feat: fail-closed guard against proprietary source in this public repo

[marc-n-dream]

Why

This repository is public. The Android and webview repos that feed it are private, and only
compiled artifacts (.aar, .aar.tmpl, .unitypackage, .bundle) plus accepted glue
(.cs, .jslib, .mm) are supposed to cross the boundary. A leak of native source here is
irreversible, so the bar is "a routine mistake cannot leak it" — not "reviewers will notice".

What

• scripts/check-proprietary-source.sh — denylist guard, bash 3.2 safe. Blocks *.java/*.kt
  (plus .orig/.bak/~ variants), *.gradle, *.pro, *.sh, *.toml, Rakefile/*.rake,
  the gradle wrapper, .gitmodules, AndroidManifest.xml and src/main/res/. Allowlists Unity
  gradle templates, the guard script itself, and content-free *.meta.
• .githooks/pre-commit — local gate. Enable with git config core.hooksPath .githooks.
• .github/workflows/no-proprietary-source.yml — the authoritative gate. Runs on every push and
  PR, and additionally rejects submodule gitlinks so the private repos can never be referenced
  from here.
• AGENTS.md — documents the boundary.

Limits

The guard is path-based. It cannot catch native source pasted into an accepted file (.cs, .md).
Documented rather than fixed.

Verification

Guard block/allow paths exercised directly, and again through the cross-repo webview sync with a
planted .java in the payload — the sync aborts. An adversarial review pass closed 10 bypasses
before this branch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ShCMtnweyHuUEJQn2EH5d6

[marc-n-dream]

⚪ No Jira key on the branch or PR — reviewed on the PR body alone.

Nothing blocking. I ran the guard instead of trusting its comment block: clean full-tree scan, all 19 deny paths blocked, all 10 accepted paths passed, delete-only commit survives set -u on stock bash 3.2, and the allowlist currently rescues exactly one file (itself). Four things to look at — the pick of them being a fail-open in a script headed "fail-closed".

Scope note: the "What" section lists four files; the diff touches eleven. The .serena/** untracking and the .gitignore rewrite appear nowhere in the body — flagged inline.

[marc-n-dream]

⚪ No Jira key on the branch or PR — reviewed on the PR body alone.

Second pass at the same commit; the four findings from the earlier review still stand and are not repeated. Three new ones, and the headline is that .mm sits on the accepted list while 2,493 lines of native Objective-C++ already ship in this public repo under it — legitimately, as it turns out, which is precisely what the documentation neglects to say.

Verified this round beyond the last pass: the gate genuinely runs and passes in CI (No proprietary source green on both the push and pull_request events for this head), the generated deny regex is well-formed, DENY_PATTERN/DENY_REASON are 13/13 so no reason-drift, and there is no Java or Kotlin anywhere in the tree — tracked, untracked, or ignored. The Android half of the boundary holds: native ships as unityandroidlibrary.aar and WebViewPlugin-*.aar.tmpl.