Add OrcaReplay to Monitoring & Analytics

[xizhuomengcontin]

Adds OrcaReplay to Monitoring & Analytics, next to watchtower — which is the closest
neighbour in that section, and the reason I want to be precise about what this adds.

watchtower inspects the API traffic between an agent and its provider. OrcaReplay records that
traffic and then serves it back: run the same Codex command again and it reaches the same
result with no model called and no key needed.

$ orca record codex -- codex exec "fix the failing test"
info recorded run=run_8f21c3 events=41 exit=0

$ orca replay last
info replaying exchanges=6 egress=blocked
info replay.done reused=6/6 exact=6 divergences=0 exit=0

Why this is Codex-specific and not a generic LLM tool

Your rule says general AI/LLM tools don't belong without specific Codex integration, so here is
the Codex-specific part rather than a claim that it "supports Codex":

• It reads your config.toml. The adapter resolves the top-level model_provider key and
  redirects model_providers.<name>.base_url for the launched process only — so a Codex pointed
  at a non-OpenAI provider is captured at the origin it actually uses, not the default one. It
  honours CODEX_HOME.
• It reads Codex's own rollout. Sessions under $CODEX_HOME/sessions are parsed as JSONL, and
  the harness preamble (skills, environment context) is separated from what you actually typed, so
  the timeline shows your turns rather than Codex's scaffolding.
• Codex is one of the adapters in the project's integration test suite: it records against a stub
  origin, kills the origin, replays, and asserts the exchange counts.

One limit, stated up front

egress=blocked means model-provider egress, not network isolation. Replay serves the model's
answers from the trace, but the recorded tool calls still execute for real — a shell command that
ran curl runs it again. Replay is not a sandbox; the entries in your Docker & Sandboxing section
are the tools for that.

Apache-2.0. Happy to move it to Cross-Agent Tools instead if you think it sits better there — the
same recorder handles Claude Code, goose, opencode and others, so either section is defensible;
Monitoring & Analytics seemed closer to what a reader would be looking for.