v0.2.0

Ferrite v0.2.0

Added

• A shared read-only GGUF mapping API that lets validated quantized tensors
  retain file ranges without a model-sized heap copy.
• Context-only single-session and batched prefill paths for non-final prompt
  tokens whose output logits are not observable.
• Exact-prompt cohort fan-out through independent KV snapshots in the
  continuous scheduler.
• Eval schema version 3 with model SHA-256 records, complete ordered server
  token-ID traces, and whole-cohort parity checks.

Changed

• The CLI and server now load quantized weights from immutable mapped model
  storage while retaining owned loading for library callers.
• Continuous batching now uses a bounded five-millisecond admission window,
  batched context-only prompt evaluation, and generic equal-token-sequence
  prompt grouping.
• Batched Q8 output argmax reuses per-worker scratch storage instead of
  allocating a vector for every vocabulary row.
• Architecture, server, CLI, evaluation, performance, and release guidance now
  document mapped-file safety and the shared-prompt execution contract.

Performance

• Raised the repeated four-request Qwen2.5 0.5B Q4_K_M shared-prompt server
  median from 93.21 to 131.45 aggregate tokens per second, a 41.03% increase.
• Reached a repeated eight-request median of 159.58 aggregate tokens per second.
• Reduced four-request server peak RSS from 956.8 MiB to 568.8 MiB and retained
  CLI RSS from 1,005.1 MiB to 556.7 MiB.
• Reduced the four-request median time to first token from 880 ms to 183 ms.

Validation

• Verified every response in the accepted request cohorts against complete
  ordered token-ID traces and exact default-route parity.
• Passed strict all-feature Rust checks, default and all-feature tests, rustdoc,
  x86_64 Linux cross-checks, package verification, dependency policy, and
  reproducible release-tool gates.