fix(toolchain): replace Wrangler/Miniflare GPL-family path

[seonghobae]

Purpose and owner boundary

Replace the former Wrangler/Miniflare/Sharp development path with pinned workerd@1.20260625.1 + esbuild@0.28.1, keeping Cloudflare lifecycle truth and foreign routing/security authority outside Noema.

Current protected-base repair — 2026-09-06 KST

Protected Noema is main@e26d771470a4ece873c367b40b3cd6cb03ac7de3. Because the lockfile policy requires exact PR-base equality, the post-#527 base movement made the prior baseSha=5b8e620... stale. Causal commit 7c8c26983f5628ccf6125ed3d3cb52edc9782f9d changes only .github/lockfile-change-policy.json.baseSha to e26d771..., preserving package/top-level digests, source URLs, target inventory and toolchain bytes. Ordinary two-parent/non-force restack 6b7f0a7b8c3069a815f74ee654620e59574bd4e1 then adopts merged #527 OIDC trust without weakening licensing/scanner policy.

Fresh exact-head ci 34027015120 and reviewer-ci 34027015135 are pending; required Security Scan 34027015129 and patch-validator-image 34027015123 are queued. Predecessor results do not transfer. Keep Draft until unchanged-head terminal GREEN and clear current review state.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[seonghobae]

Evidence: lockfile regenerated, 3 stale test fixtures repaired — 5c17c82

Fresh exact-head CI (ci 33772121594 / job 100734009932, and lockfile-reproducibility job 100733942817) surfaced real, non-congestion failures on c09f1cb. Diagnosed via job logs and confirmed locally on the exact pinned toolchain (downloaded Node.js 24.19.0 / npm 11.17.0 to match devEngines):

1. package-lock.json was stale — exactly the causal gap the PR description already flagged. Regenerated with npm install --package-lock-only --ignore-scripts --no-audit --no-fund --legacy-peer-deps=false --install-links=false in a fresh directory (same recipe as lockfile-reproducibility.yml); the result is byte-identical (cmp clean) to what that workflow's own fresh-directory regeneration produces. Confirmed test/cloudflare-toolchain-license-boundary.test.ts now passes and the lockfile no longer contains node_modules/wrangler, node_modules/miniflare, node_modules/sharp, @img/sharp-libvips-*, or any LGPL-3.0 entry.
2. test/upload-artifact-node24-integrity.test.ts — the reviewed-workflow inventory was missing the new lockfile-reproducibility.yml. Its two actions/upload-artifact@ uses already pin the reviewed SHA (043fb46d... / v7.0.1), so this is inventory-tracking only, not a new review decision.
3. test/patch-validator-image-contract.test.ts — asserted the old single-line npm pkg delete devDependencies.@cloudflare/workers-types devDependencies.wrangler; the current workflow correctly deletes workerd/esbuild too (this PR's new devDependencies) across multiple continuation lines. Updated the assertions to match, and added the corresponding node_modules/esbuild / node_modules/@esbuild absence checks the workflow already performs.
4. test/cloudflare-worker-config.test.mjs ("fails closed when a root field would be silently omitted") — this fixture appended the new key after the [vars] section header. TOML is section-scoped, so it was actually exercising the (intentionally open-ended) vars path, not the root-key allowlist. Moved it ahead of [vars]; the parser's root-key rejection was already correct and required no change.

Verified via npm run typecheck and the full vitest suite on the pinned toolchain. Remaining local failures (acquisition symlink/owner-only-mode checks, one SIGTERM-reaping timing test) only reproduce because this sandbox runs as root/uid 0 — untouched, and consistent with CI reporting exactly 4 failed files before this push.

Pushed non-force, fast-forward from c09f1cb → 5c17c82. CHANGELOG.md Unreleased updated. No security/license/deployment-boundary logic was changed — this is lockfile regeneration plus test-fixture repair only.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BV96rXhqoR3tYZ9AeAVur4

---

Generated by Claude Code

[seonghobae]

Fresh current-head authority supersedes the stale body snapshot: protected main is e1ac9d50f6c646f04be8c137c8acdc7200182fcd, exact #540 head is 190908ee21b71c062d263bd337231c3f0a1e77cc, open / Draft / mergeable. Exact-head ci 33856520016, required Security Scan 33856520006, reviewer-ci 33856520098, patch-validator-image 33856520066, and lockfile-reproducibility 33856520023 all completed cancelled; reviewer job 100971000064 ended with steps=[], runner_id=0, empty runner name, so no repository code executed. This is not a source/test RED or GREEN and predecessor evidence does not transfer. Keep Draft and preserve the toolchain/license gates; cancellation belongs to the runner/control-plane owner path rather than a leaf source change.

[seonghobae]

2026-09-06 KST post-#546 repair finding: protected base is now main@85b17014b8d46eacc95e096ca114568c321d0263, but this branch's .github/lockfile-change-policy.json still binds baseSha to e1ac9d50f6c646f04be8c137c8acdc7200182fcd. Because #546 did not change package.json/package-lock.json but did change protected reviewer/CodeGraph/CHANGELOG surfaces, this is not a Close condition and not a safe wholesale tree merge. Keep Draft; the next repair must semantically restack the toolchain/workflow/CHANGELOG/baseline delta on 85b17014..., rebind the lockfile policy only after re-verifying the before/after bytes against the new protected base, and generate fresh exact-head gates. No force push, predecessor GREEN transfer, or lockfile hand edit.