ConvoMem (Salesforce, Apache-2.0, ~75K QA pairs) ships as pre-mixed
test cases — self-contained conversation haystacks plus questions —
that map 1:1 onto the grouped runner mode.

- datasets/convomem.py: the full dataset is multi-GB (one size-300
  batch is ~850MB), so fetching is selective. Batch files within each
  <category>/<N>_evidence/ dir are ordered by case context size, and an
  HTTP Range tail-probe (last 4KB) reads each file's final contextSize
  without downloading it. Only files matching the requested sizes are
  fetched; index.json records every probe result including files NOT
  downloaded, so the selection itself is auditable. Probes are
  throttled with retries (HF CDN resets rapid bursts).
- converters/convomem_to_corpus.py: stratified deterministic sampling
  by (category, contextSize) with a fixed seed; sampling.json records
  seed, per-stratum population, and sample counts so a published number
  states exactly which slice of ConvoMem it covers. Leakage scrub:
  containsEvidence/model_name dropped, conversation ids remapped to
  neutral positional ids. Ground truth maps evidence conversation ids
  through the remap; abstention evidence referencing absent
  conversations yields empty ground truth by design.
- CLI: datasets fetch --dataset convomem --context-sizes; convert
  convomem --sample-per-stratum/--seed/--context-sizes. justfile
  recipes + README section.

Live-verified against the real dataset (user_evidence, size 10):
50 files probed, exactly 2 downloaded, 3 cases sampled -> 30 docs /
30 queries (size-10 cases pack 10 questions per haystack, each
targeting a distinct evidence conversation), all ground truth non-empty
and remapped. 7 new unit tests; suite green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Drew Cain <groksrc@gmail.com>