[devin-ai-integration]

What and why

The CircleCI Go race tests (build 12113) and Go coverage shards (build 12104) jobs on develop died on go test's 30-minute per-package budget at d8b60ed, the merge of #1319 (usages subset the base library usage of their kind), and had passed on b60a18c just before. This PR measures where the time went and removes the remaining cost of that change on the hot paths.

Root cause of (a). After #1319 every ImplicitGenerals / DirectSupertypes derivation calls implicitUsageBaseFeature, implicitKerMLFeatureBase, implicitBaseUsage and declaredChainReachesKindBase, each of which re-walks the declared generalization chain (declaredGeneralizationReaches) and re-resolves the library name (LookupQualified), five to eight times per symbol. Counting over tests/stressmodel on develop: 178 300 computeImplicitGenerals calls (2.66 M memo hits), none provisional, so the memo itself works — each miss is simply expensive. The fix memoizes the kind base once per symbol, journaled and guarded exactly like implicitBases and ImplicitGenerals; semantics and outputs are unchanged.

Overlap with open PRs. None memoizes this. #1349 (conformsByName, filter.go), #1232 (API JSON, touches implicit.go for derived properties only), #1327 (crossing.go), #1314 (gather.go/model.go recording) and the runtime PRs #1340/#1308/#1310/#1311/#1312 do not touch the kind-base lookup.

Step 1 — where the time went (last 26 develop merges)

Three causes, separated:

• (a) feat(semantics): usages subset the base library usage of their kind #1319 semantics change — the only step change in the series. CircleCI race job: internal/exec/runtime exploration run 1168–1415 s on the 19 merges before → 1803 s = timeout at d8b60ed; tests/stressmodel 157–190 s → 513 s (3.0×); tests/model 1478–1877 s → 2189 s. Coverage job: cmd/sysml 1353–1773 s → timeout; internal/exec/runtime 1018–1299 s → 1671 s; tests/model 853–959 s → 1373 s. Packages with no semantics-heavy tests (tests/migrate, tests/corpus, internal/exec/smt) did not move at that commit. The six develop merges after it (fix(semantics): derive a declaration's implied generals once per symbol #1350, perf(runtime): compute an expression's footprint once per model #1347, perf(resolve): enumerate imported members without walking their routes #1348, perf(runtime): compute a feature's effective types once per model #1351, perf(repl): count a diagnostic's UTF-16 column without encoding the line #1352, perf(semantic): compute a feature's redefinition closure once per model #1353 — CircleCI builds 12114–12175, run on 2026-10-21) show what the already-merged perf PRs recovered: internal/exec/runtime race 1294–1426 s and cmd/sysml coverage 1511–1571 s are back in their pre-feat(semantics): usages subset the base library usage of their kind #1319 bands from fix(semantics): derive a declaration's implied generals once per symbol #1350 on (a662c2b race: 1359 s), while the semantics-heavy packages are not — tests/stressmodel race 182–213 s against 157–190 s before (+10–28 %), coverage 113–122 s against 101 s (+12–21 %), tests/model race 1582–1853 s at the top of its band. That residue is what this PR removes.
• (b) suite growth — gradual, not a step: tests/corpus race 633 s (feat(check): connection and flow definition ends must be usages #1301) → 1125 s (fix(lower): project a named terminate node into the transition footprint #1326) as corpus gates were added; tests/migrate 1932–2383 s on a 55 m budget throughout; cmd/sysml under -coverpkg=./... 1353–1773 s, i.e. under 2 minutes of headroom on a 30 m budget before feat(semantics): usages subset the base library usage of their kind #1319, so any regression tips it over.
• (c) CI configuration — unchanged over the window: both jobs resource_class: large, parallelism: 5, same shard list, same cache. GitHub pr.yml shards are too noisy to attribute (the runtime shard takes 21–41 min on unchanged code), which is why the CircleCI columns are the evidence. The develop pipelines after d8b60ed sat unrun until 2026-10-21 13:51 UTC, which is why their rows carry that date.

CircleCI Go race tests (resource_class large, parallelism 5): per-node wall minutes and per-package go test seconds

develop commit	PR	CircleCI build	job	runtime: node min	runtime: exec/runtime explore	runtime: exec/runtime rest	runtime: tests/wasm	runtime: tools conformance	model: node min	model: tests/corpus	model: exec/smt	model: workspace/model	export: node min	export: tests/export	export: tests/model	export: check/passes	rest: node min	rest: tests/stressmodel	rest: tests/codegen	rest: tests/diagnostics	migrate: node min	migrate: tests/migrate
25909c6	#1300	11874	success	45	1315	–	714	178	17	763	899	–	30	1015	1768	837	32	157	820	237	36	2108
94ea870	#1301	11897	success	43	1217	–	685	214	14	633	731	–	30	1026	1756	845	32	162	854	233	35	2059
12156af	#1303	11898	success	43	1240	–	691	191	16	751	883	–	28	920	1627	748	34	178	971	283	34	2000
e3ec3c2	#1304	11921	success	44	1254	–	735	193	17	788	960	–	29	981	1686	813	34	168	911	261	34	1981
43a831e	#1306	11925	failed	43	1240	–	703	190	16	738	790	–	30	949	1739	787	35	190	992	289	35	2056
4317c4a	#1316	11943	failed	46	1339	–	725	207	19	895	984	–	29	907	1676	759	35	166	907	262	37	2204
1195153	#1305	11957	failed	46	1378	–	747	213	16	714	884	–	32	1051	1872	887	36	184	970	275	38	2209
9c04259	#1302	11960	success	45	1297	–	753	202	19	911	1010	–	32	1048	1877	826	36	170	918	271	35	2078
47a75de	#1315	11981	success	45	1247	–	744	201	21	1017	1037	–	29	983	1713	788	34	179	944	291	34	2015
2bb7df3	#1317	11985	success	47	1335	–	795	197	16	797	725	–	29	966	1699	778	33	163	877	240	36	2142
c5a59e1	#1318	11997	success	48	1415	–	812	198	18	879	919	–	26	912	1496	730	33	175	890	247	36	2103
f9f6d95	#1323	12007	success	45	1297	–	689	190	20	977	1031	–	32	1043	1854	823	31	181	897	282	36	2094
2de4131	#1325	12021	success	44	1250	–	710	183	18	900	916	–	29	1019	1669	823	32	171	882	244	38	2197
a270e66	#1313	12033	success	43	1168	–	754	189	19	923	1019	–	29	1048	1711	860	34	173	909	240	37	2149
4667955	#1326	12053	success	43	1178	–	748	189	22	1125	1138	–	30	1067	1735	884	32	176	922	270	33	1932
50885da	#1329	12065	success	47	1351	–	763	204	22	1093	1164	–	28	1008	1648	822	35	179	939	266	41	2383
fce39e1	#1334	12066	success	47	1368	–	746	203	18	861	891	–	29	995	1685	806	34	183	912	274	35	2055
f53120a	#1335	12078	success	43	1266	–	694	180	19	936	951	–	30	991	1752	778	36	186	1008	307	38	2250
b60a18c	#1330	12092	success	44	1322	–	666	163	19	897	953	–	25	899	1478	720	32	166	877	265	37	2192
d8b60ed	#1319	12113	failed	37	1803† timeout	–	–	–	19	855	835	–	37	998	2189	903	35	513	941	263	39	2292
5b0c985	#1350	12116	success	46	1326	–	769	198	17	798	866	–	31	984	1786	852	36	194	941	281	35	2062
2ce843b	#1347	12128	success	47	1426	–	738	178	19	940	990	–	32	1025	1852	881	36	213	1007	282	36	2125
1e51a00	#1348	12138	success	44	1294	–	700	183	17	838	880	–	27	945	1582	814	33	200	916	249	34	2015
c2a6582	#1351	12153	success	45	1357	–	724	182	20	978	1025	–	30	1032	1764	890	35	211	927	280	37	2211
e5bdb8f	#1352	12164	success	44	1341	–	679	179	20	995	965	–	30	1058	1778	908	34	200	917	254	35	2095
a662c2b	#1353	12174	success	45	1359	–	694	177	18	880	910	–	29	1003	1702	874	34	181	875	243	36	2108

CircleCI Go coverage shards (large, parallelism 5, -coverpkg=./...): per-node wall minutes and per-package seconds

develop commit	PR	CircleCI build	job	runtime: node min	runtime: exec/runtime	runtime: tests/wasm	model: node min	model: tests/corpus	model: exec/smt	model: workspace/model	export: node min	export: tests/export	export: tests/model	export: check/passes	rest: node min	rest: cmd/sysml	rest: tests/codegen	rest: tests/stressmodel	rest: frontend/grpc	migrate: node min	migrate: tests/migrate
25909c6	#1300	11875	success	23	1102	702	14	666	814	461	15	461	870	440	28	1553	704	95	435	21	1199
94ea870	#1301	11886	success	22	1066	668	13	684	728	479	15	468	860	448	24	1353	632	92	376	20	1148
12156af	#1303	11899	success	24	1155	752	14	722	823	522	15	482	871	456	27	1489	710	105	391	21	1217
e3ec3c2	#1304	11910	success	24	1173	736	14	724	816	520	16	493	900	478	28	1548	733	106	410	20	1171
43a831e	#1306	11922	failed	24	1152	757	13	634	727	445	16	504	911	486	25	1428	693	98	375	21	1200
4317c4a	#1316	11935	failed	24	1160	738	17	759	954	565	15	469	870	450	28	1564	733	97	407	21	1201
1195153	#1305	11947	failed	23	1135	715	13	681	766	516	16	505	921	491	27	1527	727	100	395	21	1203
9c04259	#1302	11959	success	26	1299	836	15	887	835	661	16	515	917	479	27	1517	713	97	429	22	1268
47a75de	#1315	11970	success	23	1140	719	18	897	1031	667	16	517	910	491	29	1600	733	99	441	21	1217
2bb7df3	#1317	11983	success	24	1172	743	16	912	870	696	15	515	896	484	26	1474	701	102	394	22	1281
c5a59e1	#1318	11995	success	23	1147	734	15	860	901	641	16	524	909	495	26	1489	715	100	411	22	1262
f9f6d95	#1323	12008	success	26	1259	809	17	955	995	710	15	513	895	480	27	1500	713	102	414	20	1197
2de4131	#1325	12020	success	23	1139	761	15	879	895	640	16	522	911	494	26	1465	690	95	398	19	1133
a270e66	#1313	12030	success	24	1178	784	16	924	906	710	16	528	918	495	26	1495	719	100	412	21	1234
4667955	#1326	12043	success	25	1209	783	13	771	640	534	15	493	853	461	28	1569	733	96	417	21	1185
50885da	#1329	12056	success	24	1185	786	15	864	840	623	15	510	891	477	32	1773	857	120	457	21	1224
fce39e1	#1334	12067	success	20	1018	619	18	1011	1019	819	16	517	898	488	25	1398	674	94	372	20	1195
f53120a	#1335	12081	success	23	1139	723	15	888	894	655	16	510	902	474	28	1579	731	103	434	22	1270
b60a18c	#1330	12093	success	24	1191	779	16	925	889	692	17	564	959	531	27	1524	708	100	416	21	1220
d8b60ed	#1319	12104	failed	32	1671	687	15	866	753	738	24	566	1373	604	31	1800† timeout	946	498	447	20	1190
5b0c985	#1350	12114	success	25	1216	729	15	866	883	621	16	500	920	507	27	1523	713	116	406	21	1240
2ce843b	#1347	12129	success	23	1154	670	15	868	796	640	16	510	937	518	27	1535	714	116	418	21	1206
1e51a00	#1348	12141	success	24	1219	722	17	912	964	665	15	464	860	468	27	1511	709	117	416	20	1158
c2a6582	#1351	12151	success	24	1188	727	16	862	912	632	15	445	860	452	27	1535	715	112	416	19	1116
e5bdb8f	#1352	12165	success	23	1110	669	16	849	902	620	16	517	948	525	27	1550	739	118	415	21	1231
a662c2b	#1353	12175	success	23	1144	693	15	878	876	614	17	537	958	544	28	1571	740	121	428	21	1224

GitHub pr.yml full-tier shards on each merged PR's head: Run Go tests step minutes

develop commit	PR	PR head	pr.yml run	runtime (min)	runtime-corpus (min)	model (min)	export (min)	rest (min)	migrate (min)
25909c6	#1300	815012a	success	31	33	18	30	39	15
94ea870	#1301	8205685	success	32	23	18	31	26	22
12156af	#1303	3c0adf6	success	21	33	16	26	40	30
e3ec3c2	#1304	a7438e3	success	23	22	18	18	34	15
43a831e	#1306	c1a8757	success	31	38	14	30	37	21
4317c4a	#1316	4a72836	success	–	–	–	–	–	–
1195153	#1305	b021628	success	31	36	12	19	36	16
9c04259	#1302	3265d5d	success	25	21	13	31	41	27
47a75de	#1315	c4c6d9d	success	33	37	20	29	29	27
2bb7df3	#1317	1676bc1	success	–	–	–	–	–	–
c5a59e1	#1318	7bdb1f1	success	–	–	–	–	–	–
f9f6d95	#1323	04d73e5	success	33	34	20	27	28	30
2de4131	#1325	e5f3a4c	success	21	36	15	22	31	30
a270e66	#1313	5e16a37	success	32	34	20	17	41	22
4667955	#1326	5d47474	success	31	37	16	31	42	30
50885da	#1329	f7841c9	success	32	35	21	29	42	30
fce39e1	#1334	cac9ca3	success	21	22	16	22	27	30
f53120a	#1335	fd355a9	success	–	–	–	–	–	–
b60a18c	#1330	dd25007	success	–	–	–	–	–	–
d8b60ed	#1319	3156524	success	40	36	21	39	34	30
5b0c985	#1350	1ceae96	success	32	33	16	19	25	30
2ce843b	#1347	2cff947	success	41	35	22	39	28	27
1e51a00	#1348	22a1a46	success	24	36	17	22	33	30
c2a6582	#1351	c291109	success	36	35	21	38	45	31
e5bdb8f	#1352	106814f	success	31	34	18	39	44	28
a662c2b	#1353	606dac8	success	24	37	21	38	35	29

Step 2 — before/after, same machine (8 vCPU, go1.25.11), go test -count=1

target	b60a18c (before #1319)	a662c2b (develop)	this PR
cmd/sysml -run TestExploreTablesTheSpacecraftRaceWithinItsBudget (package ok seconds)	135.5 s	145.7 s (+7.5 %)	124.2 s (−8.3 %)
internal/exec/runtime -run TestExploreWithIsExploreOverTheConformanceCorpus (wall)	40.4 s	41.8 s (+3.5 %)	36.6 s (−9.4 %)
tests/stressmodel (package ok seconds)	15.3 s	19.3 s (+26 %)	16.3 s (+6.5 %)

go tool pprof -top -cum -diff_base develop vs b60a18c on tests/stressmodel: ImplicitGenerals +4.21 s, computeImplicitGenerals +4.04 s, implicitUsageBaseFeature +2.24 s, DirectSupertypes +1.86 s, candidateBaseFeatures +1.64 s, declaredGeneralizationReaches +1.50 s, implicitBaseFeatures +1.34 s, declaredChainReachesKindBase +0.62 s; implicitBases −1.30 s (already memoized by #1350). On the spacecraft exploration and the conformance-corpus exploration the develop-vs-before diff is GC noise plus ≤0.5 s in TimeEventConforms; the perf PRs already merged recovered those.

Step 3 — CI configuration

With (a) removed, one package was still living on the budget's edge before #1319 for reasons (b)/(c): cmd/sysml under -coverpkg=./... -covermode=atomic took 1353–1773 s of the 1800 s budget on every one of the 19 green coverage runs before #1319 (1773 s at #1329, 27 s of headroom) and still 1511–1571 s on the six after it with the perf PRs in, while sharing the rest node's 4 vCPUs with ~2200 s of other packages (tests/codegen 709 s, internal/frontend/grpc 416 s, internal/frontend/repl 285 s, tests/diagnostics 174 s, internal/workspace/libs 141 s, cmd/sysml-grpc 127 s, …). The rest shard is also the longest GitHub pr.yml shard (25–45 min). So cmd/sysml becomes a shard of its own, cli:

• scripts/race-shard.sh: cli → cmd/sysml, which rest no longer contains.
• .circleci/config.yml: both jobs parallelism: 6, shard list runtime model export rest migrate cli; the node-count guard stays.
• Makefile: COVERAGE_SHARDS gains cli; test-shard, test-quick and coverage-shard accept it.
• .github/workflows/pr.yml: full tier gains the cli shard, the quick tier runs it with rest migrate, and the merge step expects 7 profiles.

No -timeout changed, nothing is skipped or marked short, and the node-count guards keep the three shard lists in step. CircleCI's build-test workflow is filtered to main/develop, so the cli node first runs on the develop pipeline after this merges; on this PR the same shard layout is exercised by the GitHub full tier (Go race tests (cli), Go race coverage merging 7 profiles).

How it was verified

• TestKindBaseIsLookedUpOncePerSymbol (internal/semantic/semantics): the memoized lookups answer as the first query did and a repeated implicitUsageBaseFeature, implicitBaseUsage, declaredChainReachesKindBase and ImplicitGenerals query allocates nothing.
• go build ./..., go vet ./..., gofmt -l . empty; go test -count=1 ./internal/semantic/semantics ./tests/stressmodel ./internal/exec/runtime ./cmd/sysml; full go test -count=1 ./... (race-free) locally.
• scripts/race-shard.sh cli prints cmd/sysml, scripts/race-shard.sh rest no longer does; both workflow files parse; make -n coverage-shard SHARD=cli.
• The before/after timings above, same machine, same session, go test -count=1.
• GitHub full tier on 6218508, all green: Go race tests runtime 34 min, runtime-corpus 34, model 23, export 31, rest 30 (without cmd/sysml), migrate 17, cli 9; Go race coverage merged the 7 profiles. Devin Review: no findings.

Checklist

• make test and make lint pass locally (race-free go test ./...; make lint unchanged surface)
• Tests added or updated for the change
• Documentation extended where it already covers the surface (see CONTRIBUTING.md) — the shard lists live in the Makefile help and scripts/race-shard.sh usage, both updated
• Changelog entry added as changes/unreleased/<slug>.<section>.md, not as an edit to CHANGELOG.md
• baselines regenerated and make docs-counts run if a gate count moved (compliance rows need nothing: the census is counted at docs build)
• No internal work-item labels (waves, slices, F4, K5) in the body, docs, or changelog

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.