v1.2.0 — Cloudflare Worker edge + all 13 providers

Added

• One package, three deployment shapes. The same noveum-ai-gateway crate
  now builds as the native binary / Docker image, a Rust library, and a
  Cloudflare Worker (wasm32, true per‑PoP edge). The Nova Guard engine,
  pricing, and request/response shaping are shared so all shapes behave
  identically. See docs/CLOUDFLARE_WORKER.md and
  docs/CLOUDFLARE_DEPLOYMENT.md.
• All 13 providers on the edge Worker: the OpenAI‑compatible set (OpenAI,
  Groq, Together, Fireworks, Mistral, Cohere, Gemini, DeepSeek, xAI, OpenRouter,
  Perplexity) plus Anthropic (/v1/messages + x-api-key, Messages→OpenAI
  conversion) and Bedrock (Converse API, AWS SigV4 signed in pure Rust;
  also accepts temporary credentials via x-aws-session-token, which the
  native path does not).
• Worker Nova Guard parity: input block + redact, output‑phase block + redact,
  NOVEUM_GUARD_BLOCK_RESPONSE_MODE (synthetic_success/provider_error),
  permissive CORS + OPTIONS preflight, header/query pass‑through, SSE
  streaming pass‑through, and an 8 MB response inspection cap.
• New shared modules routing (provider routing + request/response transforms)
  and sigv4 (pure‑Rust AWS SigV4; unit‑tested against RFC 4231).

Changed

• Native Anthropic now passes 4xx/5xx error bodies through unchanged instead
  of converting them to an empty chat.completion.
• Output‑phase enforcement picks the response shape from the actual body (so
  Anthropic responses already converted to OpenAI shape are still inspected).
• Output redaction now rewrites every choice / content block (was first‑only).
• jsonschema built without its HTTP/file $ref resolvers (wasm‑safe; no
  behavior change for inline schemas).

Live edge: https://noveum-ai-gateway.shashank-1a4.workers.dev
crates.io: https://crates.io/crates/noveum-ai-gateway/1.2.0

🤖 Generated with Claude Code