+f' ({len(common)} matched folds)')
+            ax.set_xticks(range(4),['5','10','20','40']);ax.set_xlim(-.25,3.25);ax.set_xlabel('Total training pairs');ax.set_ylabel(label);ax.grid(axis='y',alpha=.16)
+        fig.legend(handles=[Line2D([],[],color=COL[a],marker=MARK[a],lw=1.8,label=LAB[a]) for a in ['lora','scsemi','pairedcal']],ncol=3,loc='upper center',bbox_to_anchor=(.5,.99))
+        self.add(fig,'healthy_budget','Strict paired budget: held-out reconstruction','Figure 3',
+            'Dots: available folds. Lines: matched folds across all three methods and four budgets, only when available. No extra validation pairs.',
+            'Healthy strict outer folds use total support budgets 5/10/20/40. Do not merge with historical n+7-validation panels. Missing folds stay missing; no interpolation or unmatched-fold mean learning curve.')
+        fig,axes=plt.subplots(2,3,figsize=ST.FULL);fig.subplots_adjust(left=.08,right=.98,top=.84,bottom=.16,wspace=.35,hspace=.5)
+        for ax,typ in zip(axes.flat,TYPES):
+            for ai,arm in enumerate(['lora','scsemi']):
+                for ni,n in enumerate([5,10,20,40]):
+                    rr=[r for r in rows if r['n']==n and r['arm']==arm]
+                    vals=[100*(r['composition']['mean_fraction_generated'][typ]-r['composition']['mean_fraction_observed'][typ]) for r in rr]
+                    if vals:ax.scatter(np.full(len(vals),ni)+(ai-.5)*.16,vals,color=COL[arm],marker=MARK[arm],s=30,alpha=.7)
+            ax.axhline(0,color='#777777',ls='--',lw=.8);ax.set_title(typ);ax.set_xticks(range(4),[5,10,20,40]);ax.set_xlim(-.25,3.25);ax.set_ylabel('Bias (percentage points)');ax.yaxis.set_major_locator(MaxNLocator(4))
+        fig.legend(handles=[Line2D([],[],color=COL[a],marker=MARK[a],lw=0,label=LAB[a]) for a in ['lora','scsemi']],ncol=2,loc='upper center',bbox_to_anchor=(.5,.99))
+        self.add(fig,'healthy_types','Strict support: composition bias across all six types','Figure 3',
+            'Each dot is one outer fold; generated mean fraction minus observed mean fraction. Zero is unbiased; this is not per-donor MAE.',
+            'T/B/Monocyte/NK/DC/Other all retained, including adverse results. Fold-specific mean fractions come from verified core score summaries. No pooling across unmatched folds.')
+    def disease(self):
+        for ds in COHORTS:
+            rr=[r for r in self.e.records if r['cohort']==ds]
+            if not rr:continue
+            if not any(r.get('contrast') for r in rr):
+                fig,axes=plt.subplots(1,2,figsize=ST.FULL);fig.subplots_adjust(left=.09,right=.98,top=.82,bottom=.2,wspace=.32)
+                for ax,key,label in zip(axes,['contrast_mean_cp10k_r_gen','deg_top50_sum'],['Disease logFC r ↑','Benchmark top-50 DEG overlap ↑']):
+                    for arm in ARMS:
+                        ordered=sorted([r for r in rr if r['arm']==arm and finite(r['headline'].get(key))],key=lambda r:r['draw'])
+                        if ordered:ax.plot([r['draw'] for r in ordered],[r['headline'][key] for r in ordered],marker=MARK[arm],c=COL[arm],lw=0,label=LAB[arm])
+                    ax.set_xticks(range(5));ax.set_xlabel('Support draw');ax.set_ylabel(label);ax.grid(axis='y',alpha=.16);ax.yaxis.set_major_locator(MaxNLocator(5))
+                    if key=='deg_top50_sum':
+                        values=[r['headline'][key] for r in rr if finite(r['headline'].get(key))]
+                        ax.set_ylim(-.1,max(1,max(values,default=0)+1));ax.yaxis.set_major_locator(MaxNLocator(5,integer=True))
+                handles,labels=axes[0].get_legend_handles_labels()
+                fig.legend(handles,labels,ncol=3,loc='upper center',bbox_to_anchor=(.5,.99))
+                fig.text(.5,.05,'Full CPU rankings and patient-aware DEG: pending',ha='center',fontsize=12,color='#6E6E6E')
+                self.add(fig,'disease_'+ds,NAMES[ds]+': completed core disease metrics','Figure 4',
+                    'Core scores only. Support draws share patients. Benchmark DEG is the saved cell-level top-50 overlap, not corrected donor-aware DEG.',
+                    'Per-draw core metrics from validated existing generation. Missing method scores are not zero; no method superiority is claimed without matched comparator results. Full partial-r/CAT/RBO and donor-aware DEG remain pending. This panel upgrades automatically when full contrast receipts are available.')
+                continue
+            fig,axes=plt.subplots(2,2,figsize=ST.FULL);fig.subplots_adjust(left=.18,right=.97,top=.94,bottom=.13,wspace=.65,hspace=.42)
+            metrics=[('headline','contrast_mean_cp10k_r_gen','Disease logFC r ↑'),('contrast','partial_r','Bulk-controlled partial r ↑'),('contrast','cat100','Same-sign CAT100 ↑'),('contrast','rbo_up_down','Rank agreement RBO ↑')]
+            for ax,(family,key,label) in zip(axes.flat,metrics):
+                sets={arm:{r['draw'] for r in rr if r['arm']==arm and finite(r.get(family,{}).get(key))} for arm in ARMS}
+                present=[v for v in sets.values() if v];common=set.intersection(*present) if present else set()
+                if not common:ax.text(.5,.5,'Matched full metrics pending',ha='center',va='center',transform=ax.transAxes,color='#6E6E6E',fontsize=11)
+                for ai,arm in enumerate(ARMS):
+                    vals=[float(r[family][key]) for r in rr if r['arm']==arm and r['draw'] in common and finite(r.get(family,{}).get(key))]
+                    if vals:
+                        ax.scatter(vals,ai+np.linspace(-.12,.12,len(vals)),color=COL[arm],marker=MARK[arm],s=45,edgecolor='#4D4D4D',linewidth=.45)
+                        ax.plot([np.mean(vals)],[ai],marker='|',markersize=14,color='#222222',lw=0)
+                        ax.text(1.02,ai,f'n={len(vals)}',transform=ax.get_yaxis_transform(),va='center',fontsize=10,color='#6E6E6E')
+                ax.set_yticks(range(5),[LAB[a] for a in ARMS]);ax.set_ylim(4.5,-.5);ax.set_xlabel(label);ax.grid(axis='x',alpha=.16);ax.xaxis.set_major_locator(MaxNLocator(4))
+            self.add(fig,'disease_'+ds,NAMES[ds]+': strict five-pair disease integrity','Figure 4',
+                'Each metric uses common draws across available methods. Dots: draws; ticks: means. Shared patients; full CPU results required for rankings.',
+                'Same strict support/test and condition for all displayed methods within each draw. Main logFC from core score; partial r/CAT/RBO from full contrast output. Each metric selects the intersection of available method draw sets; all unfiltered values remain in CSV. No statistical significance from overlapping-draw means.')
+    def extra(self):
+        full=[r for r in self.e.records if r['state'] in FULL]
+        for ds in COHORTS:
+            rr=[r for r in full if r['cohort']==ds]
+            if not rr:continue
+            fig,axes=plt.subplots(1,3,figsize=ST.FULL);fig.subplots_adjust(left=.16,right=.96,top=.92,bottom=.18,wspace=.65)
+            for ax,(fam,key,label) in zip(axes,[('headline','coverage_dm','Depth-matched coverage ↑'),('extras','c2st_knn_auc','C2ST kNN AUC (0.5 ideal)'),('direct','en','Sample energy score ↓')]):
+                for ai,arm in enumerate(ARMS):
+                    vals=[float(r.get(key,np.nan) if fam=='direct' else r.get(fam,{}).get(key,np.nan)) for r in rr if r['arm']==arm]
+                    vals=[v for v in vals if np.isfinite(v)]
+                    if vals:ax.scatter(vals,ai+np.linspace(-.12,.12,len(vals)),c=COL[arm],marker=MARK[arm],s=42,edgecolor='#4D4D4D',linewidth=.4)
+                ax.set_yticks(range(5),[LAB[a] for a in ARMS]);ax.set_ylim(4.5,-.5);ax.set_xlabel(label);ax.xaxis.set_major_locator(MaxNLocator(4));ax.grid(axis='x',alpha=.16)
+                if key=='c2st_knn_auc':ax.axvline(.5,ls='--',c='#4D4D4D',lw=.8)
+            self.add(fig,'extras_'+ds,NAMES[ds]+': full realism and depth controls','Supp S07',
+                'Only completed full CPU evaluations. Each point is a support draw; no error bars over overlapping draws.',
+                'C2ST raw-mode/R1 kNN AUC, depth-matched coverage and existing PC/squared-distance sample energy score are different diagnostics. Small core-only differences are not substituted for full results.')
+        fig,axes=plt.subplots(2,2,figsize=ST.FULL);fig.subplots_adjust(left=.12,right=.98,top=.84,bottom=.13,wspace=.4,hspace=.55)
+        for ax,ds in zip(axes.flat,COHORTS):
+            for ai,arm in enumerate(['lora','scsemi']):
+                for ti,typ in enumerate(TYPES[:5]):
+                    vals=[float(d['top50_overlap_count']) for r in full if r['cohort']==ds and r['arm']==arm for d in r.get('donor_deg',[]) if d['celltype']==typ and d['status']=='ok' and finite(d.get('top50_overlap_count'))]
+                    if vals:ax.scatter(np.full(len(vals),ti)+(ai-.5)*.15,vals,c=COL[arm],marker=MARK[arm],s=32,alpha=.75)
+            ax.set_title(NAMES[ds]);ax.set_xticks(range(5),['T','B','Mono','NK','DC']);ax.set_xlim(-.35,4.35);ax.set_ylabel('Top-50 overlap ↑');ax.set_ylim(bottom=-.8);ax.yaxis.set_major_locator(MaxNLocator(4));ax.grid(axis='y',alpha=.16)
+            if not any(r['cohort']==ds and r.get('donor_deg') for r in full):
+                ax.text(.5,.5,'Full CPU results pending',ha='center',va='center',transform=ax.transAxes,color='#6E6E6E',fontsize=12)
+        fig.legend(handles=[Line2D([],[],color=COL[a],marker=MARK[a],lw=0,label=LAB[a]) for a in ['lora','scsemi']],ncol=2,loc='upper center',bbox_to_anchor=(.5,.99))
+        self.add(fig,'donor_deg','Patient-aware DEG: same support and test samples','Supp S08',
+            'Unique-donor pseudobulk; repeated visits handled by the saved model. Dots are support draws, not independent patients.',
+            'Only donor_deg rows with status=ok enter this plot. Non-estimable or pending populations remain missing, not zero. These corrected limma-voom overlaps differ from historical cell-level Welch DEG summaries.')
+        fig,axes=plt.subplots(1,2,figsize=ST.FULL);fig.subplots_adjust(left=.24,right=.97,top=.93,bottom=.16,wspace=.55)
+        for ax,key,label in zip(axes,['comp_l1','pb_r'],['Composition L1: scSemi minus LoRA','Expression r: scSemi minus LoRA']):
+            plotted=[]
+            for r in full:
+                if r['arm']!='lora':continue
+                for q in r.get('bootstrap',[]):
+                    if q.get('arm')=='scsemi' and q.get('baseline')=='lora' and q.get('metric')==key and q.get('status')=='ok' and all(finite(q.get(k)) for k in ['estimate','ci_low','ci_high']):plotted.append((r,q))
+            for i,(r,q) in enumerate(plotted):
+                v=float(q['estimate']);lo=float(q['ci_low']);hi=float(q['ci_high']);ax.plot([lo,hi],[i,i],c='#999999',lw=1.2);ax.scatter(v,i,c=COL['lora'],s=28)
+            labels=[(NAMES[r['cohort']]+(' d'+str(r['draw']) if r['draw'] is not None else ' f'+str(r['fold'])+' n'+str(r['n']))) for r,q in plotted]
+            if len(labels)>22:
+                ax.clear();ax.text(.5,.5,'Per-unit donor intervals\nare in the exported CSV',ha='center',va='center',transform=ax.transAxes);ax.set_yticks([])
+            else:ax.set_yticks(range(len(labels)),labels,fontsize=10);ax.invert_yaxis()
+            ax.axvline(0,c='#555555',ls='--',lw=.8);ax.set_xlabel(label);ax.xaxis.set_major_locator(MaxNLocator(4))
+        self.add(fig,'donor_intervals','Paired donor uncertainty: saved 2,000-bootstrap intervals','Supp S08',
+            '95% percentile intervals within each support draw. No averaging of intervals across overlapping draws; all rows are exported.',
+            'Only scSemi-minus-LoRA paired donor bootstrap rows with status=ok. Positive composition delta favors LoRA; positive expression delta favors scSemi. These intervals are not simultaneous or multiplicity-adjusted.')
+
+
+        for ds in COHORTS+['immicro']:
+            items=[(r,q) for r in full if r['cohort']==ds and r['arm']=='lora' for q in r.get('individual_bootstrap',[]) if q.get('status')=='ok' and all(finite(q.get(k)) for k in ['estimate','ci_low','ci_high'])]
+            if not items:continue
+            fig,axes=plt.subplots(1,2,figsize=ST.FULL);fig.subplots_adjust(left=.16,right=.98,top=.93,bottom=.16,wspace=.4)
+            for ax,key,label in zip(axes,['comp_l1','pb_r'],['Composition L1 ↓','Mean sample expression r ↑']):
+                available=[(r,q) for r,q in items if q.get('metric')==key]
+                for i,(r,q) in enumerate(available):
+                    ax.plot([float(q['ci_low']),float(q['ci_high'])],[i,i],c='#777777',lw=1.2)
+                    ax.scatter(float(q['estimate']),i,c=COL['lora'],s=35)
+                labels=[('Draw '+str(r['draw']) if r['draw'] is not None else 'Fold '+str(r['fold'])+' / '+str(r['n'])+' pairs') for r,q in available]
+                ax.set_yticks(range(len(labels)),labels,fontsize=11);ax.invert_yaxis();ax.set_xlabel(label);ax.grid(axis='x',alpha=.16)
+                if not available:ax.text(.5,.5,'Interval not estimable',transform=ax.transAxes,ha='center',va='center',color='#6E6E6E')
+            self.add(fig,'individual_ci_'+ds,NAMES[ds]+': PBISC x1 donor uncertainty','Supp S08',
+                'Single-arm estimates and 95% donor-bootstrap intervals. These are not paired method differences; draws share patients.',
+                'Saved individual_donor_bootstrap_2000.csv, unique donors with all visits retained. Expression estimate is the donor-average sample r, not the core table median. No cross-method CI is fabricated when only one method is available.')
+
+
+def build(args):
+    A.mkdir(exist_ok=True);initialize_base();stamp=datetime.now(timezone.utc).strftime('%Y%m%dT%H%M%S%fZ');b=A/'builds'/stamp;b.mkdir(parents=True)
+    e=Evidence(b).load(Path(args.status));parent=js(PARENT)['slides'];assert len(parent)==156
+    pm=js(PARENT.parent/'manifest.json');protected=js(PARENT.parent/'BASE_VERIFIED.json')['protected']
+    for s in parent:
+        for f in s.get('figures',[]):
+            p=str(D/f)
+            if sha(p)!=pm['sources'][p]['sha256']:raise ValueError(f'Parent figure changed: {p}')
+    finger=hashlib.sha256(json.dumps({'evidence':{p:v['sha256'] for p,v in e.sources.items() if p!=str(Path(args.status))},'units':e.units,'parent':sha(PARENT),'builder':sha(__file__)},sort_keys=True).encode()).hexdigest()
+    old=A/'LATEST.json'
+    if args.if_changed and old.exists() and js(old).get('fingerprint')==finger:
+        savejs(b/'NO_CHANGE.json',{'fingerprint':finger,'previous':js(old)['build']});print('No numerical evidence delta; skipped rendering');return
+    savejs(b/'SOURCE_MANIFEST.json',{'schema':'pbisc.x1.priority.deck-sources.v1','sources':list(e.sources.values()),'units':e.units,'rejected':e.failures,'optional_interval_gaps':e.optional_gaps,'fingerprint':finger})
+    flat=[];boots=[];degs=[];individual=[]
+    for r in e.records:
+        flat.append({k:r[k] for k in ['unit','cohort','arm','state','n','fold','draw','support_sha256']}|r['headline']|{f'contrast_{k}':v for k,v in r.get('contrast',{}).items()})
+        individual.extend({'unit':r['unit'],'cohort':r['cohort'],**x} for x in r.get('individual_bootstrap',[]))
+        boots.extend({'unit':r['unit'],'cohort':r['cohort'],**x} for x in r.get('bootstrap',[]));degs.extend({'unit':r['unit'],'cohort':r['cohort'],**x} for x in r.get('donor_deg',[]))
+    pd.DataFrame(flat).to_csv(b/'ALL_METRICS.csv',index=False);pd.DataFrame(boots).drop_duplicates().to_csv(b/'DONOR_INTERVALS.csv',index=False);pd.DataFrame(degs).drop_duplicates().to_csv(b/'DONOR_DEG.csv',index=False);pd.DataFrame(individual).drop_duplicates().to_csv(b/'INDIVIDUAL_DONOR_INTERVALS.csv',index=False)
+    plots=Plots(e);plots.status();plots.healthy();plots.disease();plots.extra()
+    additions=defaultdict(list)
+    for s in plots.slides:additions[s['deck_group']].append(s)
+    last={s['deck_group']:n for n,s in enumerate(parent,1)};slides=[]
+    for n,ss in enumerate(parent,1):
+        s=copy.deepcopy(ss);s['parent_156_page']=n;slides.append(s)
+        if n==last[s['deck_group']]:slides.extend(additions.get(s['deck_group'],[]))
+    mapping={s['parent_156_page']:n for n,s in enumerate(slides,1) if 'parent_156_page'in s}
+    for s in slides:
+        if s['kind']=='plan_card':s['rows']=[(h,c,re.sub(r'\d+',lambda m:str(mapping[int(m.group())]),p) if re.match(r'^pp?\.',p) else p) for h,c,p in s['rows']]
+    slides[0]['subtitle']=f'{len(slides)} slides | Preserved full master | Strict x1 results and CPU-statistics snapshot'
+    prs=Presentation();prs.slide_width=Inches(G.SLIDE_W);prs.slide_height=Inches(G.SLIDE_H)
+    funcs={'title':G.build_title,'agenda':G.build_agenda,'divider':G.build_divider,'content':G.build_content,'plan_card':draw_card,'evidence_note':draw_note}
+    for n,s in enumerate(slides,1):funcs[s['kind']](prs,s) if s['kind']=='title' else funcs[s['kind']](prs,s,n)
+    pp=b/(OUT+'.pptx');pdf=b/(OUT+'.pdf');prs.save(pp)
+    savejs(b/'spec.json',{'deck':{'output_pptx':str(pp),'output_pdf':str(pdf)},'slides':slides})
+    pd.DataFrame([{'page':i,'parent_156_page':s.get('parent_156_page'),'group':s['deck_group'],'title':s['title'],'addition':s.get('addition_key','')} for i,s in enumerate(slides,1)]).to_csv(b/'page_map.csv',index=False)
+    subprocess.run(['libreoffice',f'-env:UserInstallation=file:///tmp/pbisc-x1-priority-lo-{stamp}','--headless','--convert-to','pdf:impress_pdf_Export:{"UseTaggedPDF":{"type":"boolean","value":"false"},"PDFUACompliance":{"type":"boolean","value":"false"}}','--outdir',str(b),str(pp)],check=True,timeout=600)
+    doc=fitz.open(pdf);fail=[];fonts=set();imageproof=[];newpages=[]
+    if len(doc)!=len(slides) or len(prs.slides)!=len(slides):fail.append('page count')
+    if sorted(s['parent_156_page'] for s in slides if 'parent_156_page'in s)!=list(range(1,157)):fail.append('parent156 coverage')
+    if sorted(s['original_slide'] for s in slides if 'original_slide'in s)!=list(range(1,77)):fail.append('original76 coverage')
+    for n,(slide,s) in enumerate(zip(prs.slides,slides),1):
+        imageproof.append(V._image_proofs(slide,s,n,fail))
+        if 'parent_156_page'in s and s.get('figures',[])!=parent[s['parent_156_page']-1].get('figures',[]):fail.append(['parentfigures',n])
+        for sh in slide.shapes:
+            if sh.left< -2 or sh.top< -2 or sh.left+sh.width>prs.slide_width+2 or sh.top+sh.height>prs.slide_height+2:fail.append(['shape',n,sh.name])
+            for tf in V._text_frames(sh):
+                for para in tf.paragraphs:
+                    for run in para.runs:fonts.update(V._font_names(run,para))
+        if s.get('addition_key','').startswith('x1_priority:'):newpages.append(n)
+    if fonts!={'Arial'}:fail.append(['fonts',list(fonts)])
+    for n,page in enumerate(doc,1):
+        for block in page.get_text('dict')['blocks']:
+            for line in block.get('lines',[]):
+                for t in line.get('spans',[]):
+                    x0,y0,x1,y1=t['bbox']
+                    if min(x0,y0)<-.75 or x1>page.rect.width+.75 or y1>page.rect.height+.75:fail.append(['pdftext',n,t['text']])
+    qa=b/'qa';qa.mkdir();review=fitz.open()
+    for n in newpages:
+        doc[n-1].get_pixmap(matrix=fitz.Matrix(1.3,1.3)).save(qa/f'page_{n:03}.png');review.insert_pdf(doc,from_page=n-1,to_page=n-1)
+    review.save(b/'NEW_RESULTS_REVIEW.pdf',garbage=4,deflate=True);review.close()
+    from PIL import Image,ImageOps,ImageDraw
+    thumbs=[]
+    for n in newpages:
+        im=Image.open(qa/f'page_{n:03}.png').convert('RGB');im.thumbnail((480,280));can=Image.new('RGB',(500,310),'#EEEEEE');can.paste(im,((500-im.width)//2,20));ImageDraw.Draw(can).text((10,5),f'Page {n}',fill='black');thumbs.append(can)
+    contact=Image.new('RGB',(1000,310*((len(thumbs)+1)//2)),'white')
+    for i,im in enumerate(thumbs):contact.paste(im,((i%2)*500,(i//2)*310))
+    contact.save(qa/'contact.png')
+    for p,h in protected.items():
+        if sha(p)!=h:fail.append(['protected source changed',p])
+    result={'schema':'pbisc.x1.priority.deck-qa.v1','status':'automated_pass_visual_pending' if not fail else 'failed','failures':fail,'slides':len(slides),'parent156_preserved':True,'original76_preserved':True,'Figure1_parent_pages7to10_preserved':True,'new_pages':newpages,'font':sorted(fonts),'image_checks':imageproof,'source_rejections':e.failures,'optional_interval_gaps':e.optional_gaps,'visual_review':'pending','artifacts':{'pptx':str(pp),'pdf':str(pdf),'preview':str(b/'NEW_RESULTS_REVIEW.pdf'),'contact':str(qa/'contact.png')}}
+    savejs(qa/'VERIFICATION.json',result)
+    if fail:raise RuntimeError(f'QA failed: {fail}')
+    for src,dest in [(pp,D/(OUT+'.pptx')),(pdf,D/(OUT+'.pdf'))]:
+        tmp=dest.with_suffix(dest.suffix+'.tmp');shutil.copy2(src,tmp);tmp.replace(dest)
+    latest={'build':str(b),'fingerprint':finger,'created_at':datetime.now(timezone.utc).isoformat(),'slides':len(slides),'new_pages':newpages,'accepted_units':len({r['unit'] for r in e.records}),'full_lora_units':sum(r['arm']=='lora' and r['state']in FULL for r in e.records),'artifacts':result['artifacts'],'protected_sources':protected,'visual_review':'pending'}
+    latest['verification_path']=str(qa/'VERIFICATION.json');latest['verification_sha256']=sha(qa/'VERIFICATION.json')
+    savejs(A/'LATEST.json',latest)
+    savejs(A/'VERIFIED.json',{'schema':'pbisc.x1.priority.latest-verified.v1','status':'automated_pass_visual_pending','build':str(b),'fingerprint':finger,'slides':len(slides),'verification':str(qa/'VERIFICATION.json'),'outputs':{str(D/(OUT+ext)):sha(D/(OUT+ext)) for ext in ['.pptx','.pdf']},'parent_base':str(PARENT.parent/'BASE_VERIFIED.json')})
+    print(json.dumps({k:v for k,v in latest.items() if k!='protected_sources'}))
+
+if __name__=='__main__':
+    ap=argparse.ArgumentParser();ap.add_argument('--status',default=str(STATUS));ap.add_argument('--if-changed',action='store_true');args=ap.parse_args();A.mkdir(exist_ok=True)
+    with (A/'build.lock').open('a') as lock:
+        fcntl.flock(lock,fcntl.LOCK_EX);build(args)