fix(codex): route native image edits through official fallback

[MattLYT]

Why

Native Codex image editing (POST /v1/images/edits, e.g. gpt-image-2) misses the compiled catalog in a schema-v2 MultiRouter containing only text models. Raw forwarding then selects the outer router's local URL, and the self-loop guard rejects the request before it reaches an upstream.

The observed log reported route_missed=true and effective_upstream_local_proxy_self_loop. Commit 8b29949 removed fallback from generic strict routing; image generations already use a separate image provider resolver, but image edits did not.

What changed

• Apply the existing image provider resolver only to native Codex POST image edits, preserving explicit third-party image routes. Ordinary Responses and generic raw routing remain catalog-only.
• Parse multipart model and stream fields using multer without rebuilding uploaded image/mask bytes or changing content type/query. Reject malformed, duplicate, file-valued, oversized, or invalid routing fields.
• Support the existing Codex path alias and exact double-v1 image-edit alias. Exclude vendor suffix paths, other methods/endpoints, external API clients, xAI OAuth, and GitHub Copilot from the official image fallback.
• Group the existing chat conversion helpers' public/upstream model parameters to resolve their two argument-count CI errors without lint suppression or behavior changes.

Verification

Regression tests cover strict routing boundaries, official image fallback, explicit image routes, external authentication precedence, malformed multipart input, and a local HTTP upstream receiving byte-for-byte image/mask uploads with the original content type/query. The official-route handler test reaches the deliberately missing test OAuth state instead of self-loop rejection.

Tests use isolated homes, in-memory databases, and synthetic uploads. No production credentials were used, the installed desktop app was not replaced, and a live authenticated image edit was not replayed.

On the merged upstream base aff5fa46, all 101 handler tests (including the 10 image-edit regressions) and 9 raw-passthrough tests passed; formatting also passed. Strict Clippy remains blocked by 13 existing diagnostics in unchanged history/usage modules, outside this PR. The full default-parallel test build exhausted local memory/pagefile; the scoped tests above were successfully compiled and run with -j 1. Full-suite success is not claimed.