v0.1.26

What's Changed

• Measure the unmeasured: a grammar property test, the three defects it found, and the Linux sandbox CI can no longer skip by @farukciftler in #1
• fix(docs): restore the README phrase the skip-count test pins by @farukciftler in #2
• fix(cli): a build that cannot run a model no longer answers as if it could by @farukciftler in #3
• feat(eval): the selection suite says how far along it is by @farukciftler in #4
• perf(grammar): the mask stopped re-deriving a constant 147k times a step by @farukciftler in #5
• fix: a tool call could take 13 minutes — the whitespace cycle, the calendar deadline, and the trace that found them by @farukciftler in #6
• fix: three ways one runaway got out — the text ceiling and the tool-call cap by @farukciftler in #7
• fix(tools,grammar): the shape only Turkish wrote, and the root cause that stayed shut by @farukciftler in #8
• Linux and Windows measured on real machines — what CI could not see by @farukciftler in #9
• Two measurements from a rented GPU: what --compare never checked, and why parallel eval wins nothing by @farukciftler in #10
• The multiplier a bigger GPU can actually give: a batched-decode measurement by @farukciftler in #11
• The model the binary downloads was not the model the numbers describe by @farukciftler in #12
• --compare read a missing tool as a broken model by @farukciftler in #13
• README: six minutes on a GPU, and where the multiplier actually is by @farukciftler in #14
• The three failures the model set actually named — and a README that had drifted by @farukciftler in #15
• A benchmark protocol, 608 cases in seven languages, and four models measured by @farukciftler in #16
• The llama architecture, two task benchmarks, and the web as a cassette by @farukciftler in #17
• Distilling a tool-caller: 0% to 61.5% correct calls, and what it cost by @farukciftler in #18

New Contributors

• @farukciftler made their first contribution in #1

Full Changelog: v0.1.25...v0.1.26