Create gateway token group

Per-group guardrail policy. Missing actions default to ignore. profiles may be listed with both dlp actions ignore (staged, inactive) but a non-ignore dlp action requires at least one profile. streaming defaults to buffered when any response action is block and passthrough otherwise; passthrough with a response block is 400. Findings are codes and counts only; content is never stored. The stored form is echoed back canonicalised (all actions explicit, profiles sorted); {} is stored as an all-ignore policy, null disables.

Per-group same-route retries. max_attempts counts the first try of each step (the primary and each fallback) and is the ceiling for the x-ltg-max-attempts request header. A step is retried only when its attempt definitely did no provider work (connection failure before dispatch, or 429/529) before the first provider event; an uncertain failure (timeout, 408/500/502/503/504) moves to the next fallback step instead. max_attempts x (1 + the longest fallback chain) must not exceed 8, else 400. The n-th retry waits retry_delay_ms (constant), retry_delay_msn (linear) or retry_delay_ms2^(n-1) (exponential), at most 5 s and never past the request deadline, then jittered uniformly down to half that value so concurrent retries spread out; a provider Retry-After of at most 5 s is honoured and a longer one moves to the next step. request_timeout_ms applies to streaming requests only: it bounds each attempt's time to its first provider event and never applies after it. Non-stream requests keep the per-attempt hop and idle bounds (their headers arrive only with the finished completion). retry_delay_ms and backoff are floors for the request knobs. Each planned attempt is reserved (at most 8 per request); unused reservation is released at settlement. Managed routes only. The stored form is echoed back with defaults filled in; null disables retries.