Add the resident AI detective

[YesterdaysLemon]

What changed

• add a same-origin, stateless OpenAI Responses route backed by gpt-5.6-luna
• run a seven-tool resident allowlist through the exact shared WebMCP board catalog
• preserve human approval for every model-authored string or group
• add per-case durable chat, compact consent, browser/model fallback states, and expressive terminal moods
• animate only clues touched by pending AI proposals and replace the ambiguous GHOST ribbon with a small AI provenance stamp
• harden public inference with attachment stripping, strict schemas, origin and size checks, network-bound rate limits, bounded tool rounds, safe logging, and deterministic fallback
• retire the unused OpenAI Sites configuration/plugin and keep the VPS delivery artifact/container coherent

Verification

pm test — 39 passing

px tsc --noEmit

pm audit --omit=dev --audit-level=high — 0 vulnerabilities

pm run build

• exact Docker image production smoke (health, HEAD, origin-trial header/meta)
• real gpt-5.6-luna inference smoke
• browser flow: consent, hosted chat, �udit_evidence → propose_connection, visible proposal, reload persistence, unsupported-WebMCP fallback copy, no console errors

Merge hold

Do not merge until the root-owned VPS app config mounts a runtime env file containing OPENAI_API_KEY. The code itself remains fully functional through the local deterministic fallback if the secret is absent.

Closes #15