Bleeding-edge evergreen upgrade -> v1.2.0

[ABCrimson]

Summary

Bleeding-edge evergreen upgrade of the entire monorepo, released as v1.2.0. Every dependency and tool moves to its newest available release (including pre-releases, per the evergreen policy), the codebase is migrated to the new APIs, and all user-facing documentation is corrected to 100% accuracy against the upgraded code.

No public API or behavior changes — this is a toolchain/dependency refresh. The one consumer-facing change is the raised minimum Node version (engines.node >= 26.4.0), which is why this is a minor (1.1.5 → 1.2.0) rather than a patch.

What changed

Runtime / tooling

• TypeScript 6.0.1-rc → 7.0.1-rc (native Go compiler)
• React 19.3 canary (Jun 26 build) + react-dom (same hash); radix-ui 1.4.4-rc → 1.6.0
• Vite 8 GA (8.1.0), Vitest 4.1 → 5.0.0-beta.5, Playwright 1.59 → 1.62.0-alpha
• Biome 2.4.6 → 2.5.1, tsdown 0.22.3, pnpm 11.9.0, VitePress 2.0.0-alpha.17
• engines.node >= 26.4.0; CI on Node 26.4.0 / ubuntu-26.04
• Rust/WASM crate: edition 2021 → 2024, wasm-bindgen 0.2.100 → 0.2.126, wasm-pack 0.15

Migrations the new versions required

• pnpm 11.9: overrides + allowBuilds moved to pnpm-workspace.yaml
• Vitest 5: benchmark fixture API (with required .run()), it.sequential → it, oxc transform
• radix 1.6: dialog now dismisses on a full outside click (unit tests updated; root-caused via a throwaway probe)
• Rust: wasm-opt feature-flag config (cached binaryen couldn't validate the new bulk-memory output)
• Cross-platform WASM build guard (node existsSync instead of POSIX [ -f ], which failed on Windows)
• CI action majors: checkout v7, pnpm/action-setup v6, codeql v4, pages v5
• lefthook hooks routed through pnpm exec (bare binaries weren't on the hook shell's PATH)

Verification — all gates green

Gate	Result
tsc --noEmit (TS 7, all packages)	✅
Biome 2.5.1 lint	✅
Unit (Vitest 5)	✅ 416 passed
E2E (Playwright 1.62, Chromium)	✅ 97 passed
Full build (7 packages)	✅
Bundle size	✅ core 6.28/6.5 kB · react 10.18/12 kB
Docs build (VitePress α17)	✅
Rust crate (cargo check + wasm-pack)	✅
Benchmarks (Vitest 5 fixture)	✅ 22 benches, real measurement

Deep review

A multi-agent review of the diff (each finding adversarially verified) caught a real regression I had introduced: Vitest 5's bench() is a factory — without .run() all 41 benchmarks "passed" in ~2 ms measuring nothing. Fixed and verified. Also hardened: the size-limit React budget now externalizes the react/react-dom/radix-ui peers and measures the adapter's own code (~10.2 kB, stable across peer bumps), and the scaffold tsconfig (ES2026/ES2025 are invalid TS targets → ESNext).

Documentation accuracy pass

A 10-agent audit cross-checked every doc against source. 52 corrections applied across the README, npm pages, the 33-page VitePress site, CONTRIBUTING/ARCHITECTURE, and CHANGELOGs — version references, bundle-size figures, test counts, and 15 verified API-example bugs (FrecencyEngine constructor, getBonus, KeyboardShortcutRegistry.register/getConflicts, ParsedShortcut shape, formatShortcut, createSearchEngine).

Left unchanged on purpose (historical artifacts, clearly dated): cmdk-complete-rewrite-specification-v2.md (the March 2026 design spec, which even predates the modern-cmdk rename), docs/plans/*, and prior CHANGELOG release entries.

Deliberate decisions (flagging for review)

• changesets/action kept on stable @v1 (not the v2-next prerelease) — it sits in the npm-publish path; v1 supports every input used.
• Scaffold template uses react ^19.2.0 (not ^19.3.0, which has no installable stable) so generated projects actually npm install.
• engines.node >= 26.4.0 is aggressive for a browser library (it's a build-env constraint). Per the evergreen directive. Easy to relax if you'd rather broaden compatibility.

Release

Targeting v1.2.0. Per the agreed plan, the git tag + GitHub Release will be created on main after this merges (npm publish via the existing release CI / NPM_TOKEN).

🤖 Generated with Claude Code