v3.20.1 — Stability audit

Stability audit release. A second four-agent deep audit plus a full feature-by-feature runtime verification (module loading, command registry, tool dispatch, end-to-end chat loop, IDE bridge). Every clear-cut finding is fixed with a regression test — suite 710 → 729.

Fixed — data loss & correctness

• Writing into a pre-existing empty file no longer deletes it on /undo or /rewind (a None-vs-"" undo-sentinel collision).
• multi_edit and diff_apply now respect batch-review mode (/accept) instead of writing to disk instantly and skipping accept/reject.
• delete_file is now undoable and rewindable for text files.
• The free-local default model no longer routes to the cloud — qwen3-coder (local Ollama DEFAULT_MODEL) was also a Qwen cloud alias, so the default hit the Qwen API. Coding-Plan users use qwen-coder or qwen3-coder-plus/qwen3-coder-next.

Fixed — MCP & session integrity

• MCP tool calls to underscore-named servers (mcp_my_server_x) were misrouted — every tool on any _-named server was unreachable.
• Image + Claude no longer poisons the session — attaching an image on an OpenAI-compat model then switching/failing-over/compacting on Claude now converts image_url → Claude's image format instead of 400-ing every later request.

Robustness & hardening

• multi_edit skips a malformed edit element instead of failing the whole call.
• OpenAPI sources degrade gracefully on malformed specs.
• A stalled Ollama pull is reported as failure, not "installed!".
• Bridge bearer token compared in constant time; OAuth/bridge sockets closed on stop.
• /clear resets the cached context-token count (no bogus auto-compact).

Full details in CHANGELOG.md.