docs(rfc): pooled HTTP Playwright MCP server for generated claws

[BoldBlackBot]

Summary

• Adds a Proposed RFC: an opt-in pooled browser backend for generated claws — a long-lived Playwright MCP server in HTTP mode (--port + --isolated) started by the container boot command, before the exec hermes gateway.
• One chromium launch, one BrowserContext per MCP client: bounds browser memory for multi-client workloads (~(N−1) × 1.0 GiB saved for N clients) and removes cross-client tab/session collisions (upstream microsoft/playwright-mcp#893 failure mode).
• Daemon lifetime is tied to the container: gateway death = container exit = the whole PID namespace (daemon + browser tree) is reaped. No orphaned browser processes are possible; recreate deployments bounce daemon + gateway atomically.
• Opt-in via stack parameter EnablePooledPlaywright (default false) — default-path output stays byte-identical, golden test regenerates only for the enabled path.

Design highlights

• Boot-command daemon with restart loop, bound to loopback (--host 127.0.0.1), version-pinned server (no @latest), chromium installed to the persistent volume via the pinned server's own install-browser.
• Client entry in agent_home/config.yaml (type: streamable-http, url: http://127.0.0.1:8931/mcp) pushed by the existing manage overlay; RFC documents the stack-parameter-before-overlay ordering and rollback.
• Memory accounting section: pooling pays only when concurrent MCP clients > 1; single-client claws are better served by idle-recycle lifecycle settings (out of scope here).

Test plan

• RFC review — no code changes in this PR
• Implementation lands as a follow-up per the RFC checklist (template.yaml parameter + command, config entry, skill phases, README, golden regen)

RFC: rfcs/2026-08-16_pooled-http-playwright-mcp.md