cleanup(core): move the affected touched-project locators into rust

[AgentEnder]

Stacked on feat/nxc-4861-files-input (NXC-4858). Review only the last commit; the base branch's 74 commits are not part of this change.

Groundwork for NXC-4859 (task-level nx affected). This is "PR 0": move the locators, change nothing else.

Current Behavior

filterAffected runs four TouchedProjectLocators in TypeScript, then reverse-walks the project graph. Three of them (getTouchedProjects, getImplicitlyTouchedProjects, getTouchedProjectsFromProjectGlobChanges) are pure path and glob matching, and they re-run minimatch per (pattern, file) pair — minimatch recompiles the pattern on every call, so the cost grows with changeset size times the number of distinct {workspaceRoot} filesets.

Expected Behavior

Those three move to a native locateTouchedProjects. Same results, same locator order, same duplicate-emitting contract.

The lockfile, npm-package and tsconfig locators stay in TypeScript — they depend on the npm/yarn/pnpm/bun lockfile parsers. They cross the boundary as ThreadsafeFunction callbacks invoked once per run, so the locator set is still extensible from JS rather than hardcoded.

Graph pruning also stays in TypeScript: the native ProjectGraph models dependencies as HashMap<String, Vec<String>>, with no edge type or source, so it cannot rebuild what filterAffected returns.

Also deletes workspace-json-changes.ts — the angular.json locator was already unregistered and has no remaining referents anywhere in the repo.

Behavior deltas worth knowing

Small, and all of them widen or correct rather than break:

• {workspaceRoot}/config/ now matches config/app.json; minimatch did not.
• {workspaceRoot}/!(dist)/** no longer matches dist/a.ts; minimatch's leading-! handling did.
• A {workspaceRoot}/some-dir/ fileset becomes some-dir/**.
• The old implicits[input] ??= [] on a plain object silently dropped any fileset named {workspaceRoot}/__proto__, constructor or toString, because the inherited property is non-nullish. The BTreeMap removes that quirk.
• A self-referencing named input used to overflow the stack; it now terminates.

Performance

Measured on a synthetic 3000-project graph (30k files, 10k external nodes), min of 11, comparing this branch against its base. The trade is a variable cost for a fixed one, so which way it goes depends on how many distinct {workspaceRoot} filesets a workspace declares.

Many distinct filesets — the nx repo itself has 151 across 149 projects:

changed files	before	after
10	373 ms	65 ms
100	1426 ms	172 ms
1000	12824 ms	1239 ms

Few, shared filesets:

changed files	before	after
1	8.7 ms	40.4 ms
1000	59.3 ms	47.0 ms

So a large win where the pattern set is big, and a ~30 ms regression on a small-pattern workspace, from the fixed cost of walking the graph natively. filterAffected also marshals the project graph, which is memoized per graph identity — nx release calls filterAffected once per commit, and without the memo that re-marshalled the whole graph every time.

Related Issue(s)

NXC-4859

[nx-cloud]

Nx Cloud AI Fix could not be generated

View fix generation details ↗

---

View your CI Pipeline Execution ↗ for commit 55d1656

Command	Status	Duration	Result
nx affected --targets=lint,oxlint,test,build,e2...	❌ Failed	11m 19s	View ↗
nx run-many -t check-imports check-lock-files c...	✅ Succeeded	4s	View ↗
nx-cloud record -- pnpm nx-cloud conformance:check	✅ Succeeded	54s	View ↗
nx build workspace-plugin	✅ Succeeded	<1s	View ↗
nx-cloud record -- nx sync:check	✅ Succeeded	18s	View ↗
nx-cloud record -- nx format:check	✅ Succeeded	<1s	View ↗

---

☁️ Nx Cloud last updated this comment at 2026-08-30 02:39:57 UTC