fix(keychain): customer QA fixes for SDK, passkey recovery and onboarding

[clawdbot-glitch003]

Summary

Customer-path QA of https://keychain.litprotocol.com using public docs and published npm 2.0.2 first, followed by fixes. No real customer credentials, paid provider calls, attestation bypass or funded transactions.

• Fix default Node SDK TLS-to-enclave binding via conditional package entry, preserving browser bundling.
• Fix packed TypeScript declarations and test strict external tarball consumers.
• Reject inconsistent identity metadata, prototype-property lookup traps, invalid MCP request IDs and destroyed-client operations; add CLI help/version.
• Fix cached-passkey/recovery login selecting a new empty vault. Explicitly loaded backups now take precedence over duplicate root-owned vaults; root backup recovery works without backend discovery, but only for an exactly matching selected credential. Authorization checks remain in place.
• Publish owner/provider guides and missing /README.md; remove broken platform links; clarify export/get versus service/use, Supabase credential shape, MCP paths, credential exposure, attestation, rotations and write retries.
• Prepare SDK 2.0.3 and align examples. Not published or deployed. Publish and verify 2.0.3 before deploying these version-pinned docs.

Verified against production

Real virtual-CTAP2 passkey registration and existing-key login; dummy secret create/approve; attestation-enabled CLI/SDK/MCP retrieval; child environment/file handling and cleanup; denied unapproved agent; disable/enable; revoke/reapprove; rotation v2 with old config; permission renewal; backup-selected recovery using only the recovery passkey; execution-key replacement (old config 401, new config succeeds).

The fixed Node SDK also completed live attestation with all six checks, including tls-certificate-in-tee. Production customer-path tests used published 2.0.2; local regression tests verify the candidate fixes, not their deployment.

Local validation

• npm test: 88 passed, 3 environment-dependent skips, 0 failures.
• PostgreSQL-backed local integration reran those API cases: 3 passed, 0 skipped; 9 Rust tests passed, 1 live Stripe test ignored; 3 application browser E2E tests passed. Lit/Stripe are local mocks in this suite.
• Dedicated passkey browser regressions: 10 passed. Real Chromium virtual-authenticator WebAuthn and application UI; public network mocked, stopping before real TEE login.
• Build, TypeScript, strict Clippy, formatting and diff checks passed.
• Packed tarball strict TypeScript, Node TLS runtime fixtures, and browser bundling passed.
• Independent review found no additional blocking SDK/initial-passkey security or logic defects; explicit-backup follow-up inspected and rerun at integration.

Repeated GitHub first-test module-import failures led to replacing the Vite dev server with an isolated prebuilt UI/identity harness. No test retries or weakened assertions. Three cold starts and a CI-mode local run each passed all 10 passkey tests; the integrating reviewer independently reran all 10 successfully. Two new harness readiness/cleanup tests pass. Exact original optimizer race is not claimed proven. All GitHub checks passed on final commit cc22df17a0909939947c03acc300363e9852f71f, including the complete Keychain job. Ready for review; not merged, deployed or published.

Honest coverage limits

Provider success calls need disposable GitHub/Stripe/OpenAI/Slack/Supabase credentials. Google/wallet login, physical/cross-device passkeys, paid checkout/cancellation, over-limit billing, actual elapsed expiry and complete backend data-loss recovery remain unverified. Two intermittent production failures are recorded without unsupported root-cause claims. This is broad customer-path QA, not proof of exhaustive security coverage.

Detailed report: user-qa-2026-09-16.md. Private keys/configs/backups excluded.