feat(railway): one-click Railway deploy + team hosting guide (Phase 2)

[Coding-Dev-Tools]

What (Phase 2 of the team-hosting plan)

One-click Railway deploy so a team admin hosts one instance on their own Railway account (their cost), and teammates just log in + connect agents — no local install. You sell Team seats; you carry zero hosting burden.

Changes (config + docs only — no code)

• railway.json — Railway service config (Dockerfile builder, /api/health healthcheck, ON_FAILURE restart). Railway reads it via the Deploy-to-Railway button.
• README — new "Host on Railway (team, no install for members)" section + a Deploy-to-Railway button pointing at railway.json, with the two required post-build steps called out: a persistent /data volume (so activated keys + memories survive redeploys) and ENGRAPHIS_FORWARDED_ALLOW_IPS=* (so the session cookie gets its Secure flag behind Railway's TLS proxy).
• docs/HOSTING_RAILWAY.md — 5-minute walk-through: deploy → /data volume → forwarded-allow-ips → custom domain + ENGRAPHIS_DASHBOARD_URL → bootstrap admin (license-exempt) → activate Team key (seat cap, server-validated) → invite members → connect agents (HTTP today; /mcp when feat(agent-connect): mount MCP-over-HTTP at /mcp on the dashboard #12 lands) → cost/limits/backups/security.

Why this shape

• Infra is one flat instance per team on the admin's Railway account (~$10–25/mo), amortized across seats — not per user. Team = $20/seat/mo; healthy margin, and you don't run hosting (uptime/backups/security stay with the admin).
• Members never install: they click the invite URL → log in (email/password, never license-gated) → mint a bearer token in Settings → paste into their agent. The instance's Team license is what unlocks /api/remember (402 without it) — that's the "a Team license is required to connect" gate.

Notes / caveats (called out in the guide)

• Railway volumes aren't auto-backed-up → guide recommends enabling volume backups or cron /api/export.
• Single-SQLite (WAL) store → fine for ~tens of concurrent agents, not hundreds; cap seat sales until a Postgres backend exists.
• Embedder (CPU all-MiniLM-L6-v2) is the cost lever; guide notes ENGRAPHIS_EMBED_MODEL → external embedding API for write-heavy teams.
• /mcp MCP-over-HTTP is referenced as "when feat(agent-connect): mount MCP-over-HTTP at /mcp on the dashboard #12 lands" (it's in the closed feat(agent-connect): mount MCP-over-HTTP at /mcp on the dashboard #12 / resubmit pending).

Verified

• railway.json is valid JSON; all new/edited files are LF (no CRLF pollution). No code changed (config + docs), so the suite is unaffected. Did not update PyPI (per your instruction). Not merging — opening for review.

[Coding-Dev-Tools]

Council Gate Verdict — APPROVE

Risk level: high | Agreement: 1.0 | Council session: council-e6ad8567-7bce-4842-a80b-a4535ad98d99

Per-model scores (rubric: correctness / safety / style / tests / complexity)

• nvidia/nemotron-3-ultra-550b-a55b: APPROVE (overall 4.8) -- correctness=5.0, safety=5.0, style=5.0, tests=4.0, complexity=5.0
• nvidia/llama-3.3-nemotron-super-49b-v1.5: APPROVE_WITH_NITS (overall 4.8) -- correctness=5.0, safety=4.0, style=5.0, tests=5.0, complexity=5.0

Engraphis reference

Council verdict persisted as mem_01KXJSQCC856WVGGGC5NEY0VYE (workspace Coding-Dev-Tools). Also logged to security-council workspace as mem_01KXJSQCCZ5XXM42SGTTBQBPJQ (auth/security PR).

Posted automatically by the council-gate PR review cron. Auth/security PRs are reviewed at risk_level=high.