… in load_tasks, compute ndcg5 in cache mode

…dowing, asserts, redundant slice

…n up common.py

…y functions

Add one-liner docstrings to all functions and methods across
benchmarks/common.py, run_benchmark.py, and sync_repos.py. Remove the D
ruff ignore for benchmarks/*.py so docstrings are enforced going forward.
Also moves count_indexed_targets into run_benchmark.py (where Chunk is
imported) to fix a pre-existing mypy Protocol error in the pre-commit env.

Full runs (no --repo/--language filters) automatically write results to
benchmarks/results/<sha>.json, keyed by the 12-char git SHA. The file
includes the full SHA, model name, per-repo rows, language aggregates,
and overall summary. Cache mode writes <sha>-cache.json. Filtered runs
are not saved.

Drop the --cache mode (cold vs warm build timing) — it was noisy and
not actionable. Instead, add index_ms to RepoResult so every full run
records index build time per repo alongside NDCG and query latency.
index_ms is included in the saved JSON and printed in the summary table.

…, inline _output

- benchmarks/common.py -> benchmarks/data.py (more descriptive name)
- BENCH_ROOT: /tmp/bench -> ~/.cache/semble-bench (survives reboots)
- Inline _output into _check_repo (single call site)
- Update README to drop --cache docs and reflect new paths