v0.8.2: expanded model evals and dashboard refresh

[0.8.2] — 2026-07-28

A maintenance and eval-publication release. Forge hardens malformed llama.cpp 500 recovery, adds compatibility with clients that use llama-server’s unversioned chat endpoint, and publishes an expanded evaluation dashboard covering 378,300 runs across 291 configurations.

Added

• POST /chat/completions proxy alias. The proxy now serves llama-server’s unversioned endpoint as an alias of /v1/chat/completions, restoring compatibility with clients such as pi-llama-cpp.
• Expanded large-model eval tiers. The published roster gains Gemma-4 large models, new Qwen3.6 replay sweeps, LFM2.5 and Mellum2 replay ladders, and a new 120B tier. The consolidated v0.8.2 dataset contains 74,100 generation-3 runs across 57 complete arms.
• Reasoning effort as an eval dimension. Batch resume identity, result rows, report deduplication, and display labels now distinguish effort variants of the same model without conflating their results.

Changed

• Evaluation dashboard regenerated across all published datasets. The dashboard now represents 378,300 runs and 291 model/backend configurations, with normalized family names for the new models.
• Replay-aware generation deduplication. Explicit replay-policy rows now supersede only older rows for the same policy, allowing independently measured none, keep-last, and full arms to carry forward across generations.

Fixed

• Malformed llama.cpp 500 rescue is stricter and more robust. Rescue now ignores incomplete tool-call skeletons, recognizes structural and truncated tool-call XML without depending on one backend error phrase, and prevents raw backend bodies from leaking into user-visible error details.