security: triage and remediate CodeQL baseline alerts

[nguyenthdat]

CodeQL is now enabled on pushes, pull requests, and a weekly schedule. Its initial baseline identified 20 open alerts that need explicit review rather than silent dismissal.

Security dashboard: https://github.com/nguyenthdat/opencode-multi-auth-codex/security/code-scanning

High-severity groups

• js/disabling-certificate-validation at src/web.ts:3028
• js/file-system-race at src/web.ts:3125
• js/remote-property-injection at src/store.ts:535, src/store.ts:550, and src/store.ts:570
• js/user-controlled-bypass at five request-handling locations in src/web.ts
• js/insecure-temporary-file in three test files
• js/bad-tag-filter in tests/web-headless/dashboard-smoke.test.ts

Medium-severity groups

• js/http-to-file-access in src/logger.ts and src/store.ts
• js/log-injection at src/auth.ts:258
• js/stack-trace-exposure at src/web.ts:2332
• js/identity-replacement at src/index.ts:312

Acceptance criteria

• Review every alert against the local-only threat model.
• Fix confirmed production findings with regression tests.
• Replace insecure test temporary paths with securely created temporary directories.
• Dismiss false positives only with a specific rationale in the CodeQL alert.
• Keep CI, CodeQL, bun audit, and the full Bun test suite green.