Naragate v0.8.0 — Per-user MCP tokens + standalone engine

Naragate v0.8.0 — Per-user MCP tokens + a standalone engine

This release closes the gap between the web UI and non-web agent surfaces.

Per-user API tokens (MCP / skills / agents)

MCP had no cookie, so every non-web run used the backend's deployment key and
shared one credit ledger. Now each user mints a token in Settings → MCP &
API Access (NARAGATE_TOKEN); the backend resolves it to the same email as
the web session, so a non-web run uses your own Sectors key, evidence cache
and credit ledger. Tokens are stored hashed and revocable.

New MCP tools: whoami, get_setup_status, bind_sectors_key, plus full
follow-up Q&A parity — ask_followup, get_followup_suggestions,
next_followup_suggestion. Tool surface: 30 tools (15 high-level + 15
tools.yaml primitives).

Standalone engine (no hosted or Redis dependency)

MCP/skills/agents depend on an engine, never on the hosted domain. The
engine is now a one-command process:

pip install naragate-engine && SECTORS_API_KEY=... naragate-engine
# or
docker run --rm -p 5678:5678 -e SECTORS_API_KEY=... ghcr.io/masdevid/naragate-engine

The Evidence Graph cache degrades to an in-process store when Redis is
unreachable, so the engine runs with zero infrastructure. When no engine
answers, MCP returns an actionable offline hint.

Fixes & docs

• Follow-up chat race condition — template chips are locked while the agent
  is still thinking, and the dismiss → answer → replace flow is single-flight.
• Reconcile the agent list across all READMEs (agents now match the 13 skills;
  added filings-agent, follow-up, renderer) and correct the MCP tool count.
• docker-compose now passes per-agent model overrides (previously dropped).