ci(test): six heavy packages split test into test and test:repo so package suites cache again

[claude]

Fixes #16466

Ruling: option D, seat comment 5565158780 on #16466 (six conditions), replacing the card's Ruling and Acceptance after the measured separability failure (table: comment 5565114830; first report: 5565121423). Contract tier, clause-② yes: what CI verifies per package at merge time changes shape. Draft PR, ordinary queue landing — the seat lands it.

What changes

Every test file stays where it is. @objectstack/spec, core, types, runtime, objectql and rest each split their suite into two vitest projects in one vitest.config.ts — local (every other test file) and repo (the tests that read outside the package, listed in a vitest.repo-tests.json sidecar) — and two turbo tasks: test keeps package-local inputs ($TURBO_DEFAULT$), test:repo takes the wide $TURBO_ROOT$ inputs that used to sit on test and re-run the whole suite on any change inside the radius. pnpm test at the root runs both tasks.

• scripts/check-cross-package-test-inputs.mjs (Layer B, condition 1): for a package whose manifest declares a test:repo script, the declared globs must hash on PKG#test:repo, PKG#test must carry NO $TURBO_ROOT$ input, and the sidecar must equal the gate's own scan in both directions (an escaping file left in test reds naming it; a listed file that no longer escapes reds; an unreadable list reds prescribing the scan). New self-test battery the SPLIT test:repo task (#16466), 16 cases, both directions, readSplit driven on disk; battery roster and floor bumped (7 → 8). Real tree: OK: 28 package(s) read outside themselves, all declared, and turbo.json hashes every declared glob (6 of them on a split "test:repo" task). — All 133 self-test cases passed.
• scripts/check-test-completeness.mjs: PKG:test:repo: summary prefixes, ::group:: and bare PKG:test:repo headers and Failed: PKG#test:repo rosters attribute to the package; test:smoke still attributes nothing; 9 new self-test cases. Real two-task turbo logs graded: types + trigger-api 602 test(s) declared and all accounted for; core + types + rest 5008; spec 13149; runtime 3361; objectql 4806.
• .github/workflows/ci.yml: the Test Core whole leg runs pnpm turbo run test test:repo $FILTERS ...; the cli slice leg stays test-only (cli is not split). turbo neither runs nor counts the NONEXISTENT test:repo of an unsplit package (Tasks: 7 successful, 7 total for types + trigger-api, zero trigger-api:test:repo lines), so the guard's run-completed rule is unchanged.
• turbo.json: base test:repo task; six PKG#test:repo overrides carrying the moved inputs (spec 30, core 1, types 1, runtime 4, objectql 2, rest 5); the six PKG#test overrides keep only the four package-local inputs. ci: e2e and live tiers move to a nightly run on main; PR and queue runs keep unit, integration and conformance (maintainer-directed, part B tiering) #16455's env: ["OS_TEST_TIERS"] stays on test and every PKG#test; test:repo declares no env because nothing under it reads one — scripts/nightly-tiers.mjs is the one reader and only packages/cli/vitest.config.ts consults it; the nightly workflow runs turbo run test alone, which is right (the repo projects hold no tiered file).
• Each project re-declares the root block's test options (disableConsoleIntercept, env.OS_REGISTRY_LOG, globals, environment, testTimeout where the root had them): a root-level value is inert for a project run on vitest 4.1.10, which check-registry-log-declared and check-console-intercept-disarm both pin (the first went red on the first cut and is green now).

Condition 2 — the CI turbo cache precondition, measured before the ci.yml edit

Test Core DOES restore a turbo cache: actions/cache/restore@v6 on .turbo/cache, key RUNNER_OS-turbo-JOB-SHARD-REF-SHA with prefix restore-keys down to RUNNER_OS-turbo-JOB-, and the save step Save Turbo cache (main only) runs if: always() && github.event_name == 'push'. So the first case of the ruling holds: the shard runs test test:repo for the selected packages and the restored per-shard cache carries an unmoved PKG#test. Stated caveat: a package's seed lives under the shard INDEX that ran it on the last main push, and a PR or merge-group run partitions a different (affected) package list, so a hit needs the same index to draw the package — LPT places the heaviest packages first, so spec (the floor) is the stable one; the partitioner's own header records that legs of one shard index range from fully replayed to fully cold. Task-level selection (the ruling's second case) is therefore still the deterministic answer for docs-only groups and is left as the follow-up question in the report, not done here.

Condition 5 — the ablation in the new shape (scratch commits, both reset away, never pushed)

Tree 22d3581888 with a warm local cache for all six split packages (each test and test:repo run once through turbo). Two scratch commits on the branch, each followed by turbo run test test:repo --dry-run=json, each reset to the pushed head with git reset --hard and proven (HEAD==PRE yes · diff-empty yes · porcelain [] · blob == HEAD yes; remote head unchanged before and after).

Docs-only scratch commit 2296808208 (content/docs/concepts/north-star.mdx): exactly 3 task hashes move → @objectstack/rest#test:repo, @objectstack/spec#test:repo, create-objectstack#test.

task	clean	docs-only commit	leaf .ts commit 935ce423fd (packages/connectors/connector-slack/src/index.ts)
@objectstack/spec#test	HIT	HIT (hash same)	HIT (hash same)
@objectstack/spec#test:repo	MISS*	MISS (hash moved)	MISS (hash moved)
@objectstack/core#test	HIT	HIT	HIT
@objectstack/core#test:repo	HIT	HIT	MISS (moved)
@objectstack/types#test	HIT	HIT	HIT
@objectstack/types#test:repo	HIT	HIT	MISS (moved)
@objectstack/runtime#test	HIT	HIT	HIT
@objectstack/runtime#test:repo	HIT	HIT	MISS (moved)
@objectstack/objectql#test	HIT	HIT	HIT
@objectstack/objectql#test:repo	HIT	HIT	MISS (moved)
@objectstack/rest#test	HIT	HIT	HIT
@objectstack/rest#test:repo	HIT	MISS (moved)	HIT
create-objectstack#test	MISS (never run locally)	MISS (moved)	MISS (same)

The leaf commit moves 11 hashes in all: the three genuinely affected packages' tasks (connector-slack, example-showcase, dogfood) plus the five test:repo tasks that hash packages/**/*.ts; the six test hashes do not move. Before this PR the same two edits moved spec#test, rest#test, create-objectstack#test (docs) and spec/core/types/runtime/objectql#test (leaf) — the whole suites (first report, comment 5565121423).

*spec#test:repo reads MISS on the clean tree because spec's declared packages/**/*.json sweeps vitest's own node_modules/.vite/vitest/ID/results.json, rewritten by every run: a diff of the task's 9404 inputs between two consecutive dry-runs moves exactly that one key. It was the whole spec suite's property until this PR confined it to the 33s task; filed as #16555; this PR does not address it.

Condition 3 — no file lost, none counted twice

vitest list --filesOnly on the merged tree (22d3581888, main 78e8485d67 merged): the pre-split count is main's own config run against the same tree (a probe config written into the package and deleted after).

package	pre-split (main's config)	local	repo	local + repo	overlap	repo == sidecar
spec	484	458	26	484	0	yes
core	51	49	2	51	0	yes
types	20	19	1	20	0	yes
runtime	238	236	2	238	0	yes
objectql	285	284	1	285	0	yes
rest	188	187	1	188	0	yes

Partitioner: unchanged. It weighs PACKAGES, the shard runs both tasks for a package, and each package's timings row was measured with the repo files inside test, so the row still covers both tasks until the next refresh; a package without a row is estimated at countTestFiles × 1.522s by the existing fallback. partition-test-shards --self-test: self-test OK (71 measured packages → 72 shard items, 6 shards, max/mean 1.00x at most 1.3x, floor 404s, bins 672/672/672/671/672/671s). The next refresh's generator keeps only task === 'test' samples and would drop test:repo seconds — filed as sub-issue #16550 of #16464, whose branch rewrites that generator.

The suites in the new shape (local, shared box, every leg through the verify lock)

Through turbo (pnpm turbo run test test:repo --filter=PKG --concurrency=4 --summarize --log-order=stream), every leg exit 0: spec local 458 files / 12759 tests (369s) + repo 26 / 390 (33s); runtime 236 / 3292 (211s) + 2 / 69 (14s); objectql 284 / 4801 (210s) + 1 / 5 (0.7s); rest 187 + 1 / 8; core 49 + 2 / 42; types 19 / 571 + 1 / 7. One local-only red on the way: packages/types/src/node.test.ts's host-only-package pin fails once packages/plugins/organizations/dist exists on the box (#16215 made that package a workspace member; CI stays green only because that dist is never built on types' shard) — reproduced CI's condition by removing that one local artefact; filed as #16552; out of scope here.

Measurement (ruling step 1, from the first delivery)

27 packages / 113 escaping files; the six heavy packages hold 33, of which 20 import private internals; every wide glob's holder on spec, runtime and rest cannot import through a public entry (verified against the built dist/*.d.ts). Escaping share of each suite (local full suites): spec 7.3%, core 4.0%, types 32.1% of a 13s suite, runtime 1.0%, objectql 0.1%, rest 0.1% — the win is the cacheability of the remaining 93–99.9%. Full 113-row table with durations, globs and import style: comment 5565114830 on #16466. Options B (four-file move) and C (publishing REST_ROUTE_LEDGER and runtime's two bindings) are out of scope per the ruling's condition 6, recorded on the card, not filed.

Serial proofs and gates

• Merge-tree, driver-less bare clone (git clone --bare --shared, merge.os-regen.driver unset there): mine f5f116e173 × ci(test): e2e and live filename tiers move to a nightly run on main #16481's then-current head 07464ab227 → exit 0, tree 3c24652828, no names (the merged tree read back: env on test and every PKG#test, the radius on PKG#test:repo, both ci.yml hunks present); origin/main b1b978c8df × mine 22d3581888 → exit 0, tree 0976e1390c, no names. ci(test): e2e and live filename tiers move to a nightly run on main #16481 landed at 08:33:55Z (f48f3f1b21) and is merged into this branch (b62f8d451b); the merged turbo.json and ci.yml compose exactly as the probe's tree did.
• node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack at b62f8d451b: 99 commands (the family gained check:sourcemap-no-sources-content from main), every exit captured before any pipe: 99 exit 0 (check:dual-build-cjs-loads refused with exit 3 while organizations/dist was deliberately absent, and reads 104 published require entry point(s) across 67 package(s) load once it was rebuilt).
• check-governed-merges --test over the 23 paths: 0 of 23 path(s) hit the register — NOT governed.
• pnpm check:nul-bytes: OK (scanned 8143 text file(s) ...). Whole-repo pnpm lint (eslint . --no-inline-config) at b62f8d451b under the verify lock: exit 0.
• Changeset: none — skip-changeset. The diff publishes nothing from any released package: every touched package publishes dist, README.md, CHANGELOG.md (spec adds its json-schema/liveness/prompts trees) — the sidecar and the vitest config are outside those whitelists, and a package.json scripts line is dev-only.

Generated by Claude Code

🤖 Generated with Claude Code

https://claude.ai/code/session_019RfFHiRCSs3JXLK4cwcfox

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

1 anchor(s) derived from 6 changed package(s); no hand-written page names any of them. ⚠️ 6 changed file(s) yielded no anchor (packages/core/vitest.repo-tests.json, packages/objectql/vitest.repo-tests.json, packages/rest/vitest.repo-tests.json, …), so the pages documenting them are NOT COVERED by this run — this is not a clean bill of health for those files.

What this run could not see

• 6 changed file(s) yielded no anchor (packages/core/vitest.repo-tests.json, packages/objectql/vitest.repo-tests.json, packages/rest/vitest.repo-tests.json, …) — pages documenting those are invisible to this run
• a page that states a rule by its inputs shares no identifier with the emitter that implements the rule, so an emitter-only diff cannot list it — not on this run and not on any run. Measured on fix(driver-sql): emit varchar(maxLength) for a text field a declared index keys on #11430: content/docs/protocol/objectql/types.mdx documents the text-family column mapping by the ObjectQL type names it maps FROM (text / textarea / html) while the diff changed createColumn; it went unlisted, and it was the page that diff falsified, in four places. No shared token exists to detect this on, so a rule your change carries has to be re-read by hand in the pages that restate it.

Coarse fallback — 141 page(s) merely mention a changed package (the pre-#9192 predicate, kept for the deliberately-wide backstop): node scripts/docs-audit/affected-docs.mjs --json ce8caba91403c8f160cb7764c63b08371a13db99 → packageMentionDocs.

Which tree this was computed on

This run read content/docs from b9626589ffe8881fee379247420a8f4991ad9c61 — the merge of head b62f8d451b1d95a7694631af30270be2d91b2975 into base ce8caba91403c8f160cb7764c63b08371a13db99, which is what actions/checkout gives a pull_request run. Not the PR head.

A worktree cut from an older main holds a different content/docs, so re-deriving there can legitimately return a different list — that is a different tree, not a wrong row. To answer on the same tree:

# while this PR is open — GitHub drops the merge commit once it closes
git fetch origin b9626589ffe8881fee379247420a8f4991ad9c61 && git checkout b9626589ffe8881fee379247420a8f4991ad9c61
# afterwards, rebuild it from the two parents, which stay fetchable
git fetch origin ce8caba91403c8f160cb7764c63b08371a13db99 b62f8d451b1d95a7694631af30270be2d91b2975 && git checkout -B drift-repro ce8caba91403c8f160cb7764c63b08371a13db99 && git merge --no-ff b62f8d451b1d95a7694631af30270be2d91b2975

node scripts/docs-audit/affected-docs.mjs --json ce8caba91403c8f160cb7764c63b08371a13db99

⚠️ That checkout carried uncommitted changes, so the commit above does not fully identify what was read.