Fix tool-call loop leaks, raise step cap, add message queueing

[MathiTz]

Summary

Fixes a recurring bug where malformed/unrecognized tool-call tags leaked raw JSON into the chat instead of executing, raises the agent loop's tool-call safety cap, and adds message queueing + live step visibility so long-running builds don't get silently cut off.

Changes

• Tag-leak fixes (core/workspace_tools.py): tolerate stray whitespace in <motion_tool> tags, recognize the <| DSML | tool:NAME> channel leak observed from deepseek-v4-flash, and add a general fallback guard so any unrecognized envelope wrapping tool-call-shaped JSON is treated as malformed (triggering a retry) instead of being shown to the user as the final answer.
• Tool-call loop (main.py): raised MAX_TOOL_STEPS from a tight limit to a generous safety ceiling (150), intended only to catch a truly stuck model rather than cap legitimate multi-file builds. If it's ever hit, completed work is reported instead of discarded. Every tool op (not just writes) now streams progress.
• UI (ui/tui.py):
  • New conversations start in read-only "plan" mode (discuss first); an explicit build-trigger phrase or Tab switches to "build" mode.
  • Plan-mode instructions no longer demand file writes; the model is nudged to produce a concrete plan instead of punting back to the user.
  • Added StepsMessage: an always-visible live tool-activity feed in chat, separate from the opt-in "thinking" view.
  • Messages submitted while the agent is busy are now queued instead of cancelling the in-flight worker; the same worker drains the queue once its current turn finishes.
• Memory system (memory/db.py, memory/retriever.py, core/learning.py): fixed inverted cosine-distance ranking in the sqlite-vec path, skip zero-norm embeddings instead of crashing, added a minimum relevance threshold, unconditional recall de-duplication, and SkillSynthesizer now computes a real (or hash-fallback) embedding instead of a hardcoded zero vector.
• Workspace resolution (core/config.py, core/learning.py, core/orchestrator.py, install.sh): anchor the harness's own persistent state (config, memory DB, skills) to its install directory instead of CWD, so motion can be invoked from any workspace directory without losing track of its own state.
• Docs: added docs/cli.md, updated docs/architecture.md / docs/roadmap.md.

Testing

Full test suite passes locally: PYTHONPATH=. .venv/bin/pytest tests/ -q → 66 passed.

Co-Authored-By: Warp agent@warp.dev