v0.8.3: proxy model routing and deterministic eval publication

[0.8.3] — 2026-08-03

A proxy-correctness and evaluation-publication release. Forge restores model identity handling across vLLM and Anthropic proxy paths while making evaluation collection and publication deterministic and provenance-safe.

Added

• Deterministic Hugging Face dataset publication tooling. A new offline tests.eval.dataset_builder command produces verified Parquet bundles with latest, snapshot, and history views, a fixed schema, source provenance, content hashes, and an upload-ready Dataset Card. Builds are verified and published atomically, with PyArrow available through the dedicated dataset-builder extra. #136

Changed

• Eval-generation provenance is explicit and resume-safe. batch_eval --generation stamps every collected row with its comparability epoch. Existing JSONL outputs are validated before collection begins, rejecting malformed data, mixed or mismatched generations, and ambiguous resume identities. Shared generation and replay-policy semantics now govern collection, reporting, and publication. #135

Fixed

• Unpinned external vLLM proxies discover their served model before listing it. The first GET /v1/models can use the caller's credential to complete lazy discovery, adopt and latch the backend's served identity, and return Forge's existing one-entry model list instead of the temporary default placeholder. Explicit pins and non-vLLM backends remain unchanged. #100, #137
• Anthropic proxy requests preserve their request-local model identity. Unpinned Anthropic backends now use each request's opaque model name instead of the fabricated claude default; an explicit --model remains authoritative, and a request with neither source fails with HTTP 400 before dispatch. Concurrent request models remain isolated, clean Anthropic passthrough preserves fields such as cache_control without mutating the caller-owned request body, and /v1/models returns an empty list when no proxy-wide model is pinned.