client_credentials (4/4): rate-limit token issuance (defense-in-depth) #163

Part 4 of 4 for #159 — req 5 of the issue's security requirements (defense-in-depth). Does not block the grant shipping (part 3), but gates closing #159.

Scope

Rate-limit client_credentials token issuance. The plugin has no rate-limiting machinery today, so this is new infrastructure — keep it minimal:

• Per-node, in-memory token bucket keyed by client_id (documented as per-node: Harper replication makes a shared counter table a hot-write anti-pattern, and the assertion replay guard + 60s exp already bound abuse).
• Configurable limit (e.g. mcp.clientCredentials.rateLimit requests/min, sensible default), 429 + error: "slow_down"-style OAuth error on exceed.
• Scope decision to settle in this issue: client_credentials-only, or also cover the existing authorization_code/refresh grants while we're here.

Tests

Under-limit requests pass; over-limit returns 429 with an OAuth error body; bucket refills; per-client isolation (one noisy client doesn't starve another).

Dependencies

Part 3 (#162 — the grant must exist to be limited).