about summary refs log tree commit diff
path: root/README.md
blob: fc9a5fc3208c1b70ff27a78f32ccf1f5eccb6646 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
# genenetwork4

A Rust/Axum-based API server for GeneNetwork.

## Endpoints

| Method | Path                                              | Description                       |
|--------|---------------------------------------------------|-----------------------------------|
| GET    | `/`                                               | Greeting.                         |
| GET    | `/version`                                        | Build info (hash, date).          |
| GET    | `/species`                                        | All species.                      |
| GET    | `/species/{species}`                              | One species by name or id.        |
| GET    | `/species/{species}/datasets`                     | ProbeSet datasets in a species.   |
| GET    | `/groups`                                         | All groups (filterable).          |
| GET    | `/mapping-methods`                                | Mapping tools offered.            |
| GET    | `/accessions/{accession}`                         | Resolve a GN accession number.    |
| GET    | `/groups/{group}`                                 | One group, by name or Id.         |
| GET    | `/groups/{group}/parents-and-f1s`                | The family's parental design.      |
| GET    | `/groups/{group}/strains`                         | Strains/samples in a group.       |
| GET    | `/groups/{group}/case-attributes`                 | Case-attribute names and values.  |
| GET    | `/groups/{group}/traits`                          | Published (phenotype) traits.     |
| GET    | `/groups/{group}/traits/{trait_id}`               | A published trait's metadata.     |
| GET    | `/groups/{group}/traits/{trait_id}/values`        | A published trait's values.       |
| GET    | `/groups/{group}/datasets`                        | ProbeSet datasets in a group.     |
| GET    | `/publications/{pubmed_id}/traits`                | Traits a publication reports.     |
| GET    | `/datasets/{dataset}`                             | One expression dataset.           |
| GET    | `/datasets/{dataset}/probesets`                   | Expression traits in a dataset.   |
| GET    | `/datasets/{dataset}/probesets/{probeset}`        | An expression trait's metadata.   |
| GET    | `/datasets/{dataset}/probesets/{probeset}/values` | An expression trait's values.     |
| GET    | `/traits/{dataset}`                               | Every trait in a dataset.         |
| GET    | `/traits/{dataset}/{trait}`                       | One trait's metadata.             |
| GET    | `/traits/{dataset}/{trait}/values`                | One trait's values.               |
| POST   | `/compute/correlation`                            | Correlate a trait (stub; see [docs/correlations.md](docs/correlations.md)). |
| POST   | `/compute/gemma`                                  | Map a QTL with GEMMA (see below). |
| GET    | `/compute/gemma/status/{run_id}`                  | Poll a gemma run's progress.      |
| POST   | `/compute/gemma/ravanan`                          | Map a QTL via ravanan (see below).|
| POST   | `/compute/rqtl`                                   | Map a QTL with R/qtl (see below). |
| POST   | `/compute/bnw`                                    | Learn a Bayesian network (BNW).   |
| POST   | `/compute/bnw/features`                           | Score each module as a trait.     |
| GET    | `/compute/bnw/examples[/{name}]`                  | BNW's worked examples, ready to post. |
| POST   | `/plot/rqtl`                                      | The R/qtl Manhattan plot, as SVG. |
| POST   | `/plot/gemma`                                     | The GEMMA Manhattan plot, as SVG. |
| POST   | `/plot/bnw`                                       | The BNW network, as SVG.          |
| GET    | `/compute/rqtl/status/{run_id}`                   | Poll an rqtl run's progress.      |
| POST   | `/get-test-token`                                 | Get a test bearer token (local).  |
| GET    | `/tokens`                                         | List short API tokens (local).    |
| POST   | `/tokens`                                         | Mint a short API token (local).   |
| DELETE | `/tokens/{id}`                                    | Revoke a short API token (local). |
| GET    | `/tokens/who`                                     | Describe the presented credential (local). |
| GET    | `/search`                                         | Search traits (free-text).        |
| GET    | `/openapi.json`                                   | OpenAPI 3.1 document.             |

Every endpoint above requires a valid gn-auth bearer token
(`Authorization: Bearer <token>`) except `/`, `/version`, `/openapi.json`
and `/get-test-token` itself -- see docs/gn-auth.md and the
"Development" section below (`POST /get-test-token`) for how to get
one for local testing.

**Short API tokens.** Beside gn-auth's JWT there is a second credential
format for agents and scripts: `gn4ai_` + 64-char opaque tokens, stored as
SHA-256 hashes in one SQLite database (`tokens.sqlite`, `GN4_API_TOKENS_DB`,
chmod 0600, gitignored). Nothing is stored *in* the token, so revocation is
deleting a row and takes effect on the next request; a token in a log line
reveals nothing; and if the registry is lost every short token stops working.
Mint, list, revoke and introspect with `scripts/gn4-token.py
create|list|revoke|who`; the token is printed once and never again.
Administration lives at `/tokens`, `/tokens/{id}` and `/tokens/who`, which
**answer only a direct caller on this machine** (TCP peer address, plus a
refusal for anything claiming to be proxied) -- minting credentials is not
reachable from the network, and there is no switch to change that. A short
token reads **public data only**: it works everywhere a token suffices, and on
`#[gnauth(secured)]` endpoints for public rows; a confidential dataset answers
403 telling the caller to use a JWT, since gn-auth cannot validate an opaque
token. Send it as `Authorization: Bearer gn4ai_...` and never in a query
string, which would land in nginx's log and in `next` links.

**What a request looks like in the log.** One JSON line per request on stderr
(`GN4_ACCESS_LOG`), carrying the credentials, never the credential:

```json
{"event":"request","ts":"2026-11-14T14:02:11Z","method":"POST",
 "path":"/compute/gemma/ravanan","status":200,"took_ms":912,
 "request_id":"…","user":"zachs","scheme":"opaque","token":"agent-ravanan-3"}
```

`user` is the gn-auth user id, so the line already answers *who*; `scheme` and
`token` say *with which credential* (absent for a JWT, which has no registry
handle). `scripts/gn4-log.py` reads such a log and shows it as a table with the
handle resolved against `tokens.sqlite` — the subject as the registry has it
now, plus the token's status *today* (`[revoked]`, `[expired]`, `[unknown]`) and
its note, which is how "an agent was working at 14:02" is told apart from "an
agent was working at 14:02 with a credential revoked since". The path is logged
without its query string, and no secret or hash is ever written.

A missing or invalid token is a 401; which
further, resource-level authorization (if any) an endpoint checks once
a token is valid is a separate, per-endpoint matter covered in each
entry below.

Every success response is wrapped in one envelope (see
[Success responses](#success-responses) below), except `/` (a
plain-text greeting) and `/openapi.json`. The example responses below
are real ones, with long lists and strings abbreviated by `...`, and
`took_ms` and `request_id` shortened to plausible values.

Details, including example responses, below.

### Following the API without memorising it

Every success envelope carries `next`: the calls that lead on from what you
just read. It is deliberately exhaustive rather than minimal -- a genotype
version leads to its frequencies, its marker names and its scans; a tally
leads to the markers and to one marker's location and tally; a marker leads
to the PheWAS at it; a trait leads to its values, its genotypes and its
group; `/version` leads to the listings. A caller that knows only `/version`
can walk to anything.

A scan is the case that matters most: its reply points at the *class counts*
behind its own peak (`.../markers/{peak}/classes?trait={trait_id}`), because
"LOD 2.93" reads the same whether it rests on two animals or a hundred, and
`/compute/qtl2` says in `notes` that a `riself` panel drops heterozygous calls
before the scan while GEMMA keeps them at dosage 0.5.

`next` is a promise that following a link is a `GET` that returns 200, which
is what makes blind following safe. So a link that would need a parameter the
server cannot fill is *not* offered (`/groups/{g}/phewas` without a locus is a
422, and is left out); a template such as `.../markers/{marker}/location`
carries a placeholder for the caller to complete; and calls that are not `GET`s
(the `POST`s that run a scan) live in `actions`, never in `next`. A response
never suggests itself. `tests/links.rs` walks the graph from a dozen seeds and
fails if any link does not answer 200.

The convention that follows from this: **a reply always offers a next step, and
an assistant relaying it should name one** — a single sentence after any job,
taken from that reply's `next`/`actions`, unless the user asks for the result
alone. The server's part of that bargain is that the suggestion is always
*reachable*; the assistant's part is that it is one sentence and it is the
link the reply actually carries. See
[docs/www/SKILL.md](docs/www/SKILL.md#always-offer-the-next-step).

- `GET /` - returns `Hello, world!`.
- `GET /version` - returns a JSON record describing the running build:

  ```json
  {
    "data": {
      "name": "GeneNetwork4", "short_name": "GN4AI", "version": 4.0,
      "release": 1, "date": "20261107", "git": "9fa0303",
      "dirty": false, "source_digest": "2f7f77ed4ace92c1", "branch": "main",
      "database": "reachable"
    },
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": []
  }
  ```

  `date` and `git` are the HEAD commit's date and short hash, embedded at
  build time (see `build.rs`); they describe the commit the binary was
  built from, not the time it happened to be compiled. `git` carries a
  `-dirty` suffix when that build had uncommitted changes -- so a binary
  can never claim to be a commit its code does not match -- and
  `source_digest` is a hash over the compiled source files, which tells
  two builds of the same commit apart where a commit hash cannot.
  `branch` is the branch HEAD was on, or `HEAD` when detached.
  `database` is `reachable` or `unreachable`, checked at most every ten
  seconds (`db::health`): a service that answers `/version` while every
  data endpoint fails is exactly the state that went unnoticed for
  twenty minutes on 2026-11-07, and this field makes it visible. `build.rs`
  declares `.git/HEAD`, the ref it points at and `.git/index` as
  rerun triggers, so a rebuild after a commit re-stamps rather than
  reusing the old hash; committing without rebuilding does not move the
  stamp, which is the point.

- `GET /species` - returns a JSON array of all species in the `Species`
  table, mirroring GeneNetwork3's `gn3.db.species.get_all_species`:

  ```json
  {
    "data": [
      {"name": "human", "menu_name": "Human (Homo sapiens, hg19)",
       "family": "Vertebrates"},
      {"name": "mouse", "menu_name": "Mouse (Mus musculus, mm10)",
       "family": "Vertebrates"},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/groups", "description": "Groups (populations); add ?species=<name> to list one species' groups"}
    ]
  }
  ```

  Needs a live MariaDB connection, configured via the `SQL_URI`
  environment variable (same config key GeneNetwork3 uses), e.g.
  `mysql://user:password@host/db_webqtl`. Defaults to
  `mysql://webqtlout:webqtlout@127.0.0.1/db_webqtl` if unset.

- `GET /species/{species}` - returns every column of the matching row
  of the `Species` table, matched against `Name`, `FullName` or
  `SpeciesName` (matching GeneNetwork2's
  `wqflask.api.router.get_species_info`, which returns the same match
  logic but only a 4-column subset). 404 if no species matches.

  ```json
  {
    "data": {
      "id": 1, "species_id": 1, "species_name": "Mouse", "name": "mouse",
      "menu_name": "Mouse (Mus musculus, mm10)", "full_name": "Mus musculus",
      "family": "Vertebrates", "family_order_id": 1, "taxonomy_id": 10090,
      "order_id": 15
    },
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/species/mouse/datasets", "description": "Expression datasets of this species"},
      {"href": "/groups?species=mouse", "description": "Groups of this species"}
    ]
  }
  ```

  Needs `SQL_URI`, same as `/species` above.

- `GET /species/{species}/datasets` - returns the expression (ProbeSet)
  datasets belonging to a species, across all of its groups -- the
  species-scoped equivalent of `/groups/{group}/datasets`.
  GeneNetwork2's own species-scoped variant of
  `wqflask.api.router.get_datasets_for_group` still requires a group
  name (species only disambiguates it); this doesn't. 404 if no species
  matches. Not gated on fine-grained gn-auth authorization, same as
  `/groups/{group}/datasets` -- this returns dataset metadata (names,
  `public`/`confidentiality` flags), not trait data -- but a token is
  still required, like every endpoint. Needs `SQL_URI`, same as
  `/species` above.

  ```json
  {
    "data": [
      {"id": 386, "probe_freeze_id": 156, "name": "B6D2ONCILM_0412",
       "name2": "B6D2ONCILM_Apr12",
       "full_name": "HEI ONC Retina B6D2 (Apr12) RankInv",
       "short_name": "HEI ONC Retina B6D2 (Apr12) RankInv",
       "create_time": "2012-05-17", "public": 1, "confidentiality": 0,
       "data_scale": "log2", "tissue": "Retina mRNA"},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/species/mouse", "description": "The species these datasets belong to"}
    ]
  }
  ```

- `GET /groups` - returns a JSON array of all groups in the `InbredSet`
  table, matching the shape of GeneNetwork2's `GET /api/v_pre1/groups`
  response:

  ```json
  {
    "data": [
      {"id": 25, "species_id": 4,
       "display_name": "Brain, Aging: AD, Normal Gene Expression (Liang)",
       "name": "AD-cases-controls",
       "full_name": "Brain, Aging: AD, Normal Gene Expression (Liang)",
       "public": 2, "mapping_method_id": null, "genetic_type": null,
       "family": null, "species": "human", "mapping_method": null},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/species", "description": "The species the groups belong to (filter groups with ?species=<name>)"}
    ]
  }
  ```

  `family` is the population category and `species` the species' short
  name as `/species` lists it. `mapping_method` is the name of the
  mapping tool offered for the group (the `MappingMethod` row its
  `mapping_method_id` points to, see `GET /mapping-methods`), null for a
  group with none. Three optional filters, combined with
  AND and all case-insensitive: `family` (contains), `species` (exact
  short name) and `q` (contained in the name, display name or full
  name). For example `/groups?family=Reference Populations` lists the
  reference populations, and `/groups?species=rat&q=hxb` the rat
  HXB/BXH sets.

  "Group" is GeneNetwork's own alias for `InbredSet` -- see
  `docs/genenetwork_datatypes.md` for why. Needs `SQL_URI`, same as
  `/species` above.

- `GET /groups/{group}` - returns every column of the matching row of
  the `InbredSet` table, matched against `InbredSetName`, `Name`,
  `FullName` or numeric `Id` (matching GeneNetwork2's
  `wqflask.api.router.get_group_info`, which matches the same way but
  not by Id, plus gn-guile's `gn.data.population:population-info`,
  which only matches by Id). The response uses the "group" alias
  instead of "inbred_set" in its field names (see
  `docs/genenetwork_datatypes.md`); only the MariaDB column names
  themselves stay `InbredSet*`. 404 if no group matches.

  ```json
  {
    "data": {
      "id": 1, "group_id": 1, "display_name": "BXD", "name": "BXD",
      "species_id": 1, "full_name": "BXD Family", "public": 2,
      "mapping_method_id": "1", "genetic_type": "riset",
      "family": "Reference Populations (replicate average, SE, N)",
      "family_order": 1, "menu_order_id": 0.0, "group_code": "BXD",
      "description": "<p>...</p>", "mapping_method": "qtlreaper"
    },
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/groups/BXD/strains", "description": "The group's strains/samples"},
      {"href": "/groups/BXD/case-attributes", "description": "Case-attribute names and per-strain values"},
      {"href": "/groups/BXD/datasets", "description": "Expression (ProbeSet) datasets of the group"},
      {"href": "/groups/BXD/traits", "description": "Published (phenotype) traits of the group (can be very large)"}
    ]
  }
  ```

  Needs `SQL_URI`, same as `/species` above.

- `GET /groups/{group}/strains` - returns the strains/samples
  belonging to a group, matching gn-guile's
  `gn.data.strains:strain-id-names`. 404 if no group matches. Takes an
  optional `used_for_mapping` query parameter to filter to only
  strains used for mapping (gn-guile's `strain-id-names` takes the
  same flag as a keyword argument):

  ```json
  {
    "data": [
      {"id": 1, "name": "B6D2F1", "used_for_mapping": false},
      {"id": 2, "name": "C57BL/6J", "used_for_mapping": false},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/groups/BXD", "description": "The group these strains belong to"},
      {"href": "/groups/BXD/case-attributes", "description": "Case-attribute names and per-strain values"}
    ]
  }
  ```

  `GET /groups/BXD/strains?used_for_mapping=true` excludes BXD's
  parent strains (`B6D2F1`, `C57BL/6J`, `DBA/2J`), which aren't used
  for mapping. Needs `SQL_URI`, same as `/species` above.

- `GET /groups/{group}/case-attributes` - returns a group's case
  attributes: extra per-strain metadata beyond genotype/phenotype data
  (e.g. age, sex, treatment) -- every attribute name defined for the
  group, and every strain's values for them, combined into one response
  (GeneNetwork3's `gn3.api.case_attributes` splits this into separate
  `.../names` and `.../values` endpoints; neither list is large enough
  on its own to need splitting). This is the only one of the three
  sibling GeneNetwork codebases that has case attributes at all --
  GeneNetwork2 delegates to GeneNetwork3 for them rather than storing
  them itself, and gn-guile doesn't have the concept. 404 if no group
  matches; a group with no case attributes defined returns `200` with
  both lists empty, not a 404. Needs `SQL_URI`, same as `/species`
  above.

  ```json
  {
    "data": {
      "names": ["Epoch", "RRID", ...],
      "values": [{"strain": "B6D2F1", "case_attributes": {"Status": "Live", "SeqCvge": "", "RRID": "JAX:100006", "Epoch": "0 "}}, ...]
    },
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/groups/BXD/strains", "description": "The group's strains/samples"}
    ]
  }
  ```

- `GET /groups/{group}/phewas` - a PheWAS around one position: every
  published trait of the group whose QTL peak (`PublishXRef.Locus`) is a
  marker within `window` base pairs of it, strongest first. Centre it on
  a marker (`?locus=rs30524230`, e.g. a peak from a mapping) or on a
  position (`?chr=9&mb=29.940961`); the window is in **base pairs** and
  centred, so `window=10000` is 10 kb up- and downstream (default 10000).
  `?genotype_version=` picks whose marker table defines the window
  (default: the group's current one). The answer carries the markers the
  window covered as well as the traits, so an empty `traits` list is
  readable as "nothing peaks that close" rather than a failure. Each hit
  has `trait_id`, `dataset`, `abbreviation`, `description`, its peak
  `locus`/`chr`/`mb`, `lrs`, `lod` (= `lrs` / 4.604), `mean`,
  `additive`, `nsamples`, `pubmed_id` and the GeneNetwork trait `url`.
  Example: `?locus=rs30524230&window=10000` returns 4 markers and 10 BXD
  traits, from LOD 99.9 down to 2.7.

- `GET /groups/{group}/genotypes` - the genotype files this group can be
  mapped with, newest first: each `version` (e.g. `BXD.8`), the `file`
  name (`BXD.8.geno`), the panel's `title`, `individuals`, `markers`,
  whether `gemma` and `rqtl` can run on it, and which one is `current`
  (what a mapping request gets by default). The order comes from the
  deployment's `genotype/<group>.json`, whose `genofile` list is
  newest-first, the same rule GeneNetwork2 uses; groups with no JSON
  (HET3-ITP) fall back to the highest version number and the listing's
  `notes` says so. `POST /compute/rqtl` / `/compute/gemma` take `group`
  plus an optional `genotype_version` instead of paths.

  `GET /groups/{group}/genotypes/{version}` is the metadata of one of
  them: the file name, the header's own `@`-labels (`type`, `mat`,
  `pat`, ...), the panel's `individuals` in file order and the `strains`
  a mapping actually uses (parents/F1s dropped), `num_individuals`,
  `num_strains`, `markers`, and per-chromosome marker counts.
  `GET /groups/{group}/genotypes/{version}/markers?page=&limit=&chr=` is
  the list of **marker names** -- one column, paged, filterable by
  chromosome -- and a position is one call per marker:
  `.../markers/{marker}/location` returns `{"name", "chr", "cM", "Mb"}` as
  named fields, so a client iterating names never carries a position table
  and a position lookup is a lookup. So a caller can see the panel, the
  markers and the header it is mapping against; the **genotype calls
  themselves are never served**, only summaries of them. See
  [docs/genotypes.md](docs/genotypes.md).

  `GET /groups/{group}/genotypes/{version}/frequencies` is how the panel
  is actually *called*, which nothing else in the API says:

  ```json
  {"data": {"B": 0.4889, "D": 0.492109, "H": 0.018991},
   "notes": ["BXD BXD.4: 3811 marker(s), 354423 genotype cell(s)",
             "352741 call(s) counted, 1682 missing (0.47 %); …"]}
  ```

  The map is call -> frequency over the calls that are present, so it
  sums to 1; missing codes (`U`, `N`, `NA`, an empty cell) are counted
  separately and reported in `notes`, not as a call. A file whose cells
  are probabilities is **binned**, 0.05 wide by default (twenty bins over
  [0, 1], keys are the bin edges: `"0.05-0.10"`), because 0.4187 and
  0.4191 are the same statement about a marker; `?bin=0.1` changes the
  width. A file with only a handful of distinct numbers (`0`, `0.5`, `1`
  for a biallelic dosage) is reported exactly, since binning three values
  would hide the thing being looked at.
  `GET /groups/{group}/genotypes/{version}/markers/{marker}/frequencies`
  is the same tally for one locus (`?marker=` on the panel route does the
  same thing), which is how a caller checks whether a flat LOD came from
  silence or from a nearly-all-missing marker: both look identical in a
  scan.

  Counting is bounded: cells go through an interner, so a
  100-million-cell panel costs one hash lookup per cell and one
  allocation per *distinct* call, and a `.geno` over 256 MB is refused
  (`413 genotype_too_large`) rather than read -- `HSNIH-Palmer_true.geno`
  is 3.7 GB (134 927 markers x 6 147 individuals), which is not a
  diagnosis anyone waits for.
  `GET /groups/{group}/genotypes/{version}/markers/{marker}/classes?trait={id}`
  answers the next question -- *what does the trait look like in each genotype
  class at this marker* -- without serving genotypes: the calls are joined to
  the trait's values inside the server and reduced to `n`, `mean`, `sd`, `min`
  and `max` per class, with that trait's authorization applied. It is how you
  tell a peak that rests on two animals from one that rests on a hundred, and
  it is what explains why GEMMA and R/qtl2 peak at different markers on DOL
  trait 10003:

      GET .../markers/JAX00418132/classes?trait=10003
      {"A": {"n": 2,   "mean": 1.8,    "sd": 0.1414, "min": 1.7, "max": 1.9},
       "H": {"n": 20,  "mean": 2.525,  "sd": 1.1652, "min": 1.7, "max": 5.3},
       "B": {"n": 171, "mean": 3.9491, "sd": 1.1038, "min": 1.7, "max": 6.7}}

  The A class is two animals and they are the trait's two lowest values;
  R/qtl2, which drops `H` on a `riself` panel (R/qtl's `convert2riself`),
  keeps 173 animals there but only 2 of one class, so the marker is nearly
  monomorphic for it. GEMMA keeps the 20 hets at dosage 0.5, which is where
  its sharper p-value comes from. Nothing in the payload names an animal.

  `GET /groups/unlisted/genotypes/{name}/frequencies` is the same tally for
  a file no group lists (a panel in preparation, a fixture, the DO's `do/`
  cross), and `.../markers/{marker}/frequencies` for one locus. `unlisted`
  is a namespace, not a group: `/groups/unlisted` has no record, only its
  genotype subtree answers, and `{name}` is the version, `.geno` name, or a
  path relative to the genotype files directory. One prefix for the whole
  genotype surface beats a second spelling for the same tally. The markers
  come the same way: `/groups/unlisted/genotypes/{name}/markers` (names)
  and `.../markers/{marker}/location` (position).

- `GET /probesets/{name}` - **every dataset that carries a probeset with this
  name** (indexed on `ProbeSet.Name`): one row per dataset with the panel
  (`group`, `species`, `platform`) and the probeset's mapping statistics
  there (`chr`, `mb`, `mean`, `lrs`, `lod`, `locus`, `additive`). Suffixed
  variants match too -- `1427571_at` also finds `1427571_at_A`, the older
  chips -- and each row's `name` says which one it was, so the reply never
  leaves you guessing. `?group=BXD`, `?species=mouse` and `?platform=M430`
  (substring: there is no platform id in this schema) narrow it; `?page=`
  and `?limit=` page it, and `notes` carries the count that matters.

      GET /probesets/1427571_at
      → 155 rows over 155 datasets (Shh on chr5 28.46 Mb), e.g.
        ["SA_M2_1104_R","BXD","mouse","Striatum_M430_V2_RMA_November04",
         "1427571_at","Shh","5",28.4579,6.039,16.318,3.544,
         "rs52243480",0.129]

  This is the lookup a probeset id needed: it is not a Xapian field, so
  `/search` cannot answer it and the only alternative was guessing a dataset
  and calling `/datasets/{dataset}/probesets/{name}`.

- `GET /genes/{symbol}/traits` - the same rows, filtered on
  `ProbeSet.Symbol` instead: **every expression trait of a gene, across
  datasets**, with a real total (unlike `/search`'s capped list), grouped by
  dataset, and `?sort=lod` / `?min_lod=` / `?group=` / `?species=` /
  `?platform=` so "BXD datasets where Shh is a strong eQTL" is one call.
  `GET /genes/Shh/traits?sort=lod` → 1 225 rows over 17 datasets and 13
  probesets, strongest first (`HXB_Liver_1208 / 1387584_at`, LOD 18.46).

- `GET /groups/{group}/traits/{trait_id}/peak-genotypes?marker=…` - **the
  genotypes of the animals at one marker of a confirmed hit**, which is the one
  place this API returns genotypes. It is fenced, because a genotype matrix is
  the thing GeneNetwork does not hand out:

  * the server **re-runs the R/qtl scan** for the trait (the same machinery and
    gate as `/compute/rqtl`) and takes the peak from *that* run;
  * the marker must lie within `?window=` Mb of it (default 10) -- a marker
    elsewhere is `422 not_the_peak`, quoting the peak's marker, chromosome, Mb
    and LOD;
  * the peak must reach `?min_lod=` (default 3.0), or the answer is
    `422 not_a_hit`;
  * only animals **that have a value** for the trait are listed, and only this
    marker's calls -- nothing else of the matrix leaves.

  ```json
  {"marker": "UNCHS026345", "chr": "9", "pos": 75.7627, "lod": 5.856,
   "peak": {"marker": "UNCHS026345", "chr": "9", "pos": 75.7627, "lod": 5.856},
   "counts": {"A": 110, "B": 8, "H": 74},
   "animals": [{"individual": "TB001", "genotype": "A", "value": 4.3}, …]}
  ```

  Which is the data behind R/qtl's Fig 4.1, and
  `POST /plot/genotypes` draws it: post that reply's `animals` (or its
  `genotypes` map) with the trait's values from
  `/groups/{g}/traits/{id}/values`, and get an SVG with one point per animal,
  genotype class across, value up, class means marked:

  ```json
  {"marker": "UNCHS026345",
   "genotypes": {"TB001": "A", "TB002": "A", …},
   "values": [{"name": "TB001", "value": 4.3}, …]}
  ```

  A class with a couple of animals in it is visible at a glance -- the same
  warning `.../markers/{marker}/classes` puts in numbers.

- `GET /genes/{symbol}/aliases` - **Wikidata's aliases for a gene symbol**, the
  list a trait page shows next to the symbol, plus the `label` (the entity's
  canonical symbol in that species) and a `links` array of **external gene
  pages** built for that symbol and species -- RGD, MGI, Alliance, NCBI Gene,
  Wikidata, UniProt, STRING, GeneMANIA and, per species, GTEx/Open Targets/
  Protein Atlas/GWAS Catalog (human) or BioGPS/Allen Brain Atlas (mouse). Two SPARQL queries against
  `WIKIDATA_ENDPOINT` (the entity, then its `skos:altLabel`s), kept in an LMDB
  file, `WIKIDATA_CACHE_DIR/gene-aliases.lmdb`, for `WIKIDATA_TTL_SECONDS`
  (seven days by default). The entity query matches the label
  case-insensitively, because human genes are labelled `SHH` and mouse ones
  `Shh`, and picks mouse, rat or human deterministically -- `?species=human`
  gets the ortholog whose labels GeneNetwork2's page shows:

      GET /genes/Shh/aliases
      {"data": {"symbol": "Shh", "wikidata_id": "Q14860079", "species": "mouse",
                "aliases": ["9530036O11Rik","Dsh","Hhg1","Hx","Hxl3","M100081",
                            "ShhNC","sonic hedgehog"],
                "fetched_at": 1794315892, "stale": false}}

  When Wikidata cannot be reached the cached answer is served with
  `"stale": true` and the envelope marked `partial` -- the last value, not an
  empty list: an empty alias list and an unreachable upstream mean different
  things, and only one of them is about the gene. `?refresh=true` forces a
  refetch (an operator's switch). The cache is one file plus LMDB's lock
  companion, written `0600`.

  The `label` is what makes the human-only services reachable from a rodent
  symbol: `?species=human` on a mouse `Shh` gives `label: "SHH"`, an RGD link
  for human, and a GTEx link that resolves -- a mouse symbol in GTEx is an empty
  report. A cached entity written before the label was kept is refreshed rather
  than served (`stale` tells you which happened).

  **Both serve metadata, not values.** A token buys the index; the values
  are per dataset and authorized per dataset, and the `next` link goes to
  `/datasets/{dataset}/probesets/{name}/values`, which checks. There is no
  endpoint that returns values across datasets, deliberately: that is bulk
  work and `/groups/{group}/traits` is the bulk call.

- `GET|POST /baskets`, `GET|PATCH|DELETE /baskets/{name}` - **your own column
  of values**, stored in your own SQLite file, and mapped by name. This is how
  data gets *changed* without changing the database: take a trait's values,
  edit one or more of them (or build a covariate column that does not exist),
  and map that instead. MariaDB is never written to.

  ```sh
  curl -s "$API/groups/BXD/traits/10001/values" > wbc.json
  # post the column back as a basket (jq: .data is the values array)
  curl -s -X POST "$API/baskets" -d '{"name":"wbc-edit","kind":"trait",
    "base_dataset":"BXD","base_trait":10001,"values":[...]}'
  curl -s -X PATCH "$API/baskets/wbc-edit" -d '{"values":[...]}'   # revision 2
  curl -s -X POST "$API/compute/qtl2" -d '{"group":"BXD",
    "genotype_version":"BXD.4","trait_name":"basket:wbc-edit","basket":"wbc-edit"}'
  ```

  Choices worth knowing:
  * **The full column is stored**, not a diff: the caller has the original
    beside it and the comparison is theirs, while a mapping needs one complete
    column.
  * **Text with a parsed copy**: `raw` is what was typed (`5.2`, `NA`), `value`
    the number or null, `status` which it is -- so `NA` and `0` stay different
    claims. A column containing neither a number nor a missing marker
    (`NA`, `x`, `-`, `.`, `?`) is refused with the offending samples named,
    never silently dropped.
  * **An automatic version history**: revision 1 is the column as first
    stored and every `PATCH` adds one, with its own digest and Unix
    timestamp — earlier revisions are kept, not overwritten, so
    `GET /baskets/{name}/revisions` is the history,
    `?revision=N` reads an old column, and a mapping can pin one:
    `{"basket": "wbc-edit@2"}` maps revision 2 even after the basket has
    moved on. Every reply records which revision and digest ran
    (`Basket 'wbc-edit' revision 2 (digest 9559a70e…)`), so two runs on one
    name can never look alike.
  * **One SQLite file per user, named for them** —
    `BASKET_DIR/<user>.sqlite` (0600, WAL), where `<user>` is the token's
    subject: tokens rotate, subjects do not, so a new token finds the same
    baskets and an operator can see whose file it is. A subject that is not
    filename-safe is sanitised and disambiguated with a short digest.
    Baskets are private: `GET /baskets` only ever lists yours, and nothing
    here needs a database the API may write to.
  * **Every compute endpoint** takes `basket` in place of its value field --
    `/compute/{rqtl,gemma,qtl2,fitqtl}` -- and giving both is a 422 rather than
    a silent choice. `PATCH` replaces the column, it does not merge: leaving a
    value out drops it, deliberately (the dropped revision is still in the
    history).

- `POST /matrices`, `GET|DELETE /matrices/{name}` - **upload new traits and
  covariates as one named table**, checked against an existing genotype file.
  The shape is the one people have: a header row, then one row per individual,
  the first column naming them.

  ```
  id,body_weight,glucose,batch
  BXD1,23.100,9.400,2
  BXD2,18.700,8.100,1
  ...
  ```

      POST /matrices {"name":"my-traits","genotype_version":"BXD.4",
                      "contents":"id,bw,glucose\nBXD1,23.1,9.4\n...",
                      "kinds":{"batch":"covariate"}}

  The upload is checked against `genotype_version`'s panel **before** it is
  stored: the same number of individuals, and every name one of them. 92 of 93,
  or a name the panel does not have, is a 422 that says which -- a mapping that
  silently drops individuals would be worse than an error. `kinds` marks
  columns that are covariates; the rest are traits.

  Each column becomes a **basket** named `<matrix>.<column>`, so a mapping names
  it the way it names any basket, on any engine, and inherits the revision
  history and digest:

      POST /compute/qtl2 {"group":"BXD","genotype_version":"BXD.4",
                          "basket":"my-traits.bw"}

  Everything lands in the uploader's own SQLite file -- MariaDB is never
  written to -- and `GET /matrices/{name}/values` reads the table back
  (`null` where a value was missing). A script writes an example table from a
  genotype file's own header, so the fixture passes the check by
  construction: `scripts/example-matrix.py --genotype <file> --out
  example-matrix.csv`.

- `POST /uploads/genotypes` - **upload a genotype matrix that updates an
  existing population**, named `$USER-GENONAME`. Tab-delimited: the header
  names the individuals from column 4 on, column 1 is the marker, 2 the
  chromosome, 3 the position in Mb.

  ```
  marker        chr  pos     BXD1  BXD2  ...
  rs6269442     1    3.4921  B     D
  ```

  This is an *update*, not an ingest, and the checks say so: the base version
  must exist, at most **50** individuals may be new, at least **one** must
  overlap (`none of the 93 individuals are in the base population: an upload
  updates a population`), and at most **200 000** markers -- beyond that *the
  sysadmin has to organize a file that size*. Values must be the same **kind**
  as the base's (single letters, whole numbers, or probabilities), because a
  mapping engine reads the result the way it reads the population it updates; a
  letters panel full of `0/1` is refused, naming both kinds.

  Stored as text in the uploader's own SQLite file, and written out on demand
  as a GN `.geno` (Chr, Locus, cM, Mb, individuals) carrying the base's
  `@`-metadata: `GET /uploads/genotypes/{name}/geno` returns the path for the
  `genotype_file` parameter of the compute endpoints, which now accepts values
  **by name** (`samples` or a `basket`) as well as positionally:

  ```
  POST /compute/qtl2 {"genotype_file":"<path>","samples":[{"name":"BXD1","value":61.4}, …]}
  ```

  An upload of one chromosome is a poor thing to scan: a leave-one-chromosome-
  out kinship has nothing to leave out, so the scan comes back flat or empty.
  `scripts/example-genotype-upload.py` therefore spreads its markers across
  chromosomes; use `"kinship": "overall"` if a file really has only one. `GET /uploads/genotypes`
  lists yours, `.../markers` pages the marker table, `DELETE` removes one.
  `scripts/example-genotype-upload.py` turns any `.geno` into a valid update
  (subset of markers, a few individuals dropped or added).

- `GET /groups/{group}/traits` - returns every published (phenotype)
  trait belonging to a group, the listing equivalent of
  `/groups/{group}/traits/{trait_id}` (same fields, same join, no
  `trait_id` filter). 404 if no group matches. If the group has any
  published traits, its `PublishFreeze` row is checked the same way as
  `/traits/{dataset}` (see docs/gn-auth.md): 403 if it isn't public
  and the caller isn't authorized for it; an empty list (most groups
  have no phenotype dataset at all) skips the check, since there's
  nothing in it to protect. A large group (e.g. BXD, with 15000+
  phenotype traits) can take several seconds and tens of megabytes to
  return in full; there's no pagination or limit yet, matching
  GeneNetwork2's own `wqflask.api.router.fetch_traits`, which defaults
  to unbounded too. Needs `SQL_URI`, same as `/species` above.

  The optional `abbreviation` parameter keeps only the traits with that
  short name, e.g. `/groups/BXD/traits?abbreviation=5HT_MB_M` finds
  trait 17985 (an unknown abbreviation is an empty list).

  ```json
  {
    "data": [
      {"id": 10001, "phenotype_id": 4, "publication_id": 116,
       "mean": 52.13529418496525, "locus": "rs48756159",
       "lrs": 13.497491147108706, "lod": 2.931687912056626,
       "additive": 2.3944443506944437, "nsamples": 34,
       "pre_publication_description": "Central nervous system, morphology: Cerebellum weight, whole, bilateral...",
       "post_publication_description": "Central nervous system, morphology: Cerebellum weight, whole, bilateral...",
       "original_description": "Original post publication description: Cerebellum weight [mg]",
       "units": "mg", "pubmed_id": 11438585,
       "title": "Genetic control of the mouse cerebellum: identification of quantitative...",
       "authors": "Airey DC, Lu L, Williams RW", "journal": "J Neuroscience",
       "volume": "21", "pages": "5099-5109", "month": null, "year": "2001",
       "dataset": "BXDPublish", "abbreviation": "Cerebellum_Weight",
       "url": "https://genenetwork.org/show_trait?trait_id=10001&dataset=BXDPublish"},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": ["lod = lrs / 4.604"],
    "next": [
      {"href": "/groups/BXD", "description": "The group these traits belong to"}
    ]
  }
  ```

- `GET /groups/{group}/traits/{trait_id}` - returns one published
  (phenotype) trait's `PublishXRef` mapping stats plus its `Phenotype`
  description/units and `Publication` reference, based on the data
  explored in `precompute-publishdata.gmi` (e.g. its "More metadata"
  section, which identifies exactly this join as worth surfacing).
  `trait_id` is `PublishXRef.Id`, scoped by group for the same reason
  as `/groups/{group}/traits/{trait_id}/values`. 404 if no group
  matches, or if the group has no such trait; 403 if the trait's
  `PublishFreeze` isn't public and the caller isn't authorized for it
  (see docs/gn-auth.md).

  ```json
  {
    "data": {
      "id": 10001, "phenotype_id": 4, "publication_id": 116,
      "mean": 52.13529418496525, "locus": "rs48756159",
      "lrs": 13.497491147108706, "lod": 2.931687912056626,
      "additive": 2.3944443506944437, "nsamples": 34,
      "pre_publication_description": "Central nervous system, morphology: Cerebellum weight, whole, bilateral...",
      "post_publication_description": "Central nervous system, morphology: Cerebellum weight, whole, bilateral...",
      "original_description": "Original post publication description: Cerebellum weight [mg]",
      "units": "mg", "pubmed_id": 11438585,
      "title": "Genetic control of the mouse cerebellum: identification of quantitative...",
      "authors": "Airey DC, Lu L, Williams RW", "journal": "J Neuroscience",
      "volume": "21", "pages": "5099-5109", "month": null, "year": "2001",
      "dataset": "BXDPublish", "abbreviation": "Cerebellum_Weight",
      "url": "https://genenetwork.org/show_trait?trait_id=10001&dataset=BXDPublish"
    },
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": ["lod = lrs / 4.604"],
    "next": [
      {"href": "/groups/BXD/traits/10001/values", "description": "The trait's value in each strain"},
      {"href": "/groups/BXD", "description": "The group this trait belongs to"},
      {"href": "/publications/11438585/traits", "description": "Every trait reported by this trait's publication"}
    ]
  }
  ```

  `dataset` is the trait's `PublishFreeze`, `abbreviation` its short
  name (`Phenotype.Post_publication_abbreviation`, else the
  pre-publication one) and `url` its page on GeneNetwork's website.
  `lod` is `lrs` / 4.604.

  Needs `SQL_URI`, same as `/species` above.

- `GET /groups/{group}/traits/{trait_id}/case-attributes/{attribute}` - one
  case attribute's values (`Sex`, `Age`, `Tissue`, `Batch`, ...) for
  exactly the samples a trait's values cover, **in that trait's order**,
  so it lines up with `/values` without re-sorting -- the covariate
  source, kept off the values endpoint. `values` carries the text as
  recorded (`F`, `Live`), `null` where a sample has none; `num_values`
  counts those that have one, and `distinct` lists each value with its
  count, commonest first, which is what says whether it is usable as a
  covariate (a constant one is not). Attributes are per group -- `BXD`
  defines `Epoch, RRID, SeqCvge, Status` and no `Sex`, while
  `BXD-Longevity` and several other BXD-family studies do -- and an
  unknown one is a 404 whose `next` is the group's attribute list.
  Example: `/groups/BXD-Longevity/traits/10001/case-attributes/Sex`
  returns 3281 values in `AgeAtDeath`'s sample order, `F` × 3101, `M` ×
  180.

- `GET /groups/{group}/traits/{trait_id}/values` - returns one
  published (phenotype) trait's value in each of the group's strains,
  matching GeneNetwork2's `wqflask.api.router.trait_sample_data`
  (the Publish branch) and gn-guile's
  `gn.data.dataset:get-bxd-publish-name-value-dict`, generalized to any
  group rather than hardcoding BXD. `trait_id` is the trait's
  `PublishXRef.Id`; it alone doesn't identify a trait uniquely across
  groups (only the pair of group and `PublishXRef.Id` does), hence the
  group in the path. 404 if no group matches, or if the group has no
  such trait; 403 if the trait's `PublishFreeze` isn't public and the
  caller isn't authorized for it (see docs/gn-auth.md -- this is a
  `#[gnauth(secured)]` endpoint, since it shares actual trait values).
  Only Publish (phenotype) traits are supported so far, not ProbeSet
  (expression) or Geno (genotype) traits.

  **`?normalize=` / `?winsor_quantile=` / `?winsor_fences=`** apply the
  same pipeline `/compute/*` takes (see "Normalize the trait first" below),
  over exactly the samples returned, and say what ran in `notes`. The
  mapping suggestions in `actions` carry the same fields, so fetching
  z-scores and then following the suggestion maps the numbers that were
  shown rather than the raw ones:

  ```sh
  curl "$API/groups/BXD/traits/10010/values?normalize=zscore"
  # notes: "Normalized: zscore ... over 28 values; mean 0.000000, sd 1.000000"
  # actions[*].body.normalize == "zscore"
  ```

  A pipeline that cannot apply (both winsorising rules at once, a log of a
  non-positive value, nothing to divide by) is a `422
  invalid_normalization`.

  ```json
  {
    "data": [
      {"name": "BXD1", "value": 61.400001525878906},
      {"name": "BXD11", "value": 53.099998474121094},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/groups/BXD/traits/10001", "description": "The trait's metadata"}
    ]
  }
  ```

  Needs `SQL_URI`, same as `/species` above.

- `GET /groups/{group}/datasets` - returns the expression (ProbeSet)
  datasets belonging to a group, matching GeneNetwork2's
  `wqflask.api.router.get_datasets_for_group`. 404 if no group
  matches. Needs `SQL_URI`, same as `/species` above.

  ```json
  {
    "data": [
      {"id": 386, "probe_freeze_id": 156, "name": "B6D2ONCILM_0412",
       "name2": "B6D2ONCILM_Apr12",
       "full_name": "HEI ONC Retina B6D2 (Apr12) RankInv",
       "short_name": "HEI ONC Retina B6D2 (Apr12) RankInv",
       "create_time": "2012-05-17", "public": 1, "confidentiality": 0,
       "data_scale": "log2", "tissue": "Retina mRNA"},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/groups/BXD", "description": "The group these datasets belong to"}
    ]
  }
  ```

  `tissue` is what the dataset was measured in (the `Tissue` its
  `ProbeFreeze` points to), also on `GET /datasets/{dataset}` and
  `GET /species/{species}/datasets`.

- `GET /mapping-methods` - the mapping tools GeneNetwork offers, the
  `MappingMethod` table in id order: `qtlreaper`, `Rqtl`, `HappyR`,
  `PLINK` and `FastMap`. It includes tools no group currently uses
  (PLINK, FastMap); a group's `mapping_method` (see `GET /groups`)
  names the one offered for it. Needs `SQL_URI`, same as `/species`
  above.

  ```json
  {
    "data": [
      {"id": 1, "name": "qtlreaper"},
      {"id": 2, "name": "Rqtl"},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/groups", "description": "Groups, each with the mapping tool offered for it (`mapping_method`)"}
    ]
  }
  ```

- `GET /accessions/{accession}` - resolves a GeneNetwork accession
  number, the "GN602" of a dataset's page, with or without the `GN`
  prefix. It names the dataset (`InfoFiles.InfoPageName`), its page
  title, its `category` (`publish` for a phenotype dataset, `probeset`
  for an expression dataset, null if the name matches neither) and the
  `group` it belongs to. 400 if the text is not a number, 404 if no
  dataset has the accession. Needs `SQL_URI`, same as `/species` above.

  ```json
  {
    "data": {
      "accession": 602, "name": "BXDPublish",
      "title": "BXD Published Phenotypes", "category": "publish",
      "group": "BXD"
    },
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/groups/BXD", "description": "The group the dataset belongs to"},
      {"href": "/groups/BXD/traits", "description": "The dataset's published traits (can be very large)"}
    ]
  }
  ```

- `GET /publications/{pubmed_id}/traits` - returns every published
  (phenotype) trait, across groups, whose publication has this PubMed
  id (the same fields as `/groups/{group}/traits`). 404 if no
  publication has that id; a publication with no traits is an empty
  list. Unlike `/groups/{group}/traits`, the datasets are
  authorization-checked as in `GET /traits/{dataset}`: traits of a
  dataset the caller may not view are left out. A paper can report
  thousands of traits (PubMed 31634382 reports 2700+ microbiome ones),
  and there is no pagination yet. Needs `SQL_URI`, same as `/species`
  above.

  ```json
  {
    "data": [
      {"id": 17967, "phenotype_id": 16502, "publication_id": 13584,
       "mean": 6038.655919692095, "locus": "rsm10000012868",
       "lrs": 10.535408722907198, "lod": 2.2883164037591652,
       "additive": -647.5889092087912, "nsamples": 34,
       "pre_publication_description": "Blood chemistry: Blood 5-hydroxytryptamine (5-HT, serotonin) level in ad...",
       "post_publication_description": "Blood chemistry: Blood 5-hydroxytryptamine (5-HT, serotonin) level in ad...",
       "original_description": null, "units": "[pg/ul]", "pubmed_id": 24102824,
       "title": "Evaluation of heritable determinants of blood and brain serotonin homeos...",
       "authors": "Ye R, Carneiro AM, Airey D, Sanders-Bush E, Williams RW, Lu L, Wang J, Z...",
       "journal": "Genes, brain, and behavior", "volume": null,
       "pages": "247-260", "month": "Mar", "year": "2014",
       "dataset": "BXDPublish", "abbreviation": "5HT_Blood_M",
       "url": "https://genenetwork.org/show_trait?trait_id=17967&dataset=BXDPublish"},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": ["lod = lrs / 4.604"],
    "next": [
      {"href": "/traits/BXDPublish/17967", "description": "Metadata of the first trait"}
    ]
  }
  ```

- `GET /datasets/{dataset}` - returns every column of the matching row
  of the `ProbeSetFreeze` table, matched against `Name`, `Name2`,
  `FullName`, `ShortName` or numeric `Id` (matching GeneNetwork2's
  `wqflask.api.router.get_dataset_info`'s matching logic for its
  ProbeSet branch). 404 if no dataset matches. A dataset name is
  globally unique on its own (unlike a group's traits), so this
  endpoint isn't scoped under `/groups/{group}`. Needs `SQL_URI`, same
  as `/species` above.

  ```json
  {
    "data": {
      "id": 1, "probe_freeze_id": 10, "name": "Br_U_0803_M",
      "name2": "BXDMicroArray_ProbeSet_August03",
      "full_name": "Brain U74Av2 08/03 MAS5",
      "short_name": "Brain U74Av2 08/03 MAS5", "create_time": "2003-08-01",
      "public": 0, "confidentiality": 0, "data_scale": "log2",
      "tissue": "Brain mRNA"
    },
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/datasets/Br_U_0803_M/probesets", "description": "The expression traits (probesets) of this dataset (can be very large)"}
    ]
  }
  ```

- `GET /datasets/{dataset}/probesets` - returns every expression
  (ProbeSet) trait belonging to a dataset, the listing equivalent of
  `/datasets/{dataset}/probesets/{probeset}` (same fields, same join,
  no `probeset` filter). 404 if no dataset matches. Gated the same way
  as its singular sibling -- a valid token is always required, and a
  non-public dataset is checked against gn-auth once for the dataset as
  a whole (`auth::authorise_dataset`) rather than once per probeset --
  see docs/gn-auth.md. Needs `SQL_URI`, same as `/species` above.

  ```json
  {
    "data": [
      {"probeset_id": 1, "name": "100001_at", "symbol": "Cd3g",
       "description": "CD3d antigen, gamma polypeptide", "chr": "9",
       "mb": 44.970689, "mean": 5.48794285714286, "locus": "rs13480619",
       "lrs": 12.590069931048, "lod": 2.7345938164743697, "p_value": 0.269,
       "additive": -0.28515625},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": ["lod = lrs / 4.604"],
    "next": [
      {"href": "/datasets/Br_U_0803_M", "description": "The dataset these probesets belong to"}
    ]
  }
  ```

- `GET /datasets/{dataset}/probesets/{probeset}` - returns one
  expression (ProbeSet) trait's `ProbeSetXRef` mapping stats plus its
  `ProbeSet` gene annotation, matching GeneNetwork2's
  `wqflask.api.router.get_trait_info`'s ProbeSet branch. `ProbeSetXRef`
  has no single Id column of its own (its primary key is `DataId`; the
  natural key is the pair `(ProbeSetFreezeId, ProbeSetId)`), so a trait
  here is identified by dataset name plus `ProbeSet.Name` rather than a
  single numeric id, unlike `/groups/{group}/traits/{trait_id}`. 404 if
  no dataset matches, or if the dataset has no such probeset.

  ```json
  {
    "data": {
      "probeset_id": 1, "name": "100001_at", "symbol": "Cd3g",
      "description": "CD3d antigen, gamma polypeptide", "chr": "9",
      "mb": 44.970689, "mean": 5.48794285714286, "locus": "rs13480619",
      "lrs": 12.590069931048, "lod": 2.7345938164743697, "p_value": 0.269,
      "additive": -0.28515625
    },
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": ["lod = lrs / 4.604"],
    "next": [
      {"href": "/datasets/Br_U_0803_M/probesets/100001_at/values", "description": "The probeset's value in each strain"},
      {"href": "/datasets/Br_U_0803_M", "description": "The dataset this probeset belongs to"}
    ]
  }
  ```

  Needs `SQL_URI`, same as `/species` above.

- `GET /datasets/{dataset}/probesets/{probeset}/values` - returns one
  expression trait's value in each strain in the dataset, matching
  GeneNetwork2's `wqflask.api.router.trait_sample_data`'s ProbeSet
  branch, addressed the same way as
  `/datasets/{dataset}/probesets/{probeset}`. 404 if no dataset
  matches, or if the dataset has no such probeset.

  ```json
  {
    "data": [
      {"name": "B6D2F1", "value": 5.742000102996826},
      {"name": "BXD1", "value": 6.414000034332275},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/datasets/Br_U_0803_M/probesets/100001_at", "description": "The probeset's metadata"}
    ]
  }
  ```

  Needs `SQL_URI`, same as `/species` above.

### `/traits` -- ProbeSet and Publish data through one URL shape

Everything above addresses ProbeSet (expression) and Publish
(phenotype) traits differently: `/datasets/{dataset}/probesets/...`
vs. `/groups/{group}/traits/...`. A trait is a trait either way --
a dataset, a trait within it, and that trait's per-strain values -- so
`/traits/...` addresses both the same way, `(dataset, trait)`,
matching GeneNetwork2's own generic
`wqflask.api.router.trait_sample_data`/`get_trait_info`. This is
additional, not a replacement: the endpoints above are unchanged and
keep working exactly as documented.

`dataset` is resolved against `ProbeSetFreeze` first, then
`PublishFreeze` (`db::resolve_trait_dataset`) -- whichever matches
decides the category, so a bare trait name/id is never guessed across
categories, only the (much smaller, auditable) set of dataset names
is. A Publish dataset can be named either way GeneNetwork2 itself
accepts: the `PublishFreeze`'s own `Name`/`ShortName`/`Id` (e.g.
`"BXDPublish"`), or its owning group's `Name`/`InbredSetName`/
`FullName`/`Id` (e.g. `"BXD"`) -- a group normally has only one
`PublishFreeze`, so the two are near-interchangeable in practice.

- `GET /traits/{dataset}` - every trait in `dataset`, tagged by
  category. 404 if no dataset matches. A non-public dataset is checked
  against gn-auth once for the dataset as a whole
  (`auth::authorise_dataset`), same as `/datasets/{dataset}/probesets`.
  This is still more than `/groups/{group}/traits` checks for the same
  phenotype data: both endpoints require a token now, but only this one
  also checks `PublishFreeze.public`/`confidentiality` and, for a
  non-public dataset, gn-auth -- `/groups/{group}/traits` (unchanged)
  stops at the token.

  ```json
  {
    "data": [
      {"category": "probeset", "probeset_id": 1, "name": "100001_at",
       "symbol": "Cd3g", "description": "CD3d antigen, gamma polypeptide",
       "chr": "9", "mb": 44.970689, "mean": 5.48794285714286,
       "locus": "rs13480619", "lrs": 12.590069931048,
       "lod": 2.7345938164743697, "p_value": 0.269, "additive": -0.28515625},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": ["lod = lrs / 4.604"],
    "next": []
  }
  ```

  or, for a Publish dataset:

  ```json
  {
    "data": [
      {"category": "publish", "id": 10001, "phenotype_id": 4,
       "publication_id": 116, "mean": 52.13529418496525, "locus": "rs48756159",
       "lrs": 13.497491147108706, "lod": 2.931687912056626,
       "additive": 2.3944443506944437, "nsamples": 34,
       "pre_publication_description": "Central nervous system, morphology: Cerebellum weight, whole, bilateral...",
       "post_publication_description": "Central nervous system, morphology: Cerebellum weight, whole, bilateral...",
       "original_description": "Original post publication description: Cerebellum weight [mg]",
       "units": "mg", "pubmed_id": 11438585,
       "title": "Genetic control of the mouse cerebellum: identification of quantitative...",
       "authors": "Airey DC, Lu L, Williams RW", "journal": "J Neuroscience",
       "volume": "21", "pages": "5099-5109", "month": null, "year": "2001",
       "dataset": "BXDPublish", "abbreviation": "Cerebellum_Weight",
       "url": "https://genenetwork.org/show_trait?trait_id=10001&dataset=BXDPublish"},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": ["lod = lrs / 4.604"],
    "next": []
  }
  ```

- `GET /traits/{dataset}/{trait}` - one trait's metadata, tagged by
  category (`ProbeSetInfo`'s gene annotation or `TraitInfo`'s
  publication reference -- the two share little beyond
  mean/locus/lrs/additive, so neither is collapsed into the other).
  `trait` is `ProbeSet.Name` for a ProbeSet dataset, or `PublishXRef.Id`
  for a Publish dataset (a non-numeric `trait` against a Publish
  dataset is a 404, not a 500 -- it just can't match). 404 if no
  dataset matches, or no such trait in it; 403 if the dataset isn't
  public and the caller isn't authorized; 502 if gn-auth itself
  couldn't be reached. Needs `SQL_URI`, same as `/species` above.

  ```json
  {
    "data": {
      "category": "probeset", "probeset_id": 1, "name": "100001_at",
      "symbol": "Cd3g", "description": "CD3d antigen, gamma polypeptide",
      "chr": "9", "mb": 44.970689, "mean": 5.48794285714286,
      "locus": "rs13480619", "lrs": 12.590069931048, "lod": 2.7345938164743697,
      "p_value": 0.269, "additive": -0.28515625
    },
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": ["lod = lrs / 4.604"],
    "next": [
      {"href": "/traits/Br_U_0803_M/100001_at/values", "description": "The trait's value in each strain"}
    ]
  }
  ```

- `GET /traits/{dataset}/{trait}/values` - that trait's value in each
  strain, as `{name, value}` pairs regardless of category (`TraitValue`
  and `ProbeSetValue` are identical in shape, so this is the one piece
  that unifies with zero loss). Same 404/403/502 cases as
  `/traits/{dataset}/{trait}`.

  ```json
  {
    "data": [
      {"name": "BXD1", "value": 61.400001525878906},
      {"name": "BXD11", "value": 53.099998474121094},
      ...
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": [],
    "next": [
      {"href": "/traits/BXDPublish/10001", "description": "The trait's metadata"}
    ]
  }
  ```
- `GET /search` - free-text search across GN traits (see
  `docs/search-engines.md` for switching between the gn3 proxy and
  the native engine); hits keep gn3's field names and are wrapped in
  the envelope like every other response. Parameters: `q`
  (xapian syntax, e.g. `serotonin AND midbrain AND group:BXD`), `type`
  (`phenotype` or `gene`), `limit` (default 20, at most 100), and
  optionally `sort` (`lrs` or `lod`, which order identically) with
  `order` (`desc`, the default, or `asc`), and `page` (default 1) for
  the next `limit` hits.

  Three fields are added to every hit: `lod` (`lrs` / 4.604) and, for
  phenotype hits, `locus` (the peak marker's name) and `abbreviation`
  (the trait's short name), both from the database (null for other
  hits, or when the database is unreachable). `pubmed_id` is gn3's. With
  `sort`, gn3 is asked for its first 1000 matches by relevance, those
  are ranked, and `page`/`limit` are applied to the ranked list, so
  `page=2` continues where page 1 stopped; hits with no `lrs` come
  last. Without `sort`, `page` is gn3's own. For example, the highest-LOD trait about serotonin in BXD
  midbrain: `/search?q=serotonin AND midbrain AND group:BXD&sort=lod`.
  Tissue and sex are not fields: they appear in each hit's description,
  so match them there.

  Needs a reachable gn3, configured via the `GN3_URL` config key
  (default `http://127.0.0.1:5003/api/`).

  ```json
  {
    "data": [
      {"name": "17994", "dataset": "BXDPublish",
       "dataset_fullname": "BXD Published Phenotypes", "group": "BXD",
       "species": "mouse", "authors": ["Ye R", "Carneiro AM", ...],
       "year": 2014,
       "description": "Central nervous system, protein expression: Midbrain serotonin transport...",
       "mean": 1627.7004466337316, "additive": 233.98946836805567,
       "lrs": 16.49128337950074, "lod": 3.5819468678324804,
       "locus": "rsm10000013922", "abbreviation": "SERT_MB_M",
       "geno_chr": "19", "geno_mb": 58.673324, "inbredsetcode": "BXD",
       "pubmed_id": 24405699,
       "pubmed_link": "http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&lis..."}
    ],
    "meta": {
      "state": "complete",
      "took_ms": 3,
      "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
    },
    "notes": ["lod = lrs / 4.604"],
    "next": [
      {"href": "/search?q=serotonin+AND+midbrain+AND+group%3ABXD&limit=1&page=2", "description": "The next page of hits"},
      {"href": "/traits/BXDPublish/17994", "description": "Metadata of the top hit"}
    ]
  }
  ```

- `POST /plot/rqtl`, `POST /plot/gemma` - the drawing step after a
  mapping: **post the mapping's own response back** (the JSON
  `/compute/rqtl` or `/compute/gemma` returned, as-is) and get the SVG the
  deployment's plotter draws from it (`scripts/plot-rqtl.py`,
  `scripts/plot-gemma.py`; `src/plot.rs` runs them). Nothing is
  recomputed -- a plot is one script run, and it can be redrawn from a
  result you kept. `/plot/rqtl` also accepts just the `data` member;
  `/plot/gemma` needs the whole response, because its plotter reads
  `meta.columns` to find the `p_lrt` column. The reply is
  `image/svg+xml`, not the JSON envelope -- an image is not JSON -- and
  the plotter's one-line peak summary comes back in `x-plot-summary`.
  R/qtl plots LOD with the permutation thresholds as dashed lines when
  the result carried them. GEMMA plots `-log10(p_lrt)` with **no line
  unless one is asked for**: `?bonferroni=true` adds GEMMA's own
  Bonferroni (computed by the plotter as 0.05 over the analysed markers),
  and `?thresholds=` adds one computed elsewhere, each labelled. The title names the trait: taken
  from the posted response's own `Trait: ...` note (which `/compute/rqtl`
  and `/compute/gemma` both write first), or from `?trait=` when you post
  only `data`; `?trait=` wins over the note, and `?title=` replaces the
  whole title. Two traits over the same panel draw
  two different traces -- the plot is drawn from the rows you posted, and
  nothing is cached between calls. Configuration: `PLOT_PYTHON` (default `python3`),
  `RQTL_PLOTTER`, `GEMMA_PLOTTER`. Both mapping responses suggest these
  calls in `actions`.

- `GET /openapi.json` - returns an OpenAPI 3.1 document describing the
  endpoints above; see [OpenAPI](#openapi) below.

### `/compute/gemma`

Maps a QTL by running the real `gemma-wrapper`/GEMMA binaries: a
K-matrix step, then an LMM association step, both `--loco`, over a
genotype file, a SNP annotation file and a list of trait values. See
`src/gemma.rs` for the invocation itself and why it's shaped this way
(found by running GeneNetwork2's and GeneNetwork3's own gemma code for
real against the BXD fixtures -- GeneNetwork3's always errors out;
GeneNetwork2's non-LOCO branch does too; only `--loco` works).

**Individuals with no genotypes are left out.** A panel can name individuals
whose genotype row is empty -- `BXD-AE.4` has 136 of 344, `BXD-Heart-Metals.4`
200 of 481 -- and an empty row makes the kinship matrix singular, which GEMMA
reports as `eigenvalues close to zero` and then a LAPACK `LUDecomp ... input
domain error`, a message that says nothing about the cause. Those individuals
are dropped before the fit and counted in `notes`; they contribute nothing
either way, and the panel maps in seconds once they are gone.

**Normalize the trait first.** `normalize` changes the values' scale before
mapping -- `none` (default), `log2`, `log10`, `sqrt`, `negate`, `zscore`,
`robust_zscore`, or `rank` -- and `winsor_quantile` / `winsor_fences`
handle the tails. They run in that fixed order (**transform, then
winsorise, then standardise**), over exactly the individuals that enter the
fit, and the response says what ran:

```json
"normalize": "rank", "winsor_fences": 1.5
```

- `rank` is GeneNetwork2's "Quantile" (`qnorm`): `Φ⁻¹((i−0.5)/n)` by rank,
  ties sharing their average rank -- *not* array quantile normalisation.
  Accepting `qnorm` and `invert` (for `negate`) is deliberate: those are
  GN2's own names. `zscore` uses the **population** standard deviation, as
  GN2 does (`scipy.stats.mstats.zscore`), so the numbers match its trait
  page; `robust_zscore` is `(x−median)/(1.4826·MAD)` and is ours.
- `winsor_quantile: 0.05` clips each tail at the 5th/95th percentile;
  `winsor_fences: 1.5` clips at `Q1 − 1.5·IQR` / `Q3 + 1.5·IQR`, the rule
  GN2 uses to *flag* outliers -- and a count of those is reported whether
  or not anything is clipped. The two are mutually exclusive (`422`).
- `log2`/`log10` on a non-positive value, `sqrt` on a negative one, or a
  standardisation with nothing to divide by is a `422`; missing values stay
  missing and are never imputed. GN2's silent `sqrt` "add 1 in [0,1]" quirk
  is deliberately not reproduced.
- `beta`, `se` and `additive` come out on the transformed scale -- a
  `zscore` additive is in standard deviations -- so quote the note with the
  number.

**Name what you are mapping.** `trait_name` is optional -- GEMMA's values
are positional, so it needs no name to compute -- but it is what the
response reports (`Trait: BXD::20010`, first in `notes`), what the PheWAS
and R/qtl suggestions carry, and what `/plot/gemma` titles the picture
with. Without it they fall back to `<group>::gemma`, so two mappings over
one panel are distinguishable in the data but identically titled.

**No permutations.** GEMMA's permutation test is not run here and
`nperm`/`permutations`/`permutate` are **refused** (`422
permutations_not_supported`), not ignored -- an ignored flag is worse than
a refusal, because the caller believes it got thresholds. The reason is
cost: `gemma-wrapper --permutate` shuffles the phenotype file and repeats
the entire LOCO association scan per permutation, so a thousand is
thousands of full scans -- hours to days on a real panel, against the one
slot this endpoint has. R/qtl does have the test, on the same panel and the
same values: `POST /compute/rqtl` with `"nperm": 100` returns the
suggestive and significant LOD thresholds, and every successful GEMMA
response already carries that call, pre-filled, in `actions`. Judge a GEMMA
peak against the smallest `p_lrt` it reached, never against a genome-wide
threshold of its own.

**Covariates.** `covariates` fits one or more covariates alongside the
trait -- the same request shape `/compute/rqtl` takes (below), each
either a **case attribute of the group** by name, which the server reads,
aligns onto the panel and codes, or explicit per-sample numbers:

```json
"covariates": [
  {"case_attribute": "Sex"},
  {"case_attribute": "Diet", "map": {"HFD": 0, "LFD": 1}},
  {"name": "age_days", "kind": "numerical",
   "values": [{"name": "F3092", "value": 183.0}, …]}
]
```

GEMMA wants numbers, positional over the genotype panel, in a file of its
own (`-c`), so `src/covariates.rs` writes one: one row per individual,
no header, no sample names, and a categorical covariate **dummy-expanded**
into k‑1 columns named after their levels (`Sex.M`) -- GEMMA knows nothing
about factors, unlike R/qtl's wrapper, which does that expansion itself.
The first level is the reference. A level none of the fitted individuals
carries is left out (an all-zero column is singular and GEMMA stops with
`LUDecomp` rather than dropping the term the way R would), and the
response says so in `notes`.

Text with 2–20 distinct values is coded 0..k‑1 in sorted order; numeric
text is passed through as `numerical`; more than 20 levels, one level, or
an explicit `map` that misses a value is a `422` asking for a `map`. An
unknown attribute is a `422` listing the group's own. A missing value is
written `NA`, which GEMMA treats as it does a missing phenotype -- it
drops that individual (measured; GeneNetwork2 instead writes `-9`, which
GEMMA fits as a value). Both the coding, how many individuals have a
value, and any column left out are reported in `notes`.

Covariates reach **only the association step**, never the K-matrix step,
and `gemma-wrapper` hashes its file arguments, so a different covariate
set is a different cache entry while the kinship it was computed against
still is not recomputed. Covariates need a `group`: individuals are
aligned by name, which an explicit `genotype_file`/`snps_file` request
does not carry (`422 covariate_unavailable`).

**Concurrency: large jobs, not all jobs.** Gating exists to keep large
computations from piling up on this box, and for GEMMA that begins at
**50 individuals** (`compute_gate::GEMMA_LARGE_PANEL`) -- i.e. at the
smallest real panel, so in practice every mapping request is gated, and
only a handful of individuals (a four-way cross's parents, a probe of the
pipeline) runs free. Two reasons, neither about how long a small run takes:
every GEMMA run goes through **one shared cache directory**
(`GEMMA_CACHE_DIR`), where `gemma-wrapper` extracts per-chromosome kinship
archives of hundreds of megabytes by content hash, so two runs at once are
two writers over the same archives; and a real BXD WGS run is ~19 s of
every core it can get, while UM-HET3's 6 382 individuals write 207 MB of
kinship per chromosome.

So GEMMA runs **one job at a time, one per caller**: a second submission
from the same token while its run is running or queued is refused at once
with `409 compute_busy` naming that run, and a submission from another
caller waits for the slot. R/qtl is gated much later (1 000 individuals;
see below) because a `scanone` over a few hundred individuals is seconds of
one core and its results are uncached scratch of its own.

**Concurrency.** A large scan -- **1 000+ individuals**
(`compute_gate::RQTL_LARGE_PANEL`), where a `scanone` starts to be minutes
rather than seconds -- takes one of `RQTL_MAX_CONCURRENT` slots (default 8), one per
caller: a second large submission from the same token while its scan is
running or queued is refused at once with `409 compute_busy` naming that
run, and a submission that finds every slot taken waits for one to come
free. Eight because a scan is a single-threaded `Rscript` and this box has
cores to spare; not more, because a scan on a 3 000-individual cross is
minutes of CPU and the same machine serves the database. The count is fixed
when the gate is built (at the first large request), so a change needs a
restart. Smaller crosses are not gated at all.

**Permutations.** `nperm` runs R/qtl's permutation test, and the response
carries `permutation: {suggestive, significant, values}`. A scan without
it suggests one with `nperm: 100` in `actions` -- the thresholds a peak
needs to be judged against -- and `next` carries a PheWAS at the peak the
run found.

**No permutation test, and requesting one is an error.** GEMMA reports a
p-value per marker; `gemma-wrapper --permutate n` does exist, but it
repeats the whole LOCO scan per permutation, so it is hours of the one
slot this endpoint has. `nperm`, `permutations` and `permutate` are
therefore refused (`422 permutations_not_supported`) rather than ignored,
and a peak is judged against the smallest `p_lrt` it reached. The
response's `actions` carries the R/qtl permutation run over the same panel
and values, pre-filled, which is where genome-wide thresholds come from
here.

**Give it a group, not a path.** `group` (an InbredSet name) plus an
optional `genotype_version` makes the server resolve the genotype files
itself -- `GET /groups/{group}/genotypes` lists what is available and
which version is current, and `src/genotypes.rs` applies GeneNetwork2's
rule for that. `trait_values` may then be replaced by named `samples`,
which are matched onto the panel by name; parents and F1s are dropped
unless `include_parents_f1` is set, as in GeneNetwork2/3. `genotype_file`
+ `snps_file` remain as an explicit override, for files outside that
layout.

**If you name the files yourself**, they are BIMBAM, in GeneNetwork2's
own layout:
`genotype/bimbam/<file>_geno.txt` (the genotype matrix) together with
`<file>_snps.txt` (marker/chromosome). The `genotype/<file>.geno` that
R/qtl takes is *not* a GEMMA input. For BXD's current panel that is
`bimbam/BXD.8_geno.txt` + `bimbam/BXD.8_snps.txt` -- 235 individuals ×
21 056 markers. `trait_values` must then have **one entry per
individual, in the genotype file's column order**, `null` where the
trait has no value; GEMMA matches by position, so a short or reordered
list is a wrong answer, not an error. Expanding a trait's values onto
the panel is the caller's job today (the API does not re-index yet);
see [docs/genotypes.md](docs/genotypes.md) for the conventions, which
version counts as current, and the measured runs -- 17.1 s for the BXD
WGS panel (21 055 marker rows) against the 19 s the 7 320-marker
fixtures take.

This runs synchronously inside the request -- there is no job queue --
and it runs sequentially, not per-chromosome in parallel. **Not
feasible past roughly 100,000 markers at this point.** The BXD
fixtures (7,320 markers, LOCO across 20 chromosomes) already take
about 19 seconds; some real GeneNetwork datasets have upwards of 1
million markers, which this endpoint cannot handle yet -- the request
would simply run for a very long time inside one HTTP call. Making
that feasible needs either parallelising the per-chromosome LOCO jobs
(`gemma-wrapper --parallel`, which gn-guile's own precompute script
already uses) or moving the computation out of the request/response
cycle entirely (a job queue, or an offline precompute step, the way
gn-guile's own `scripts/precompute/run-gemma.scm` already does it).

```json
{
  "data": [
    [1.0, "rs31443144", 3010274.0, 0.0, "X", "Y", 0.276, 0.1533063, 0.1255921, 0.00001, -0.07665315, 0.2178407],
    ...
    [1.0, "rs8253327", 191908118.0, 0.0, "X", "Y", 0.284, 0.4281751, 0.1158946, 0.00001, -0.21408755, 0.0003520103]
  ],
  "meta": {
    "state": "complete",
    "took_ms": 18456,
    "request_id": "ade51981-4417-47cd-93d7-70dbd9da1ecc",
    "columns": ["chr", "rs", "ps", "n_miss", "allele1", "allele0", "af", "beta", "se", "l_mle", "additive", "p_lrt"],
    "num_rows": 7320,
    "num_cols": 12
  },
  "notes": [
    "Rows come from GEMMA's own .assoc.txt output, positional (see meta.columns for names); column names depend on the lmm value used.",
    "Kinship: gemma-wrapper --json --cache-dir /tmp/gn4-gemma --loco -- -g .../BXD_geno.txt.gz -p /tmp/gn4-gemma/492514-...-pheno.txt -a .../BXD_snps.txt -gk",
    "GWA: gemma-wrapper --json --cache-dir /tmp/gn4-gemma --loco --input /tmp/gn4-gemma/492514-...-k.json -- -g .../BXD_geno.txt.gz -p /tmp/gn4-gemma/492514-...-pheno.txt -a .../BXD_snps.txt -lmm 9"
  ],
  "next": []
}
```

(`data` has one row per marker -- 7,320 rows in this real run against
the BXD fixtures, trimmed to two here. Each row is positional, in
`meta.columns`' order, rather than a `{"chr": ..., "rs": ..., ...}`
object per row -- cuts the repeated key names out of a response this
large.

**The parental design is an endpoint, not a guess, and `/groups/{group}`
carries its headline.** That reply's note reads *"Parental design: C57BL/6J
(maternal) x DBA/2J (paternal), F1s B6D2F1 / D2B6F1; a mapping excludes the
parents and F1s by default"* — the two strain names answer "which cross is
this?" without a second request — and `GET
/groups/{group}/parents-and-f1s` reports a family's parents and F1s from the
deployment's `parents_and_f1s.json` (ported from GeneNetwork2's `ParInfo`, all
24 groups), plus the `@type` of its genotype file, the individuals each role
names when the metadata names them, and — the part a caller actually needs —
which members of the *current panel* the default rule drops, and which the `F1`
name heuristic catches. For an unlisted group it says `source: "name
heuristic"` and warns that its parents are unknown, so nothing is assumed. For
an RI set it also explains a subtlety: the metadata names the *design's* F1s
(`B6D2F1`) while the panel carries *specific* intercross F1 individuals
(`C57BL/6JxBXD037F1`), so the exclusions come from the name rule and
`excluded_by_name` lists precisely which.

**Choosing individuals.** Parents and F1s are excluded by default, as in
GeneNetwork2/3 (`include_parents_f1: true` turns that off). On top of that rule
any request that maps over a group's panel takes a selection:

```json
{"group": "BXD", "include_parents_f1": true,
 "individuals": {"include": ["C57BL/6JxBXD037F1"], "exclude": ["BXD48a"]}}
```

`exclude` wins over `include`, and an unknown name is a `422
invalid_individuals` whose message lists the panel — a silently dropped animal
changes a kinship matrix without anyone noticing. The reason both exist: the
parent/F1 list is *data* (`parents_and_f1s.json`, seven groups) that a request
cannot reach, while F1 names differ per group (`B6D2F1` for BXD, `BHF1`/`HBF1`
for BXH, `C57BL/6JxBXD037F1` for intercross F1s).

Worth knowing before turning it on: **including parents/F1s changes the
kinship, not the fit.** GEMMA computes kinship without reading phenotypes, so
they shape the relatedness structure; the association still uses only the
individuals that have a trait value. The reply says both numbers — *"197
individuals … ; the kinship uses all 197, the association uses the 33 with a
trait value"* — and a request can therefore legitimately have 234 individuals
in the genotype file and 34 in the fit.

`additive` is the effect size, sitting next to the score: **`−beta/2`**,
which is what GeneNetwork's own tables and CSV call the additive effect
(`gn2/wqflask/marker_regression/gemma_mapping.py`). Halved because half
the difference between the two homozygotes is the usual additive effect;
negated because GN2's BIMBAM conversion codes `@mat` as 1, so GEMMA's
`beta` is per *maternal* copy while GeneNetwork's published sign is
anchored on the other allele. `beta` and `se` are left where GEMMA put
them -- `se` is what makes the effect interpretable (`beta ± 1.96·se` is
a 95 % interval) -- and `p_lrt` stays the **last** column. These describe
*this run* on these samples and this genotype version; they are not the
`additive` that trait info reports, which came from the published
mapping (trait 20010: this run `−0.54`, published `+2.118`). The last two `notes` are the exact gemma-wrapper invocations
used (paths shortened here for readability) -- `-gk` only on the
kinship step, `-lmm` only on the GWA step, never both together; see
`src/gemma.rs`'s `compute_gwa` doc comment for why that distinction
matters. `lmm` defaults to 9, matching GeneNetwork2's and gn-guile's
own convention; see the field's own doc comment in
`src/routes/gemma.rs` for what's actually been verified about it.)

One further note is GEMMA's own variance component for the **null**
model, which it writes only when asked (`-pve`, which this build passes
on the GWA step; accepted alongside `-lmm 9`, verified against GEMMA
0.98.6). Under `--loco` there is one per chromosome -- twenty different
kinship matrices -- so the note carries **every** value, not a summary:

    GEMMA -pve, null model, per chromosome (LOCO): chr1 0.6674±0.3451,
    chr2 0.8426±0.3668, ..., chrX 0.8563±0.3525. That is pve 0.3118 to
    0.9549 over 20 chromosomes; lowest chr8 pve 0.3118 ± 0.5009 (vg
    23.3659, ve 12.3389), highest chr7 pve 0.9549 ± 0.3570 (vg 76.0933,
    ve 0.8661).

The unrounded numbers are also appended to the run log as `variance:
chr7 pve=0.954938 se_pve=0.357032 vg=76.0933 ve=0.86612` lines, so a
`?id=` run's `/compute/gemma/status` stream shows them while the run is
still in flight; and gemma-wrapper keeps each chromosome's own
`.log.txt` beside its archive (`5eb53fa`), named by the same hash. `pve`
is the proportion of variance the kinship explains -- the same quantity
as `/compute/bulklmm`'s `h2`, but per chromosome instead of genome-wide --
and it is scale-free, so compare it with `h2` rather than comparing `vg`
against `σ²_e`: a trait measured in thousandths has variances in
thousandths. When the estimate sits on a boundary GEMMA prints `-nan`
for `se(pve)`; that is reported as "not available", because a boundary
estimate has no standard error, and a `pve` of exactly 0 or 1 is a
direction rather than a proportion.

#### `/compute/gemma/status/{run_id}` -- polling a run's progress

Given the endpoint above can genuinely take minutes, pass `?id=` on
the `POST` and poll this on a second connection while it's in flight
to show progress, the same log-tailing shape GeneNetwork2/3 use for
R/qtl (`gn3/computations/streaming.py`, `gn3/api/streaming.py`): the
run's `id` becomes a log file (`{cache_dir}/{id}.log`); gemma-wrapper's
own stderr (its `--json`-mode progress banner, chromosome list,
"Executing: ...", per-step info -- its stdout carries only the one
final JSON record, so there's nothing to stream there) is appended to
it line by line as both the kinship and GWA steps run; once the whole
request finishes, `"GN4: computation complete"` is appended as a
sentinel so a poller can tell "still running" from "done". `run_id`
must be 1-64 ASCII letters/digits/`-`/`_` -- it's a path component, so
this is enforced on both this endpoint and the `?id=` that names it.
`id` must also be unique: if a log already exists for it (another run
still in progress, or a finished one never polled/cleaned up), the
`POST` is refused with 409 rather than silently overwriting it.

```
POST /compute/gemma?id=my-run-1       (on one connection; blocks until done)
GET  /compute/gemma/status/my-run-1?peek=0       (on another, while the first is in flight)
```

```json
{
  "data": {
    "run_id": "my-run-1",
    "lines": [
      "gemma-wrapper 1.00-pre1 (Ruby 3.3.9) by Pjotr Prins 2017-2024",
      "Using GEMMA 0.98.6 (2022-08-05) by Xiang Zhou, Pjotr Prins and team (C) 2012-2022",
      "Executing: time -v .../gemma -o ... -g ... -p ... -a ... -gk -outdir ..."
    ],
    "pointer": 612
  },
  "meta": {"state": "complete", "took_ms": 1, "request_id": "..."},
  "notes": [],
  "next": []
}
```

Pass `pointer` back as `?peek=` on the next poll; `lines` is empty
(not an error) when nothing new has been written since. An unknown
`run_id` (never started, or the main request hasn't written its log
yet) is a 404, not an empty response -- genuinely different from
"nothing new yet".

#### `/compute/gemma/ravanan` -- the same mapping, run per-chromosome via ravanan

Same request/response shape as `/compute/gemma` -- `group` (plus
`genotype_version`), an uploaded genotype (`genotype_upload`), or
`genotype_file` + `snps_file`; values as `samples` (named, aligned by
name) or `trait_values` (positional), or a `basket` -- resolved by the
same code as the direct runner (`resolve_inputs`), so the two cannot
drift. Two fields are **refused rather than ignored**, because this
workflow has no step for them: `covariates` (`422
covariates_not_supported`, use `/compute/gemma`) and `normalize` (`422
normalization_not_supported`). The same two steps (LOCO kinship, then
LOCO GWA), but run by [ravanan](https://github.com/vepiteski/ravanan)
driving a CWL workflow compiled on every call (via `ccwl compile`, see
below) from the ccwl source at `cwl/gemma/gemma-loco.scm`, instead of
this server's own direct subprocess calls -- an alternative runner for
the same computation, kept side by side with `/compute/gemma` so the
two can be compared (correctness, timing).

**Ravanan's cache is only as durable as `RAVANAN_CACHE_DIR`.** Ravanan's reuse
key is `(script, inputs)`, and both live in its `--store` under that directory,
so the cache survives a container restart if — and only if — the directory is on
a host path the container bind-mounts. Measured: 13.6 s cold, then 3.6 s for the
identical request after restarting the container, store unchanged. The default
is under `/tmp`, which is neither durable nor visible inside the container; the
reply carries a note when it sees that. Same reasoning for `GN4_API_TOKENS_DB`
(`./tokens.sqlite` by default) and `GEMMA_CACHE_DIR`. See
[docs/ravanan-slurm.md](docs/ravanan-slurm.md#what-survives-a-restart).

**K is shared across traits.** The `kinship` step is given a constant-valued
phenotype file carrying only the trait's missing-data pattern, because
`gemma -gk` uses `-p` solely to learn which individuals are analysed; `gwa`
gets the real values. Two traits over the same animals therefore reuse all
twenty K matrices -- on the BXD fixture a second trait adds ~1 MB of store and
20 GWA jobs rather than ~13 MB and a second kinship scatter.

**This runner calls `gemma` itself, not `gemma-wrapper`** (the direct
runner still calls the wrapper, which is what makes the comparison
worth making). GEMMA 0.98.6 has `-loco <chr>` natively for both `-gk`
and `-lmm`, taking the chromosome to leave out by name and requiring
the BIMBAM annotation file (`-a`) that every genotype shape here
supplies -- so gemma-wrapper's LOCO orchestration only duplicates the
tool. Its parallel fan-out was never used on this path anyway: ravanan
scatters one `gemma -gk|-lmm ... -loco <chr>` job per chromosome, and
the kinship matrix flows from the kinship job to the GWA job as a CWL
`File` output rather than through a wrapper cache. The wrapper's
content-hash K cache is the one thing given up; a repeat request
recomputes, which is what the per-chromosome scatter costs anyway. This
workflow is written in
[ccwl](https://ccwl.systemreboot.net/) at `cwl/gemma/gemma-loco.scm`,
not raw CWL; `compute_ravanan` compiles it to CWL with `ccwl compile`
fresh on every call (cheap -- pure macro expansion, no Guix build
involved) into a generated, throwaway file under `ravanan_cache_dir`,
never checked into the repo. `compute_ravanan` computes the chromosome
list from `snps_file` itself (`chromosomes_from`) and passes it as the
workflow's `chromosomes` input, since CWL's `scatter` needs a concrete
array at invocation time -- and GEMMA's `-loco` takes exactly that
chromosome name. `ravanan_cmd`, `ccwl_cmd`, `gemma_ravanan_workflow`
(the ccwl source, not a CWL file), `gemma_ravanan_manifest`
(ravanan's own required `--guix-manifest`, a Guix manifest naming
`gemma`, at `cwl/gemma/manifest.scm`) and `ravanan_cache_dir`
(its own cache directory, separate from `gemma_cache_dir`, for pheno
files, run logs, the compiled workflow, and ravanan's own `--store`)
are configurable the same way as the rest of `src/config.rs`'s layered
config. Its genotype and SNP inputs follow the same convention as
`/compute/gemma` -- the `bimbam/<file>_geno.txt` + `<file>_snps.txt`
pair, not the `.geno` file (see [docs/genotypes.md](docs/genotypes.md)) --
and the values must likewise line up one for one with that file's
individuals.

ravanan's batch system is configurable too: by default each CWL step is
a local child process, and `RAVANAN_BATCH_SYSTEM=slurm-api` submits
every step to a Slurm cluster through slurmrestd instead (endpoint, JWT
file, partition, nice and scratch directory via `RAVANAN_SLURM_*` /
`RAVANAN_SCRATCH`). With slurm, `ravanan_cache_dir` must be on a
filesystem the compute nodes can see, since ravanan's `--store` and
scratch live there. See [docs/ravanan-slurm.md](docs/ravanan-slurm.md)
for the keys, the fail-closed rules and the slurm-side auth
requirements, and [cwl/README.md](cwl/README.md) for how a request gets
from the API to a job on a node (and what the nodes themselves need).
`GET /compute/gemma/status/{run_id}` is shared between both endpoints
and checks whichever cache directory the `run_id` actually logged to.

Verified end to end on 2026-11-12 against the Slurm container from
`gn-machines` (`specials/slurm-container.scm`): a `group: BXD` request
became 40 Slurm jobs (20 kinship + 20 GWA, scattered over the SNP map's
chromosomes), all completed, 21 055 rows. See
[docs/ravanan-slurm.md](docs/ravanan-slurm.md) for what the host and the
nodes must provide -- including the two requirements that bit during
that run: `ravanan` must be a build with the `--slurm-api-endpoint`
fix, and the *server's* `HOME` must be writable by the job user inside
the cluster.

The reply carries the same variance-component note `/compute/gemma` does
(`-pve`, every chromosome). Since this workflow runs `gemma` itself,
each job's `.log.txt` is a first-class CWL output and the note is read
straight out of those files -- no archive to unpack and no wrapper-named
copy to depend on. See [docs/ravanan-slurm.md](docs/ravanan-slurm.md).

**Concurrency.** Like `/compute/gemma` -- same 50-individual threshold,
since ravanan writes into the same cache directory -- only a large request
claims a slot. One ravanan workflow at a time, one
per caller: a second large submission while one runs (or is queued) is
refused at once with `409 compute_busy`, naming the run and linking
`/compute/gemma/status/{id}` (which ravanan shares with `/compute/gemma`);
a submission from another caller waits for the slot. One slot because
ravanan already fans its own
per-chromosome steps out -- over the cluster, when the slurm backend is
selected -- so a second request would only queue behind the same scratch
directory. Its gate is separate from `/compute/gemma`'s: they are
different runners of the same computation with their own cache
directories.

Exploratory, and genuinely rough around the edges -- three real bugs
turned up building this, none of them in `/compute/gemma`'s own
direct-subprocess path:

- ravanan resolves `gemma-wrapper` by building a Guix profile from
  `--guix-manifest` against the *ambient* Guix process
  (`GUIX_PACKAGE_PATH`, no `--guix-channels`) rather than spinning up a
  separately pinned inferior -- the latter forces a whole-channel
  package-cache rebuild that breaks on any unrelated package in
  `guix-bioinformatics` too new for the pinned base Guix. This needs
  `GUIX_PACKAGE_PATH` set in whatever environment starts the
  genenetwork4 server itself (inherited by the `ravanan` child process
  genenetwork4 spawns); it is not something `compute_ravanan` sets.
- ravanan's `--store` needs `files/`, `logs/` and `data/`
  subdirectories to already exist; it does not create them itself
  (`compute_ravanan` does now).
- A step scattered over exactly one input (`kinship`'s `chr`) needs an
  explicit `scatterMethod` even though CWL itself only requires one
  for *multiple* scattered inputs -- ravanan's scheduler has no
  fallback case for a missing one and silently misbehaves without it.
- Every chromosome's `gwa.cwl` run shares gemma-wrapper's own
  content-hash basename for its output archive (the hash doesn't
  factor in `--chromosomes`), so the 20 scattered `archive` outputs
  all collide under ravanan's optional `--output-directory`
  symlink-everything-into-one-flat-dir step; `compute_ravanan` simply
  doesn't pass that flag, since it was already reading every output's
  path straight out of ravanan's own `--store` via its stdout JSON
  record, never through `--output-directory`.

A real end-to-end run against the BXD fixtures (same rows as the
`/compute/gemma` example above, confirmed byte-for-byte including
`rs31443144`'s exact values) took **16.7s** wall-clock scattered across
20 chromosomes this way, vs. `/compute/gemma`'s own ~19s and the
earlier sequential (non-scattered) ravanan workflow's ~21s -- a real,
if modest, win from letting ravanan's own scheduler fan the
chromosomes out instead of going through gemma-wrapper's internal GNU
`parallel` call, on top of no longer needing that path's own
`PARALLEL_SHELL` workaround at all (each scattered job's
`--no-parallel` gemma-wrapper call never invokes GNU `parallel` in the
first place, so the bash-`time`-keyword bug it worked around does not
arise here).

Those 16.7s are with ravanan's default **`single-machine`** batch system,
where a step is just a local child process. With
`RAVANAN_BATCH_SYSTEM=slurm-api` every step additionally pays Slurm
dispatch -- submit, scheduler tick, launch, poll -- which on a small
input dwarfs the mapping itself: three chromosomes, measured on the test
container (`gn-machines/specials/slurm-container.scm`, one node, two
CPUs), are **2.3s** cold / **0.7s** warm on `/compute/gemma`,
**4.7s** / **3.5s** on ravanan single-machine, and **33s** on
ravanan+Slurm with a 5s scheduler interval (**158s** with Slurm's
defaults), while each job's own `gemma-wrapper` run took under a second.
Same rows in every case. See [docs/ravanan-slurm.md](docs/ravanan-slurm.md)
for the breakdown and what to tune.

### `/compute/qtl2`

A genome scan with **R/qtl2**: genotype probabilities from a hidden Markov
model, and a linear mixed model with leave-one-chromosome-out kinship. It is
the middle ground between `/compute/rqtl` (Haley-Knott, no kinship) and
`/compute/gemma` (an LMM, but biallelic BIMBAM with mean-imputed genotypes),
and it is how multiparent populations are analysed.

```json
{
  "group": "BXD", "genotype_version": "BXD.4",
  "trait_name": "BXD::10001",
  "samples": [{"name": "BXD1", "value": 61.4}, …],
  "kinship": "loco",        // loco (default) | overall | none
  "model": "normal",        // normal (default) | binary
  "cores": 4
}
```

Rows are `marker, chr, pos (Mb), lod` — the same shape `/compute/rqtl`
returns, so **`/plot/rqtl` draws it unchanged**.

On DOL trait 10003 the three engines tell the same story with different
emphasis, which is what modeling relatedness buys: R/qtl's HK scan peaks on
chr9 at LOD 5.86 and does not see chr16 at all; GEMMA's LMM prefers chr16 at
−log10 p 7.15; R/qtl2's LMM finds chr9 (LOD 4.85), chr19 (4.76) and chr16
(4.24) on one footing.

**Autosomes only.** The X needs a sex covariate to be fitted properly, and
the Y and mitochondrion degenerate the per-chromosome kinship — GEMMA's LOCO
run is given the same restriction, for the same reason. The `notes` say how
many markers were left out.

**Multiparent crosses.** A multiparent population cannot come from a
`.geno`: that file carries two codes per marker (`B`/`D`, `A`/`H`/`B`), which
is a collapse, not a founder allele. It has to ship its own qtl2 control file
with founder genotypes, so the endpoint also takes a cross by name:

```json
{"rqtl2_cross": "do", "trait_name": "DO::WBC",
 "samples": [{"name": "JAXW202.2129", "value": 4.34}, …]}
```

The name is looked up as `genotype/rqtl2/{name}.json`, then
`{name}/{name}.json`, then `{name}.json` under `GENOTYPE_FILES`; `do/` is the
one such cross this deployment has (742 individuals, 8 founders A–H, 6 413
markers, WBC and NEUT). The samples are matched to the cross by name and
anything outside its population is left out (and counted in `notes`).

A multiparent scan is a **different claim** from a biallelic one: every marker
is fitted with one coefficient per founder, so the LOD asks whether *any*
founder differs, not which. Reading a multiparent LOD as a strain effect is the
usual mistake; per-founder effect sizes need `scan1coef`, which this endpoint
does not return yet.

Measured on the DO (`WBC`, 8 cores): genotype probabilities 174 s and 1.39 GB,
cached per cross and keyed by the control file *and its directory*; LOCO
kinship 15 s, cached; the scan itself 21 s. `scan1` is run on one core for a
multiparent cross deliberately — its forked path stalls here on allele
probabilities (25+ minutes with 8 cores and no CPU consumed, against 21 s
serial). The autosomes-only rule applies here too: the DO's X is left out (369
markers) rather than fitted without a sex covariate.

**Panel conversion.** The cross is generated from this deployment's own
`.geno` file and cached (`QTL2_CACHE_DIR`), because the shipped
`genotype/rqtl2/` control files are not usable: they name paths relative to a
directory the data is not in, declare no genotype-code map (every genotype
reads as missing) and no orientation (individuals read as markers) — four
traps that each produce a plausible flat scan rather than an error. The
conversion also drops heterozygous calls on a `riself` panel, as R/qtl's
`convert2riself` does. Configuration: `QTL2_SCRIPT_PATH`,
`QTL2_CACHE_DIR`.

### Gene link-outs

A gene symbol is useful beyond our own pages, and SKILL.md carries the table of
external services that still resolve (checked 2026-11-11): RGD for **both mouse
and rat**, MGI, the Alliance of Genome Resources, NCBI Gene, Wikidata (the QID
that `/genes/{symbol}/aliases` already returns), UniProt, Ensembl, GTEx,
STRING, GeneMANIA, Open Targets, the Human Protein Atlas, the GWAS Catalog,
BioGPS, the Allen Brain Atlas and GeneCup — with the retired ones named so they
are not offered by mistake (HomoloGene, UniGene, Genotation, GeneBridge,
PhenoGen).

Two conventions that keep the links honest: **quote the species** the symbol came
from, and prefer RGD/GeneMANIA(rat) for a rat symbol and GTEx/Open Targets/
Protein Atlas for a human one — a mouse symbol in GTEx resolves to an empty
report, because GTEx is human-only.

### Why R/qtl and R/qtl2 disagree on a `riset` panel

They read the same file and can end up fitting different predictors, so a peak
in one is not a bug in the other. The mechanism, with BXD trait 16627 as the
worked case (BXD.8, 35 phenotyped animals, all measured on production):

| marker | Mb | R/qtl | R/qtl2 `riself` | R/qtl2 `f2` |
|---|---|---|---|---|
| `rsm10000000889` | 5.652 | 9.59 | 3.79 | 9.98 |
| `rs33299937` | 5.653 | 9.59 | 3.73 | 9.68 |
| `rs13476332` | 6.311 | 9.59 | 3.70 | — |

**What R/qtl does.** `scripts/rqtl_wrapper.R` reads the cross with
`genotypes = c(@mat, @het, @pat)` — codes 1, 2, 3 — and then, for a `riset`
panel, calls `convert2riself`, which is *adaptive*:

```r
gtab <- table(pull.geno(cross))
usethree <- !(gtab["3"] < gtab["2"])   # is code 3 at least as common as code 2?
if (usethree) { dat[!is.na(dat) & dat != 1 & dat != 3] <- NA; dat[dat == 3] <- 2 }
else          { dat[!is.na(dat) & dat != 1 & dat != 2] <- NA }
```

It keeps the **two most numerous codes** and drops the third (warning
`Omitting N genotypes with code==2`), then recodes to 1/2. No HMM, no
probabilities: `scanone` fits each marker's own calls.

**What R/qtl2 does.** `riself`'s state space is `AA, BB` (two states, no het),
so a heterozygous call is not representable and becomes missing; the HMM then
re-imputes it from flanking markers, and `scan1` fits *probabilities*, not
calls. On BXD.8 the conversion drops **610 684 calls, 12.7 % of the matrix**
(the reply says so). At a marker with 14 % H, that is a different predictor
from R/qtl's, which is why the peak survives in one engine and thins out in the
other — not a failure of either.

`genotype_model: "f2"` keeps the calls as a third genotype (`AA, AB, BB`) and
recovers R/qtl's number (9.68 against 9.59). It is a **diagnostic**: it fits
het as a real genotype *and* assumes an outbred intercross design, which an
inbred line is not. Its value is that it separates "the conversion cost the
information" from "the model cannot see the locus".

**What the proper model is, and where it stands.** qtl2 has general RIL types —
`genril2` / `genail2`, from qtl2 0.18: *“handle any number of founders; include
the number of founders in the cross type… `genril` assumes the progeny are
inbred lines (RIL), while `genail` assumes the progeny have two random
chromosomes (AIL)”*, with `cross_info` of length `1 + n_founders`: the first
column the number of **generations**, the rest founder frequencies. Both are
supported by the qtl2 0.42 this deployment runs (checked through
`qtl2:::check_crosstype()`), and `genril2` is the model that fits a `riset`
panel's residual heterozygosity honestly.

It is **not wired yet**, and the reason is concrete rather than a matter of
effort: a founder-based cross type expects individual genotypes coded as
*alleles/pairs* of founders (as the DO's `A`–`H` are), not the `1/2/3` of a
two-parent cross; a prototype with numeric codes had every genotype read as
missing. The remaining work is therefore exact: build `founder_geno` from the
two parental alleles (for a two-parent cross every marker is homozygous
founder 1 vs founder 2 by definition of the coding), write `cross_info` with
the generation count, and code `H` as the founder *pair* rather than code 2.
Then compare trait 16627 against 9.59/9.68 — and check DOL through the same
path, because its `H` is not residual heterozygosity at all: it is a collapsed
DO, where `H` means "two different founders" and the honest model is
`crosstype: "do"` with founder genotypes the collapse destroyed.

**What to do when the engines disagree**, in order: `.../markers/{marker}/classes?trait=`
to see the class sizes and means; `.../frequencies` for the H share at the
marker; `genotype_model: "f2"` to see what keeping the calls would give; and
quote which model ran, because a reply says `Cross: riself` or `Cross: f2`.

### `/compute/bulklmm`

A genome scan that can be asked for **thresholds**. GEMMA's LOCO scan is the
reference for a biallelic panel, but a GEMMA permutation repeats the whole
scan, which is why `/compute/gemma` refuses `nperm`; the usual stand-in is
R/qtl's `nperm`, whose null shuffles phenotypes as if the individuals were
unrelated. [BulkLMM.jl](https://github.com/senresearch/BulkLMM.jl) fits one LMM
per marker with a global kinship matrix and permutes by *rotating* the data
through the kinship's eigen decomposition — relatedness included — so a
threshold costs seconds:

```json
{"group": "BXD", "genotype_version": "BXD.8", "trait_name": "BXD::16627",
 "samples": [{"name": "BXD1", "value": 1.5}, …], "nperm": 200, "seed": 42}
```

BXD trait 16627, 35 phenotyped animals, 21 056 markers, 22 s:

```
peak rsm10000000888 chr2 4.92 Mb  LOD 9.59  beta 0.0027  h2 3.0e-16
permutation thresholds (200 rotations, seed 42): alpha 0.10 → 3.51, 0.05 → 3.81, 0.01 → 4.47
```

which is the trait's published locus (`PublishXRef.Locus` `rsm10000000888`, LOD
9.59) and a genome-wide-significant peak against a threshold this API could not
compute before. Rows are `marker, chr, pos (Mb), lod, beta, h2`.

**Thresholds on another engine's plot.** `/plot/gemma` draws **no reference
line by default** — a line on a plot is an annotation, not part of the result,
and a reader should be told where each one came from. GEMMA's Bonferroni line is
one more opt-in labelling (`?bonferroni=true`); a threshold computed here
travels with the plot, labelled, because the two are different nulls:

```
POST /plot/gemma?thresholds=3.50,3.65&threshold_labels=bulklmm%20alpha%200.10,bulklmm%20alpha%200.05
{ the /compute/gemma reply }
```

Each line is drawn dashed in the series colour with its label and value
(`bulklmm alpha 0.05: 3.65`), and the Bonferroni line only appears when it is
asked for — so a plot shows exactly the lines someone chose, each named.
`/compute/bulklmm`'s reply already offers that href with its α = 0.05 threshold
filled in, so the picture is one link away.

`/plot/bulklmm` draws two reference lines for it: the genome-wide α = 0.05 and
the stricter α = 0.01. The reply also reports α = 0.10, but that is the loosest
bar in the set — drawing it invites reading it as the line to clear — so it
appears only on request (`?all_thresholds=true`, or an explicit
`?thresholds=3.50&threshold_labels=alpha 0.10`). GEMMA's Bonferroni line is
opt-in in the same spirit: a reference line is an annotation, not a result.

**About cores.** BulkLMM's single-trait `scan()` — what this endpoint calls —
is **serial in Julia by design**: the package threads `bulkscan()` (many traits)
and its helpers, not `scan()` (checked in its source: `Threads.@threads` appears
in `bulkscan.jl` and `bulkscan_helpers.jl`, not in `scan.jl`). What was
parallel before was BLAS, not the scan. The driver therefore splits the markers
into independent blocks and scans them on Julia threads (`--blocks`, default =
`BULKLMM_THREADS`), which is exact — the LODs and permutation thresholds are
identical to a serial run, verified bit-for-bit at 1 and 4 blocks — and caches
the parsed matrix and the kinship per panel:

    cold, 1 block : read 1.3 s, scan 9.0 s, total 20.9 s
    cold, 4 blocks: read 0.4 s, scan 7.5 s, total 18.3 s   (same LOD 9.589938)

The gain is modest at this size because the LMM loop is only about 8 s of the
run and each block re-estimates the variance components; the rest is Julia
startup. **Where the cores pay is `bulkscan()`** — a hundred traits in one
process, threaded over trait blocks. That is the shape an eQTL endpoint wants,
and it is the next thing worth adding here.

Two differences from GEMMA are stated in every reply rather than implied:

* **a global kinship matrix, not leave-one-chromosome-out**, so its p-values are
  not GEMMA's;
* **`h2_null` says how much relatedness it found**. Here it is ~0, so the fit is
  effectively a fixed-effects model — which is exactly why it agrees with R/qtl
  on this trait where GEMMA's LOCO scan differs.

See [docs/bulklmm.md](docs/bulklmm.md) for the model comparison, the measured
core behaviour and the packaging options. The engine is a Julia subprocess
(`BULKLMM_CMD`, `BULKLMM_ENV`, `BULKLMM_DEPOT`,
`BULKLMM_SCRIPT`, `BULKLMM_THREADS`, `BULKLMM_CACHE_DIR`) with the driver in
`scripts/bulklmm_scan.jl`. The Julia depot is prepared once, deliberately, and
is small: 79 packages, 139 MB, no GPU/CUDA stack — BulkLMM's closure is CSV,
DataFrames, Distributions, Optim and what those need. Every request runs with
`JULIA_PKG_OFFLINE=true`, so a scan can never download anything.

### `/compute/bnw`

A Bayesian network over a set of variables, learned by BNW (the Bayesian
Network Webserver): which traits belong together under a causal ordering. The
computation runs in BNW's own headless driver — the `bnw` program in the
`gn-bnw` repository — so this endpoint only prepares a matrix, calls out, and
parses the reply (`BNW_CMD`, `BNW_BIN_DIR`, `BNW_CACHE_DIR`).

```json
{
  "group": "BXD", "traits": [10001, 10003, 10038],
  "maxparents": 4, "k": 1, "tiers": 3, "threshold": 0.5
}
```

or, for a matrix the database does not hold (a basket's columns, a handful of
probe sets):

```json
{
  "values": [
    {"name": "Trait1", "samples": [{"name": "BXD1", "value": 1.2}, …]},
    {"name": "Trait2", "samples": [{"name": "BXD1", "value": 0.4}, …]}
  ]
}
```

Rows are positional with `meta.columns`:

```
from, to, posterior, in_structure
```

One row per candidate edge (variables²−variables of them: self-edges are
dropped), strongest posterior first. `posterior` is the share of the *k*-best
networks that contain the edge; `in_structure` says whether it survived
`threshold`. The four parameters default to BNW's own (`maxparents` 4, `k` 1,
`tiers` 3, `threshold` 0.5 — from `create_tiers_gom_part1.php`), and the reply
names the values used, because a network at threshold 0.5 and one at 0.9 are
different answers.

**A posterior is not a p-value.** It says how often the search put that arrow
there in *this sample* — a statement about correlation under a causal
ordering, not about surprise, and not an effect size. At most 40 variables
(the search is exponential in in-degree) and at least 3 samples complete across
every variable: BNW models a complete multivariate normal, so a sample missing
any value is dropped and the reply says how many survived.

`POST /plot/bnw` draws it from the posted reply, the same way `/plot/gemma`
does — an SVG back, nothing recomputed, the plotter's one-line summary in
`x-plot-summary`. The default figure is the **network**: nodes on a circle, one
arrow per edge that survived the threshold, arrow width and label = the edge
posterior, and the below-threshold edges as faint dotted lines so a reader sees
what was *considered* as well as what survived. `?kind=posterior` draws every
candidate edge as a cell in a matrix instead, which is the view that says how
close a rejected edge came to the threshold. Pure-Python SVG (`BNW_PLOTTER`,
run by `PLOT_PYTHON`), so no graphviz is needed in the deployment.

**From a network to a mappable trait: `POST /compute/bnw/features`.** Every
node with at least one parent at the threshold is a *module* — the node and its
parents — scored per sample by the **first principal component** of the module's
standardised members, sign-corrected so the score rises with the node. Without
`?module=` the reply is one row per module, `module, members, size,
variance_explained` (the share of the module's variance the component explains,
which says whether the score is a fair summary). With `?module=NAME` it is that
module's per-sample scores in the `{name, value}` shape every mapping endpoint
already takes — post them as `samples` to `/compute/rqtl`, `/compute/gemma` or
`/compute/qtl2` and map the module like any trait. That is the chain BNW exists
for in GeneNetwork: structure → module → QTL. A module is a correlation
structure, not a QTL: what maps is where its members' shared variation comes
from, not which member causes which.

**BNW's own worked examples: `GET /compute/bnw/examples`** and
`GET /compute/bnw/examples/{name}`. The five datasets its tutorial and examples
pages present (synthetic 2-genotype/6-trait, the scientific 5-node network at
100 and 500 samples, BXD spleen immune genes, the time series) are listed with
their sizes, and each one is served as a matrix in the `values` shape
`/compute/bnw` takes — with a ready-to-post action carrying the body. `?columns=`
restricts the matrix, which the spleen example needs (14 variables, and a caller
may want fewer). `genotype_column` warns that a column is a genotype code: this
driver is continuous-only, so such a column enters the network as a number, as
the web application's GeneNetwork path treats it. The files are read from
`BNW_EXAMPLES_DIR` (a gn-bnw checkout's `sourcecodes/example_datasets`); a
deployment without it answers `404 no_examples` rather than pretending.

The driver's contract is documented in
[gn-bnw/info_files/gn4-integration-plan.md](../gn-bnw/info_files/gn4-integration-plan.md);
`tests/bnw.rs` checks the endpoint against the answer the BNW *web
application* gives for its own six-trait example, so "it works" means the same
edges, not merely a 200.

### `/compute/fitqtl`

`/compute/rqtl` says *where* a QTL is (a LOD per marker); this says *how
much*, by fitting R/qtl's `fitqtl` at the markers you name -- normally the
peaks a scan already reported, via `GET
/groups/{group}/genotypes/{version}/markers`. It is separate because it is a
separate question, asked about a few positions rather than a genome: `fitqtl`
fits one model per marker, so a genome-wide version would be thousands of
fits (the cap is 100 markers per request).

```json
{
  "group": "BXD", "genotype_version": "BXD.4",
  "trait_name": "BXD::10001",
  "samples": [{"name": "BXD1", "value": 61.4}, …],
  "markers": ["mCV24409492", "rs8253516"]
}
```

Rows are positional with `meta.columns`:

```
marker, chr, pos, lod, pct_var, additive, se, t, p
```

`additive` is fitqtl's estimate for the QTL term. R/qtl codes the QTL
genotype ±1, so this is half the difference between the two homozygote
classes -- the same convention GeneNetwork's own `additive` uses, and on
BXD trait 10001 it reproduces the published value to six decimals
(`2.394472` here against GN2's `2.394444`, from the same peak region).
`pct_var` is the model's percentage of variance explained and `p` its
p-value; `se` and `t` belong to the effect. Markers not in the genotype
file, or that fitqtl cannot fit, are absent rather than reported as zero.

**Covariates** work exactly as on `/compute/rqtl` -- `covariates:
[{"case_attribute": "Status"}]`, inline values, or an explicit `map` -- and
are fitted alongside the QTL, so the effect reported is the one *after*
conditioning. Two details worth knowing: a covariate that is spread across
the panel but constant among the individuals a trait covers is left out, with
the columns named in `notes` (R/qtl stops with "system is exactly singular"
otherwise), and the LOD then no longer equals the scan's, because it is a
different model.

The cross-check that matters: without covariates, `fitqtl`'s `lod` at a
marker **equals** `scanone`'s LOD at that marker, to every digit the CSV
carries -- the test suite pins it. Same alignment rules as `/compute/rqtl`
(values matched onto the panel by name, parents/F1s excluded unless
`include_parents_f1`), the same `normalize`/`winsor_*` pipeline (and the
effects then come out on that scale), and the same R/qtl gate. Configuration: `FITQTL_SCRIPT_PATH`
(default `scripts/qtl_fitqtl.R`), `RQTL_CACHE_DIR` for its scratch files.

### `/compute/rqtl`

Maps a QTL by running GeneNetwork3's own `scripts/rqtl_wrapper.R` (via
`Rscript`), copied into this repo verbatim: a `scanone` genome scan,
plus permutation thresholds if `nperm > 0`. Same request/response shape
as `/compute/gemma` -- synchronous, `?id=`/`GET .../status/{run_id}`
progress-streaming reusing the same infrastructure (`src/streaming.rs`)
-- see `src/rqtl.rs` for the invocation itself, its documented scope
(no interval mapping or pair-scan `scantwo` -- just the common
single-scan case), and a fix made relative to GeneNetwork3:
its own cache-key hashing of the genotype/phenotype files never
actually looks at their contents (a key-name mismatch between
`generate_rqtl_cmd` and its real caller), so two different phenotypes
mapped with the same other options could silently reuse each other's
cached result there; fixed here by always hashing both files' contents.

**Covariates.** The same `covariates` list as `/compute/gemma` (above)
-- case attributes by name, or per-sample values -- fitted here with
`--addcovar` and a generated `--covarstruct` file. R/qtl's wrapper takes
categorical covariates as a single column and dummy-expands more than two
levels itself, so a level with no observations is its problem, not a
singular column: no columns are dropped on this path. A missing value is
`x` (the marker the wrapper calls missing), so R/qtl drops those
individuals; the coding and the count of complete cases come back in
`notes`.

**Give it a group, not a path.** `group` (an InbredSet name) plus an
optional `genotype_version` resolves the `.geno` server-side; see
`GET /groups/{group}/genotypes`. `samples` are then matched onto the
panel **by name** -- unknown names dropped, missing individuals `null` --
and parents/F1s are excluded unless `include_parents_f1` is set. The
resolved file, version, panel size and how many values aligned come back
in `notes`, so a run is reproducible without the caller ever naming a
path. `genotype_file` is still accepted as an explicit override.

**If you name the file yourself**, R/qtl takes the GN genotype file
itself -- `genotype/<file>.geno`, e.g. `genotype/BXD.8.geno` (235
individuals × 21 057 markers) or `genotype/HET3-ITP.12.geno` (6 447
individuals × 894 markers) -- and reads its format directly. Which
engine a group offers, and which of its genotype files is current, are
GeneNetwork2/3 conventions summarised in
[docs/genotypes.md](docs/genotypes.md); `InbredSet.MappingMethodId` is
what decides the engine (BXD = both, HET3-ITP = R/qtl only -- UM-HET3
is a 4-way cross). Measured on this deployment: BXD with the WGS panel,
235 samples of which 34 carry values, 21 056 markers in 7.5 s; UM-HET3
lifespan over 6 447 individuals, 893 markers in 83 s, peak LOD 7.95 at
chr 1, 24.04 Mb.

**Important:** R/qtl's cross file format has no per-sample name
matching of its own -- `samples` must list values in exactly the
genotype file's sample-column order, and it must have as many entries
as the genotype file has individuals. The second half of that used to go
unchecked (as it does in GeneNetwork3): nine values scanned against a
six-individual cross came back as a complete, plausible-looking set of
LOD scores belonging to nobody. `scripts/rqtl_wrapper.R` now stops with
"the phenotype file has 9 samples but the genotype file has 6
individuals" instead, which surfaces as a `502 rqtl_failed`. It still
cannot tell whether the *order* is right -- that is the caller's
responsibility, and the reason a trait's values and its genotype file
have to come from the same panel.

```json
{
  "data": {
    "markers": [
      ["rs31443144", "1", 3.010274, 0.0923496],
      ["rs6269442", "1", 3.492195, 0.1131095]
    ],
    "permutation": null
  },
  "meta": {
    "state": "complete",
    "took_ms": 420,
    "request_id": "...",
    "columns": ["name", "chr", "pos", "lod"],
    "num_rows": 6,
    "num_cols": 4
  },
  "notes": [
    "markers come from R/qtl's scanone output, positional (see meta.columns for names); permutation is null unless nperm > 0 was requested.",
    "Rscript scripts/rqtl_wrapper.R --geno .../rqtl_sample1.geno --pheno .../pheno.txt --model normal --method hk --nperm 0 --scale mb --outdir /tmp/gn4-rqtl --filename <hash>-output.csv"
  ],
  "next": []
}
```

(Real output, against the tiny 6-marker/6-sample BXD riset fixture in
`tests/test_data/`, trimmed to two markers. Each marker row is
positional, in `meta.columns`' order, rather than a
`{"name": ..., "chr": ..., ...}` object per row -- same
space-saving shape `/compute/gemma` uses for its own (much larger)
rows; a 4-way cross's extra mean/SE effect columns, if present, trail
each row and aren't named in `meta.columns` (see `MarkerResult`'s doc
comment in `src/rqtl.rs`). `permutation` is `null` unless `nperm` was
set above 0; when it is, it's
`{"values": [...], "suggestive": ..., "significant": ...}`, the 67th-
and 95th-percentile thresholds GeneNetwork3 itself uses -- computed
here with `numpy.percentile`'s default linear interpolation,
reproduced in `src/rqtl.rs`'s `percentile`.)

## Genotype files

`/compute/rqtl` and `/compute/gemma` take **paths on the server**, so
which panel file is current, in which format, and in which sample order
is a deployment question rather than an API one. That convention (GN2's
per-group `genotype/<Group>.json`, whose `genofile` list is newest-first;
`genotype/bimbam/<file>_geno.txt` for GEMMA; the `.geno` header for the
sample order; `InbredSet.MappingMethodId` for which engine a group
offers) is written up in [`docs/genotypes.md`](docs/genotypes.md),
together with the three mappings measured against this deployment.

## OpenAPI

`GET /openapi.json` serves an OpenAPI 3.1 document covering every route.
It's built with [utoipa](https://docs.rs/utoipa) and
[utoipa-axum](https://docs.rs/utoipa-axum):

- Each handler carries a `#[utoipa::path(...)]` attribute describing its
  method, path, parameters and responses; each response/request struct
  derives `utoipa::ToSchema` (see `VersionInfo`, `db::Species`,
  `db::SpeciesDetail`, `error::ApiError` in `src/lib.rs`/`src/db.rs`/
  `src/error.rs`).
- Routes are registered with `utoipa_axum::routes!(handler)` instead of
  plain `axum::routing::get(handler)`, on an `OpenApiRouter` instead of a
  plain `Router`, so each route's metadata is collected as it's added.
- `ApiDoc` (a `#[derive(utoipa::OpenApi)]` struct) supplies the
  document's `info` (title, description, license); without it the
  generated spec would carry utoipa-axum's own placeholder metadata
  instead of this API's.
- `OpenApiRouter::split_for_parts()` splits the builder back into a
  plain `axum::Router` (returned from `app()`) and the assembled
  `utoipa::openapi::OpenApi` value, which the `/openapi.json` handler
  then just serializes with `Json`.
- `ApiDoc`'s `#[openapi(security(("bearer_token" = [])), ...)]` makes a
  token the default for every operation, so an endpoint that needs one
  doesn't repeat `security(...)` itself -- only the handful that must
  NOT require one (`/`, `/version`, `/get-test-token`) override it with
  an empty `security()` of their own. `error::Unauthorized` (a
  `#[derive(utoipa::IntoResponses)]` type) is the shared 401 response
  every gated endpoint references by name in its own `responses(...)`
  list (`responses(Unauthorized, (status = 200, ...), ...)`) instead of
  writing out `(status = 401, description = "...", body = ApiError)` by
  hand each time.

Because the attributes are declared alongside the handlers they
describe, the spec can't silently drift out of sync with the code the
way a hand-maintained spec file could -- adding or changing a route
without updating its `#[utoipa::path(...)]` is a compile error (the
macro checks the response/schema types exist) or an obviously incomplete
diff, not something to notice weeks later.

The document is assembled once, when `app()` runs at process startup
(not at `cargo build` time, and not freshly on every request): the
`#[utoipa::path]`/`#[derive(ToSchema)]` macros expand to Rust code at
compile time, but the actual merging of all routes' metadata into one
`OpenApi` value happens the first time `app()` executes. Every
`GET /openapi.json` request after that just serializes the same
already-built value.

## Success responses

Every success response is one JSON envelope (`src/response.rs`): the
data, next to facts about the response, remarks on the data, and
suggested follow-up calls. It is the success counterpart of the error
shape below.

```json
{
  "data": {"name": "mouse", "full_name": "Mus musculus", "...": "..."},
  "meta": {
    "state": "complete",
    "took_ms": 3,
    "request_id": "d4f1f9a0-62d7-4a52-9d51-62b0c8bfa5f5"
  },
  "notes": [],
  "next": [
    {"href": "/species/mouse/datasets", "description": "Expression datasets of this species"},
    {"href": "/groups?species=mouse", "description": "Groups of this species"}
  ]
}
```

- `data` is the response proper: exactly what each endpoint below
  documents.
- `meta.state` is `complete`, or `partial` when the response is usable
  but something was left out or unavailable -- `notes` says what. For
  example `/search` is `partial` when the database could not be reached
  (hits come back without `locus` and `abbreviation`) or when a sorted
  search ranked a full 1000-match window and more matches exist;
  `/publications/{pubmed_id}/traits` is `partial` when traits of
  datasets the caller may not view were left out. The HTTP status stays
  200: the state is about completeness, not failure.
- `meta.took_ms` is the server's time for the request, from when it
  started handling it; it excludes the network. `meta.request_id` is the
  request's id, also in the `x-request-id` response header and in any
  error body, so a report can be correlated with logs.
- `x-gn4-user` names the caller on every **authenticated** response: the
  token's `sub` (e.g. `pjotrp`). It is there for the proxy's access log --
  `user=$upstream_http_x_gn4_user` in the `gn4` log format -- so the
  deployment can answer "who is using this, successfully" rather than
  guessing from addresses and user agents. Unauthenticated responses
  (`/version`, `/openapi.json`, a `401`) carry no such header, and the
  caller only ever sees an identity it already presented.
- The same identity goes into this server's own access log: one JSON line per
  request on stderr — `{"event":"request","method","path","status","took_ms",
  "request_id","user"}` — with `user` absent when no token was verified, the
  path stripped of its query string, and the token itself never logged.
  `GN4_ACCESS_LOG=false` turns it off. The `event: error` lines for 5xx go to
  the same stream and carry a `request_id`, so a failure and the request that
  caused it are one `grep` apart.
- `notes` are server-authored remarks on `data`, e.g. `lod = lrs /
  4.604` on responses that carry `lod`. They are never derived from
  database content or request input, so a model reading them is not
  reading anyone's text but this server's.
- `next` lists follow-up calls, each a relative `href` that resolves
  with a 200 for the response it comes from, and what it gives you (the
  test suite follows them, except the two listings of thousands of
  rows -- a group's traits and a dataset's probesets -- whose
  descriptions say they can be very large). `GET /search` suggests the
  next page and the top hit's metadata.
- `actions` lists suggested calls that are **not** plain `GET`s, so they
  cannot be `next` links: each carries `method`, `href`, a
  `description`, and a `body` where the server can build one. It is
  present only when there is something to suggest, so most responses
  keep the four members above. Today the per-strain values endpoints
  (`/groups/{group}/traits/{id}/values`, `/datasets/{dataset}/probesets/
  {probeset}/values`, `/traits/{dataset}/{trait}/values`) use it to hand
  back the values they just returned, ready to map:

  ```json
  "actions": [
    {"method": "POST", "href": "/compute/rqtl",
     "description": "Map 'BXDPublish::10001' with R/qtl (scanone): this body carries the 34 values above as `samples`, in the same order. Replace the `genotype_file` placeholder with a path that exists on the server. …",
     "body": {"genotype_file": "<path on the server, e.g. a genotype file it already holds>",
              "trait_name": "BXDPublish::10001",
              "samples": [{"name": "BXD1", "value": 61.400001525878906}, "…"]}},
    {"method": "POST", "href": "/compute/gemma", "description": "…",
     "body": {"genotype_file": "<…>", "snps_file": "<…>",
              "trait_values": [61.400001525878906, "…"]}}
  ]
  ```

  The file paths stay bracketed placeholders on purpose: they have to
  exist on the server's own filesystem, which the server cannot know on
  the caller's behalf (see the `/compute` sections below).

`/` returns the plain text `Hello, world!` and `/openapi.json` the
OpenAPI document itself; neither is wrapped. In `/openapi.json` every
other success response is an `Envelope_*` schema (a test enforces this).

## Error handling

Every endpoint uses the same `ApiError` JSON shape on error, modeled on
RFC 7807 (Problem Details for HTTP APIs) plus a few extra fields (see
`src/error.rs`). `/species` and `/species/{species}` can actually produce
one today (database errors, or a 404 for an unknown species); `/` and
`/version` document the same 500 response in `/openapi.json` for
consistency, even though neither has a failure path currently (no
database access, no I/O):

```json
{
  "type": "/errors/species_not_found",
  "title": "Species not found",
  "status": 404,
  "code": "species_not_found",
  "detail": "No species matching 'nonexistent'.",
  "operation_id": "species_by_name",
  "spec_url": "/openapi.json",
  "request_id": "20974d91-52c0-41fe-ba2b-4df399a4cc4f",
  "took_ms": 2,
  "next": "/species"
}
```

`next` is omitted (`null`) when there's no obviously related endpoint to
suggest, e.g. a hypothetical error from `/`:

```json
{
  "type": "/errors/internal_error",
  "title": "Internal error",
  "status": 500,
  "code": "internal_error",
  "detail": "example failure for /",
  "operation_id": "hello_world",
  "spec_url": "/openapi.json",
  "request_id": "20974d91-52c0-41fe-ba2b-4df399a4cc4f",
  "took_ms": 1,
  "next": null
}
```

- `type` and `spec_url` are relative paths, not absolute URLs, since this
  service has no fixed public domain; RFC 7807 explicitly allows `type`
  to be a relative URI reference. `spec_url` is always `/openapi.json`,
  on every error from every endpoint.
- `operation_id` is the OpenAPI operation that failed (same name as in
  `/openapi.json`).
- `request_id` is the request's `x-request-id`, generated per request
  (or echoed back if the client already sent one) by
  `SetRequestIdLayer`/`PropagateRequestIdLayer` in `src/lib.rs`, and
  included on the response too, for correlating a failure with logs.
  Every error has one, including the 401 for a missing or invalid token.
- `took_ms` is the server's time for the request up to the error, the
  same measure as a success response's `meta.took_ms` (see
  [Success responses](#success-responses)). RFC 7807 allows extension
  members like this one, `request_id`, `code` and `next` next to its
  standard ones.
- `next` is an optional hint at a related endpoint to try instead; e.g.
  a 404 from `/species/{species}` sets `next` to `/species`, so a
  caller with a wrong species name can discover the valid ones.

## Running with cargo

Requires a Rust toolchain with cargo (edition 2024).

```
cargo run
```

The server listens on `0.0.0.0:3000` by default. Set `GENENETWORK4_PORT`
to use a different port, or `GENENETWORK4_BIND` to bind a specific
address instead of every interface:

```
GENENETWORK4_PORT=8080 cargo run
GENENETWORK4_BIND=127.0.0.1 cargo run
```

### This deployment runs two instances: 3000 (production) and 3001 (development)

The reference deployment keeps a development instance next to the live
one, on the same box, so work can be tried out without disturbing what
the site is serving:

| instance | port | binary | run caches | proxied by nginx |
|---|---|---|---|---|
| **production** | `3000` | `target/release/genenetwork4` | defaults (`/tmp/gn4-gemma`, `/tmp/gn4-rqtl`, …) | yes, at `/gn4/` (see [docs/nginx.md](docs/nginx.md)) |
| **development** | `3001` | `target/debug/genenetwork4` | `/tmp/gn4-dev/{gemma,rqtl,ravanan}` | no |

They differ only in the port, the build profile and those cache
directories; both use the same read-only `SQL_URI`, the same Xapian index
and the same auth stub, so a request behaves the same on either. The
separate caches matter: a GEMMA or R/qtl run on the development instance
must not fill, or be sped up by, production's warm cache.

```sh
P=$(guix shell -D -f guix.scm -- bash -c 'echo $GUIX_ENVIRONMENT')   # needs R + gemma-wrapper on PATH

# production
PATH=$P/bin:$PATH R_LIBS_SITE=$P/site-library/ \
GENENETWORK4_BIND=127.0.0.1 GENENETWORK4_PORT=3000 GENENETWORK4_TEST_TOKEN=false \
AUTH_SERVER_URL=http://127.0.0.1:8081/ \
SQL_URI=mysql://webqtlout:webqtlout@127.0.0.1/db_webqtl \
XAPIAN_DB_PATH=/home/wrk/tmp/ai-root/factory/gn4-index SEARCH_ENGINE=native \
nohup ./target/release/genenetwork4 > /tmp/gn4-3000.log 2>&1 &   # pid in /tmp/gn4.pid

# development: same, plus the port and its own caches
PATH=$P/bin:$PATH R_LIBS_SITE=$P/site-library/ \
GENENETWORK4_BIND=127.0.0.1 GENENETWORK4_PORT=3001 GENENETWORK4_TEST_TOKEN=false \
AUTH_SERVER_URL=http://127.0.0.1:8081/ \
SQL_URI=mysql://webqtlout:webqtlout@127.0.0.1/db_webqtl \
XAPIAN_DB_PATH=/home/wrk/tmp/ai-root/factory/gn4-index SEARCH_ENGINE=native \
GEMMA_CACHE_DIR=/tmp/gn4-dev/gemma RQTL_CACHE_DIR=/tmp/gn4-dev/rqtl \
RAVANAN_CACHE_DIR=/tmp/gn4-dev/ravanan \
nohup ./target/debug/genenetwork4 > /tmp/gn4-dev.log 2>&1 &   # pid in /tmp/gn4-dev.pid
```

Both need the `PATH`/`R_LIBS_SITE` pair (see "Building with cargo, outside
a bare toolchain" below) or R/qtl fails with `there is no package called
'optparse'`, and both need the auth stub on `127.0.0.1:8081` to serve
`/auth/public-jwks` (see "Tests").

Anything the server keeps in memory is per instance: the pool, the JWKS
and authorisation caches, and any in-process compute limit. So the two
instances do not serialise against each other — a mapping on `:3001` runs
alongside one on `:3000` — which is the point of the split, and a good
reason not to treat the development instance as a capacity test of what
production will do.

### Building with cargo, outside a bare toolchain

`cargo build` in the package's own dependency shell is the supported
way to build; a bare `cargo build` against a distro toolchain fails,
and one of the reasons cannot be fixed by installing packages, since it
is a Xapian version conflict (see [Running with Guix](#running-with-guix)
for the pin in `guix.scm`).

```
GUIX_PACKAGE_PATH=<path-to-guix-bioinformatics> \
guix shell -D -f guix.scm -- bash -c '
  export LIBCLANG_PATH=$GUIX_ENVIRONMENT/lib       # autocxx: libclang 13, not 22
  export DUCKDB_LIB_DIR=$GUIX_ENVIRONMENT/lib      # libduckdb-sys: no duckdb.pc
  export DUCKDB_INCLUDE_DIR=$GUIX_ENVIRONMENT/include
  cargo build --release'
```

- `GUIX_PACKAGE_PATH` (or `-L`) is for `(gn packages gemma)`; without
  it `guix.scm` fails to load with "no code for module (gn packages
  gemma)".
- The three exported variables are exactly what `guix.scm`'s
  `set-duckdb-paths` phase exports for the packaged build. Without
  `LIBCLANG_PATH` the build stops in `xapian-rs`'s build script with
  "Unable to find libclang"; without the `DUCKDB_*` pair
  `libduckdb-sys` links a `-lduckdb` with no search path.
- The resulting binary resolves every shared library from the shell's
  store paths on its own (Guix's linker wrapper records them as a
  RUNPATH), so running it needs no `LD_LIBRARY_PATH` -- but it does stop
  working once those store paths are garbage-collected.

### Project memory: why this code looks like this

The repository carries an `ai-memory` store (`ai-memory.sqlite`, gitignored)
that links each commit to the reasoning behind it, plus an optional Xapian
topic index over both. It is how a session that starts with an empty context
recovers what the last one worked out — measured at 16 ms for a text query and
about 100 ms through the topic index:

```sh
ai-memory -C . search 'pve'                     # reasoning, full-text
ai-memory -C . why gemma.compute_compute        # commits that touched a symbol, with its reasoning
ai-memory -C . --index ~/.cache/gn4-topics search 'topic:ravanan-slurm'
```

Operational runbooks live there too — for example the Slurm container from
`gn-machines`, its JWT, the environment `/compute/gemma/ravanan` needs, and
the fastest endpoint test — because the commands are the part a reader cannot
reconstruct from a diff. The store is local and per repository: it survives a
cleared session, not a lost machine, so every conclusion is also in a commit
message and in these docs.

## Development

Every endpoint needs a bearer token (see the note under
[Endpoints](#endpoints)), which normally means getting one from
gn-auth yourself. For local development, genenetwork4 can hand you one
instead: `POST /get-test-token` calls gn-auth's own password grant
(`POST /auth/token`) for gn-auth's well-known development user and
returns the result. It needs no setting and is not tied to a mode.

```
cargo run
```

Get a token, then use it against a real endpoint:

```
curl -X POST http://127.0.0.1:3000/get-test-token
```

```json
{
  "data": {"access_token": "eyJhbGc...", "token_type": "Bearer", "scope": "profile group resource"},
  "meta": {"state": "complete", "took_ms": 41, "request_id": "..."},
  "notes": [],
  "next": []
}
```

```
TOKEN=$(curl -s -X POST http://127.0.0.1:3000/get-test-token | python3 -c 'import json,sys; print(json.load(sys.stdin)["data"]["access_token"])')
curl -H "Authorization: Bearer $TOKEN" http://127.0.0.1:3000/species
```

```json
{
  "data": [
    {"name": "mouse", "menu_name": "Mouse (Mus musculus, mm10)", "family": "Vertebrates"},
    ...
  ],
  "meta": {"state": "complete", "took_ms": 2, "request_id": "..."},
  "notes": [],
  "next": [
    {"href": "/groups", "description": "Groups (populations); add ?species=<name> to list one species' groups"}
  ]
}
```

The endpoint is registered by default and listed in `/openapi.json`
(`GENENETWORK4_TEST_TOKEN=false` switches it off, see below), and decides
per request whether to answer: only a caller on this machine gets a
token. That means the TCP connection's peer address is in
`127.0.x.x` (or is `::1`), and the request carries no `X-Forwarded-For`,
`Forwarded` or `X-Real-IP` header. Anything else, including a request
whose peer address is unknown, is a `403`. The peer address comes from
the connection itself, never from a header a client can set; the proxy
headers are used only to refuse a request that claims to have been
forwarded. This is why the check also refuses forwarded requests: a
reverse proxy running on the same host connects from `127.0.0.1` for
every client, so without it the whole internet would look local. A proxy
that forwards without setting any of those headers would defeat the
check, so don't put one in front of a server whose gn-auth has the dev
user (on a production gn-auth the dev user normally doesn't exist, and
the request just fails with a `502`).

### Behind nginx or in a container

[`docs/nginx.md`](docs/nginx.md) is the full deployment recipe: serving
genenetwork4 under a path prefix such as `https://host/gn4/` from a
loopback port, the long `proxy_read_timeout` `/compute/*` needs, and how
the root-relative links in responses are kept under the prefix. What
follows here is only about `POST /get-test-token`.

The check looks at who connected to genenetwork4, so what it sees
depends on the deployment (the Docker behavior below is how Docker
networking works; it was not tested here):

- **Container on a bridge network, port published.** The server sees
  the bridge gateway (something like `172.17.0.1`) as the peer, not
  `127.0.0.1`. Every request through the published port is a `403`,
  whether it comes via nginx or from the host itself. It still works
  from inside the container, e.g. `docker exec <container> curl -X POST
  http://127.0.0.1:3000/get-test-token`.
- **Container sharing the host's network, nginx on the same host.**
  nginx connects from `127.0.0.1`, so the peer check passes for every
  client. The server then relies on refusing requests that carry
  `X-Forwarded-For`, `Forwarded` or `X-Real-IP`, and a minimal
  `proxy_pass` sets none of them. Without one of the measures below,
  anyone who can reach the site could get a token.

For a deployed instance, don't rely on the check alone. Any one of
these is enough, and a production deployment should use the third:

1. **Block the path in nginx**, independent of genenetwork4:

   ```nginx
   location = /get-test-token { return 404; }
   ```

2. **Make nginx say it is forwarding**, so genenetwork4 refuses proxied
   requests (any one of the three headers is enough):

   ```nginx
   proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
   ```

3. **Switch the endpoint off in genenetwork4 itself** with
   `GENENETWORK4_TEST_TOKEN=false` (an environment variable, or a
   `GENENETWORK4_TEST_TOKEN=false` line in `GN4_CONF`). The route is then
   not registered at all: `POST /get-test-token` is a `404` and the
   endpoint is absent from `/openapi.json`. The setting fails closed:
   only `true`, `1`, `yes` or `on` (any case) keep the endpoint on, and
   anything else -- `false`, an empty value, or a typo such as `flase`
   -- turns it off. It is on by default, so a local setup needs nothing.

Also relevant: the token is for gn-auth's dev user
(`GN_AUTH_DEV_USERNAME`). On a production gn-auth that user normally
doesn't exist, so a request that does get through fails with a `502`
instead of returning a token. That depends on gn-auth's data, not on
anything in genenetwork4.

The credentials used (gn-auth's dev client id/secret and dev user
email/password, from `flask init-dev-clients`/`gn_auth/wsgi.py`) are
publicly known and the same for every gn-auth checkout by default --
override `GN_AUTH_DEV_CLIENT_ID`, `GN_AUTH_DEV_CLIENT_SECRET`,
`GN_AUTH_DEV_USERNAME` and `GN_AUTH_DEV_PASSWORD` if your gn-auth
instance's dev fixtures differ. Needs a gn-auth reachable at
`AUTH_SERVER_URL` with `AUTHLIB_INSECURE_TRANSPORT` set (gn-auth itself
refuses OAuth2 grants over plain HTTP otherwise -- a gn-auth-side dev
setting, not something genenetwork4 needs to know about).

The token has no `exp` claim, and neither genenetwork4 nor gn-auth
expires it by time or for lack of use. It stops verifying only when the
key that signed it leaves gn-auth's published key set: gn-auth signs with
a key for 7 days (`JWKS_ROTATION_AGE_DAYS`) and deletes it 14 days after
its creation (`JWKS_DELETION_AGE_DAYS`), so a token lasts roughly 7 to
14 days, depending on where in the key's life it was issued.
genenetwork4 caches the key set for an hour. When requests start
failing with `401 invalid_token`, fetch a new token. (`exp` is optional
in genenetwork4's own validation, which is why a token without one like
this lives as long as its signing key does -- see "Tokens are
stateless" below.)

The test suite uses this same endpoint automatically -- see
[Tests](#tests) below.

Configuration is layered: built-in defaults, then `GN4_CONF` and
`GN4_SECRETS` files of `KEY=VALUE` lines, then environment variables
(which always win). Keys: `SQL_URI` (MariaDB), `SQL_CHECK_HEALTH` (see
below), `GN3_URL` (GeneNetwork3
base URL for proxy endpoints), `GN3_TIMEOUT_SECS` (upstream call
timeout, default 30), `GENENETWORK4_PORT`, `GENENETWORK4_BIND`, and
`GENENETWORK4_TEST_TOKEN` (`false` turns `POST /get-test-token` off; see
"Behind nginx or in a container").

### Tokens are stateless

Nothing about a bearer token is stored server side: there is no session
table, no list of issued tokens, no `jti` denylist. Verification is
signature + `iss` + `exp`/`nbf`, and that is all, so **restarting
genenetwork4 loses no tokens** -- a token minted before a redeploy keeps
working after it, as does one minted by a stub that has since been
restarted (its key is fixed). What the process does keep is two caches,
and restarting only empties them:

- the signing keys (`JWKS_TTL`, an hour) -- how long a fetched copy of
  the auth server's key set is reused before being re-fetched;
- authorisation answers (`AUTHORISATION_TTL`, a minute), per
  `(subject, resource)`.

Neither is a lifetime. The token's own `exp` is the only thing that ends
its life: it is validated on every request (`validation.validate_exp` in
`src/auth.rs`, with jsonwebtoken's default 60-second leeway), it is never
extended, and there is no refresh token. The hour matters in one case
only: if the auth server is unreachable, a cached key set keeps
verification working for up to `JWKS_TTL`, and past that every request
is `401 invalid_token` with a `fetching .../auth/public-jwks` detail --
the tokens are still valid, the key lookup is not. Run the auth server
under the same supervision as genenetwork4 rather than from a shell.

Consequences worth knowing:

- **Restarting is not revocation.** Since no state exists, restarting
  genenetwork4 or the auth server leaves every outstanding token valid.
  To withdraw one, wait for its `exp` (hence short lifetimes for tokens
  you may need to pull, e.g. `scripts/gn-auth-stub.py --expires-in`). To
  withdraw all of them at once, change the signing key or
  `AUTH_SERVER_URL` -- `iss` is compared against that string verbatim,
  so a changed host, scheme or trailing slash invalidates everything
  with `InvalidIssuer`.
- **`exp` is optional.** `required_spec_claims.clear()` means a token
  without `exp` is accepted and never expires. gn-auth and
  `scripts/gn-auth-stub.py` always set one; a hand-minted token (the
  test key is in this repository) that omits it is a permanent
  credential.

Stub-minted tokens behave the same way (see [Tests](#tests)): `exp` is
whatever `--expires-in` said at mint time, and restarting the stub does
not invalidate them -- it always signs with the committed test key.

### Request body size

The API accepts bodies up to **64 MiB** (`BODY_LIMIT_BYTES` in
`src/lib.rs`). axum's own default is 2 MiB, which a legitimate request
exceeds: `/plot/{rqtl,gemma}` is *posted a whole mapping response* (a
real GEMMA run on the BXD WGS panel is ~1.9 MB; a genome-sized one is
tens of MB), and `/compute/*` carries one value per individual plus up to
eight covariates. Over the limit, axum answers `413 Failed to buffer the
request body: length limit exceeded` before any handler runs.

Two consequences worth knowing:

- a proxy in front must allow at least as much
  (`client_max_body_size 64m` in `etc/nginx-vesuvio.pu.ac.ke.conf`), or
  it refuses first with a 413 the API never sees;
- a request does not need to be that large. For covariates, prefer
  `{"case_attribute": "Sex"}` -- the server reads the attribute and
  aligns it, so nothing is posted -- over sending a value for every
  individual in the panel. Only the samples the trait has are used, so
  posting values for the rest is pure payload.

### Database connections

`SQL_URI` is used through one shared `mysql::Pool`. Two of the driver's
per-checkout round trips are configurable:

- `SQL_CHECK_HEALTH` (default `true`) pings a connection taken from the
  pool and replaces it when the ping fails. Leave it on unless the
  database is far away: turning it off saves a round trip, but then a
  connection the server closed while idle surfaces as `500
  database_error` on whichever request happens to draw it -- measured
  here as ten consecutive failures, one per dead pooled socket, until the
  process was restarted.
- `reset_connection` is off unconditionally: a `COM_RESET_CONNECTION` per
  checkout, with nothing to reset (the queries are read-only).

Even with the ping on, a socket can die between the ping and the query;
when a database error names a connection loss (`Broken pipe`, `Connection
reset`, `server has gone away`, ...), the pool is dropped so the next
request builds a fresh one (`db::is_connection_loss` / `db::forget_pool`,
called from `ApiError::from_db`, the one place database errors pass
through). `tests/db_recovery.rs` kills every pooled session and asserts
the following requests still succeed.

## Running with Guix

```
guix build -f guix.scm
```

builds the package (running both test suites, see below) and prints the
resulting store path; the binary is at `<store path>/bin/genenetwork4`.

The `xapian` input is pinned to 1.4.29 (`xapian-1.4` in `guix.scm`): the
`xapian-rs` git revision `Cargo.toml` pins was written against Xapian
1.4, generating bindings for `Xapian::InMemory::open`, and Xapian 2.x
removed it (`xapian/dbfactory.h` no longer declares the namespace; 2.x
spells it as the `Xapian::DB_BACKEND_INMEMORY` flag). 2.x also needs
C++17, where autocxx hands the parser a fixed `-std=c++14` and 1.4's
headers are happy with it. Guix's own `xapian` is 2.1.0 now, so without
the pin the `xapian-rs` build script stops with "the 'generate' or
'generate_pod' directive for 'Xapian::InMemory::open' did not result in
any code being generated" before compiling anything of this project's.
The pin is a plain `url-fetch`/`sha256` for `xapian-core-1.4.29.tar.xz`
with 1.4's own `make check-inmemory check-remoteprog check-multi
check-glass check-chert` test targets; drop it once `xapian-rs` moves to
Xapian 2. The same pin is what `guix shell -D` and a manual `cargo
build` see (see "Building with cargo, outside a bare toolchain" above).

```
guix shell -D -f guix.scm
```

drops you into a shell with all build dependencies available (cargo,
rust, python, git), for running `cargo build`/`cargo test`/`cargo run`
directly. This shell is not isolated from the network or from your
normal environment, which `cargo build`/`cargo run` need to reach
crates.io.

A fully isolated container (`-C`) normally needs `-N` to still reach the
network from inside it, but on at least one guix build
(1.5.0-6.64d4de2) combining `-D` with `-N` causes `-D` to silently stop
providing the package's build dependencies (the shell then has only
`genenetwork4` itself, not cargo/rust/etc.). If `guix shell -C -D -N -f
guix.scm` fails with "cargo: command not found", drop `-C`/`-N` and use
the plain `guix shell -D -f guix.scm` above instead.

## Tests

Two test suites exercise the same endpoints:

- `tests/*.rs` - Rust integration tests that call the router in-process
  (no socket), via `tower::ServiceExt::oneshot`. Run with:

  ```
  cargo test
  ```

  Both suites check the envelope on every success response: the Rust
  helper `tests/common::data(&body)` and the Python helper
  `server_test_case.envelope_data(body)` verify its shape and return
  `data`. `tests/envelope.rs` (and `test/python/test_envelope.py`)
  follow every `next` link the endpoints suggest;
  `tests/envelope_partial.rs` checks the `partial` state without a
  database; `tests/openapi.rs` checks that every JSON success response
  is documented as an envelope.

- `test/python/test_*.py` - Python tests that spawn the actual compiled
  binary and talk to it over real HTTP, since Python can only reach the
  server that way. Run with:

  ```
  cargo build --release
  GENENETWORK4_BINARY=target/release/genenetwork4 python3 -m unittest discover -v test/python
  ```

Either suite can fetch its own token from a live gn-auth automatically
(see "Development" above) -- set `AUTH_SERVER_URL` before running
either command above, and nothing further needs setting up by hand. A
run without a reachable gn-auth still works, skipping (not failing)
every test that needed a token it couldn't get -- set
`GN4_TEST_AUTH_TOKEN` instead if you'd rather provide one manually.

Both suites also run automatically as part of `guix build -f guix.scm`
(the Rust tests in the `check` phase, the Python tests in a
`python-rest-test` phase after install).

Every test under `tests/*.rs`/`test/python/test_*.py` for an endpoint
that touches the database (which is most of them -- `hello_world`,
`version`, `openapi` and `get_test_token`'s own disabled/wrong-config
cases are the exceptions) checks for a live MariaDB on `127.0.0.1:3306`
first (see `SQL_URI` above) and skips with a message rather than
failing when none is reachable, since the Guix build sandbox has no
network/database access. Likewise, any test for an endpoint that
requires a token (every endpoint except those four) checks for one
first and skips if none is available -- see "Development mode" above
for how the test suite gets one automatically.

The Rust tests in `tests/*.rs` that hit the database issue their
requests through `tests/common::get`, which times the request and
prints a warning to stderr if it takes over a second -- e.g.
`WARNING: GET /groups/BXD/traits/10001 took 7.1s (> 1s)`. Warnings only
show with `cargo test -- --nocapture` (or on a failing test), since
cargo captures stdout/stderr for passing tests otherwise.

### Against a real database and index

Those skips are the sandbox's, not the code's. Supply the three things
the tests look for -- a live MariaDB, a Xapian index, a token -- and the
commands above run the whole suite rather than a small subset. In a
development environment with all three, `cargo test` reports 225 tests
passed / 0 failed (the handful of skips left are listed below) and the
Python suite 94 tests, OK, no skips.

1. **Database**: whatever `SQL_URI` points at, by default
   `mysql://webqtlout:webqtlout@127.0.0.1/db_webqtl`.
   `tests/common::db_available()` really connects, so wrong credentials
   skip exactly like an absent server does.
2. **Index**: `XAPIAN_DB_PATH` for a dev index, plus `GN4_FULL_INDEX`
   for a full/production one. Only a full index is treated as a count
   oracle: `tests/search_native_db.rs` then requires the engine's
   `dataset:bxdpublish` match set to equal the database's `PublishXRef`
   ids for BXD exactly, id for id -- the invariant a dev snapshot
   famously fails (it is missing ~1.7k traits and holds ~350 ids the
   database no longer has). A dev index only has to overlap, and its
   counts are then printed as reference data, never asserted. When both
   are set the full one wins.
3. **Token**: either a real gn-auth at `AUTH_SERVER_URL` (see
   "Development"), or `scripts/gn-auth-stub.py`, which stands in for
   one: it serves the JWKS matching the committed
   `test/data/test_rsa_private.pem` (the same fixture pair
   `search_native_db` spins up in-process), answers `/auth/token` so
   `POST /get-test-token` works, and answers both authorisation
   endpoints with "no privileges" -- the stranger case the non-public
   endpoints' 403 tests assert. It is test scaffolding, not a gn-auth
   replacement: a committed test key, and no knowledge of users or
   groups. `--grant-resource-id ID` gives its token
   `group:resource:view-resource` on one resource id, for the owner-side
   tests, which want `GN4_TEST_RESOURCE_ID=ID` to match.

   Left ungranted, a stub token is a **public-data-only token**, which is
   useful beyond the tests (handing a caller or an agent something to
   browse the API with). Only-public is not a property of the token:
   genenetwork4 decides from the data's own
   `public`/`confidentiality`/`create_time` columns, and asks this
   authorisation lookup about anything that is not public (`src/auth.rs`'s
   `check`). No privileges therefore means public datasets read normally
   and every non-public one is a 403 -- the token's `scope` is never
   consulted for data access. Dataset *metadata*
   (`/species/{species}/datasets`, `/groups/{group}/datasets`) stays
   visible either way: those endpoints are documented as returning names
   and flags, not trait data, and are not gated on fine-grained
   authorisation.

   `--mint-only` mints for one subject and exits without listening: how a
   token is issued for one more user without disturbing the running stub
   (see docs/www/operator-checklist.md, "Token issuance").

   `--expires-in` sets the token's lifetime (`30d`, `24h`, `60m`, or
   plain seconds; default 24h). A month-long one is fine because the
   stub's key is fixed and never rotates: nothing invalidates the token
   but its own `exp`, and it stays verifiable for as long as anything
   serving `test/data/test_jwks.json` is up (genenetwork4 caches the key
   set for an hour, so a brief gap is survivable):

   ```
   scripts/gn-auth-stub.py --port 8099 --expires-in 30d --token-file /tmp/public-only.token &
   AUTH_SERVER_URL=http://127.0.0.1:8099/ GENENETWORK4_PORT=3999 cargo run
   ```

   Note the port: `iss` has to match the server's `AUTH_SERVER_URL`, so a
   token minted for `:8099` is rejected by a server left on the default
   `:8081`, and vice versa.

```
scripts/gn-auth-stub.py --grant-resource-id 424242 --token-file /tmp/gn4-token.txt &
export GN4_TEST_AUTH_TOKEN=$(cat /tmp/gn4-token.txt)
export GN4_TEST_RESOURCE_ID=424242
export GN4_FULL_INDEX=/path/to/full-index XAPIAN_DB_PATH=/path/to/full-index

GUIX_PACKAGE_PATH=<path-to-guix-bioinformatics> \
guix shell -D -f guix.scm -- bash -c '
  export LIBCLANG_PATH=$GUIX_ENVIRONMENT/lib \
         DUCKDB_LIB_DIR=$GUIX_ENVIRONMENT/lib \
         DUCKDB_INCLUDE_DIR=$GUIX_ENVIRONMENT/include
  cargo test --no-fail-fast -- --nocapture'

cargo build --release
GENENETWORK4_BINARY=target/release/genenetwork4 \
  python3 -m unittest discover -v test/python
```

The stub listens on `127.0.0.1:8081`, which is the default
`AUTH_SERVER_URL` -- including its trailing slash, since that string is
compared verbatim against the token's `iss`. `--port`/`--issuer` cover
the case where `AUTH_SERVER_URL` says something else.

What still skips with all three in place, and why:

- one `config` unit test asserts what an *unset* `XAPIAN_DB_PATH`
  means, so setting the variable at all skips it;
- `src/indexer/cache.rs`'s RDF-cache test needs an authoritative RDF
  cache directory, which this does not provide;
- `compute_gemma_ravanan_*` needs ravanan, gemma-wrapper and ccwl on
  `PATH`;
- `tests/search_native_dev_corpus.rs` records a dev corpus, so it needs
  `GN4_DEV_CORPUS_DIR` as well as an index.