feat(server): OpenAI-compatible serve path for forged agents

[writeameer]

Add forge-agent: an executable that serves forge Agents behind a local OpenAI-compatible endpoint, so an existing host CLI (Codex, Grok Build) can point at it via OPENAI_BASE_URL and run "warm from prompt 1" — the Engine owns the loop and the agent's scaffold is the system prompt.

• server/: http.Handler implementing GET /v1/models, POST /v1/chat/completions, and POST /v1/responses (Codex speaks the Responses API, not Chat Completions), both with SSE streaming. Translates OpenAI <-> forge messages.
• Request-body decompression (zstd/gzip) via Content-Encoding + magic-byte sniffing — Codex ships the body zstd-compressed, which was the real failure.
• Default-agent fallback: host GUIs send their own model ids (e.g. "gpt-5.5") that aren't agent names; unknown models resolve to a configured default agent instead of 404.
• Request logging for live debuggability of host-CLI integration.
• cmd/forge-agent: wires the north-star demo agents (vanilla_reviewer vs forged_reviewer) and serves them.
• scripts/: serve.ps1 (launch) + demo.sh (curl driver).
• Tests cover models/chat/responses, streaming, content-array + string inputs, zstd-encoded bodies, and the default-agent fallback.

Validated live end-to-end through the real Codex CLI and GUI.