fix: qualify real relay and validate dependency audits

[sarthakagrawal927]

The real relay qualifier now exercises actual WebSocket behavior and requires proof that the inspector request is published by resolved Undici 6.29.0. Previously the fixture escaped its module root and failed at startup; after fixing that, Node’s shared global dispatcher still selected its internal publisher. This corrects both boundaries and preserves main’s dependency-security fixes.

• Reconcile current main without restoring expired audit exceptions; validate audit JSON and Swift-format command results with fail-closed helpers.
• Patch the existing PartyKit entry to delegate to installed Undici 6.29.0. The inspector request uses its own Agent, closed after consuming the response or on failure. Retained embedded factories/source-map bytes are explicitly inert, not an upstream-update claim.
• Keep temporary configuration, state and generated files in an owned fixture, bind loopback, separate relay/inspector ports and use the read-only checkout as the module root. Sensitive reads, checkout writes and symlink escapes are denied by the cooperative Node guard.
• Qualify native V8 publisher call sites, executable/map identities and source hashes. Synthetic channel publications, unmatched records and source-map strings cannot establish publisher proof.
• Run the actual relay in a Linux network namespace in CI. Keep the installed server-consumer regression in its owning workspace, with explicit Vitest/Knip discovery and no new dependency or ignore.

Validation at 8f12bce61980826d8dbc10de9d538a7dbb45e74e: exact pnpm 10.33.2 frozen install; Node 22.23.1; 26 focused tests; full pnpm check with 37 tests, lint, server/web typechecks, web build, strict Knip, cycles and all existing health ratchets; workflow actionlint and diff checks. Final actual relay passes handshake/presence, ping/pong, normalized pose/start, duplicate-display rejection, controller replacement, malformed input and disconnect checks. The actual inspector publisher is external Undici 6.29.0 with source SHA-256 79c1d8234b602f1ccae9f38517851fc0cb7d67f7b4a20217d3a297016dbb863a.

Fresh hosted CI passed at this exact source head: full code-health checks, the actual relay in a Linux network-isolated namespace, the unsigned iOS simulator build, and Swift quality ratchets.

Residuals: 0 critical/high and 4 moderate dependency advisories (esbuild, PostCSS, Vitest and its mocker). Ordinary audit still exits 1. Existing debt remains: 6.43% line coverage, 3 complexity violations and 5 duplication groups; no ratchet was weakened. Login/deployment, physical camera/ReplayKit, device play feel, signing and release remain unqualified.

Refs #26. Keep #26 open. This PR publishes source for review; it does not deploy or release the product.

[sarthakagrawal927]

Hardened the false-clean Swift-format failure case without weakening debt ceilings, and retained bounded network-disabled embedded factory checks plus a guarded synthetic relay harness. Root independently ran all30tests, format and full pnpm check successfully; exact PR CI37049677869 is green. The actual earlier local relay verified four protocol groups but failed the bundled-inspector provenance assertion; compatibility/security attribution remains unqualified. PR39 stays draft and issue26 open. No dependency, protocol, native app, provider or release changes.

[sarthakagrawal927]

Rebase check (Claude): this branch already contains main (merge commit 8f12bce includes #40), is MERGEABLE, and CI is green, so no rewrite or force-push was done. Overrides on the branch resolve to brace-expansion 5.0.12 and undici 6.29.0 (via miniflare>undici instead of #40's undici@^5.28.4 key; lockfile result is identical). It also adds the partykit patch and packageManager: pnpm@10.33.2, which duplicates #41 (merged, same pin and CI change). Left for the owner; the PR is not marked draft, and the owner may want to convert it to draft.

🤖 Generated with Claude Code