[feature] add zeroentropy embeddings provider

[flamerged]

Summary

This adds first-class ZeroEntropy zembed-1 embeddings support to Hindsight.

zembed-1 is a state-of-the-art retrieval embedder, and this integration lets Hindsight use it natively instead of forcing users through an OpenAI-compatible shim or LiteLLM proxy. It also preserves ZeroEntropy's asymmetric retrieval semantics: retained memory content is embedded as document input, while recall/search text is embedded as query input.

Why this matters

• Better retrieval stack: Hindsight already supports ZeroEntropy reranking; this completes the native ZeroEntropy retrieval path with zembed-1 embeddings.
• No proxy tax: Users can point Hindsight directly at ZeroEntropy's /v1/models/embed endpoint, with no shim layer translating away provider-specific features.
• Correct asymmetric embeddings: ZeroEntropy exposes separate query/document modes. Native support lets Hindsight use those modes directly instead of treating all text the same.
• Production-friendly dimensions: Hindsight defaults zembed-1 to 1280 dimensions so it works with pgvector HNSW's 2000-dimension index limit out of the box, while still allowing 2560/1280/640/320/160/80/40 for deployments that want different tradeoffs.

What changed

• Added a native ZeroEntropyEmbeddings provider for zembed-1.
• Added HINDSIGHT_API_EMBEDDINGS_ZEROENTROPY_* config/env settings for API key, model, base URL, dimensions, encoding format, latency, and batch size.
• Routed retained content through encode_documents() and recall/search text through encode_query() when the embeddings backend supports asymmetric modes.
• Added support for both float and base64 ZeroEntropy embedding responses, normalizing both to float vectors before storage.
• Fixed the existing ZeroEntropy reranker factory to honor its configured base URL.
• Updated env examples, docs, and generated docs-skill references.
• Added focused coverage for config parsing, provider factory wiring, request payloads, response decoding, query/document routing, and reranker base URL wiring.

Validation

• uv run pytest tests/test_zeroentropy_embeddings.py -q -> 12 passed
• uv run ruff check hindsight_api tests/test_zeroentropy_embeddings.py -> passed
• uv run ty check hindsight_api/ -> passed
• ./scripts/hooks/lint.sh -> passed
• CodeRabbit CLI review final pass -> 0 findings
• Live ZeroEntropy API smoke test: document and query embeddings returned 1280-dimensional float vectors
• Live retain/recall smoke test through MemoryEngine against a scratch pgvector database: retained 1 memory unit and recalled 1 result with embedding_dim=1280

[nicoloboschi]

LGTM. Solid integration — Pydantic-typed request/response, asymmetric query/document split with a safe default on the base Embeddings class so existing providers don't need touching, and base64 response decoding handled. Bonus catch on the reranker base_url not being threaded through the factory.

Cleanups (encoding_format fallback dead branch, duplicated dimension validation, unused usage field, getattr-vs-Protocol mismatch in embedding_utils, aligning URL handling with the existing ZeroEntropyCrossEncoder) will land in a follow-up PR.