Kindex F26 benchmark build

This tag marks the exact code behind the F26 benchmark report (docs/benchmarks/kin-ask-f26-report.html) and the Labs post on Kindex.

It is not a PyPI release. Its package metadata still reads 0.48.1. Nothing is published from this tag.

Which version to use

• Default settings (LoCoMo 95.13%, LongMemEval-S 97.80%, BEAM-1M 80.69%): the default answer path behind these scores shipped unchanged in Kindex 0.49.0. Later releases move it to ask.preset: benchmark (#93). Pass each question's date with --as-of.

• BEAM with the dataset's annotations:

  Split	Score
  100K	91.66
  500K	89.57
  1M	90.61
  10M	84.15

  These runs need this build. They pass each question's annotations through caller hints: answer_question(hints=..., hint_queries=...) with ask.hint_mode: lead_both. That interface was removed before 0.49.0 (91616af).

  • The answer model was gpt-5.6-luna with reasoning off.
  • The shared answer settings and per-category overrides are in #78.