# Remainder run for the wiki_dpr forced-seqscan baselines: ONLY the 50%
# filter cells.
#
# WHY A SUBSET CONFIG EXISTS: the full baselines_exact_real run of 2026-09-11
# 20:40 EDT (started at commit 29316bc, tie-aware recall) completed the
# unfiltered series and all 0.1%/1%/10% filter cells (22/28 points, every one
# recall==1.0, plan-verified seqscan_forced), then crashed at the first 50%
# cell at ~01:04: pg_pinecone-integration commits landed in this tree MID-RUN
# (00:34/00:39), and multiprocessing spawn workers re-import code from disk --
# the freshly spawned pool for the 50% sweep returned the new 5-tuple worker
# results to a parent still running the old 4-tuple unpack. The completed
# cells are preserved in
# results/baselines_exact_real_20260912T004017Z.crashed_partial.json (all
# measured consistently under 29316bc; the recall definition there -- tie-aware
# -- is identical to HEAD's). The mid-run-code-change guard added after this
# (runner.run's start/end content-hash stamps; save_results refuses on
# mismatch) makes this failure mode loud in the future.
#
# The composite record for the real-corpus exact baselines is therefore:
#   0.1%/1%/10% + unfiltered  -> the crashed_partial (29316bc)
#   50% cells                 -> this run's results file
# and results/README.md's summary table cites both provenance stamps.
#
# `dataset:`/`seed`/`warmup_queries`/`concurrency` are byte-identical to
# baselines_exact_real.yaml, so this run reuses the same cached artifact,
# k-means clustering, dataset key, and ground-truth caches. skip_unfiltered:
# the unfiltered exact series is already on record in the partial (~2h of
# 9-18s scans; nothing new to learn).
name: baselines_exact_real_r50
seed: 42
k: 10
warmup_queries: 20
hardness_recall_threshold: 0.97
min_throughput_seconds: 5.0

postgres:
  maintenance_work_mem: 8GB
  min_wal_size: 1GB
  max_wal_size: 4GB

dataset:
  kind: real
  artifact: wikidpr_768d_1m.npz
  n: 1000000
  dim: 768
  n_centers: 1000
  kmeans_sample: 100000
  kmeans_iters: 15
  n_queries: 500

metric: l2

filters:
  run:
    - {selectivity: 0.50, mode: correlated}
    - {selectivity: 0.50, mode: uncorrelated}

systems:
  - name: exact_seqscan
    type: exact
    sweep_param: none
    sweep_values: [0]
    skip_unfiltered: true

concurrency: [1, 8]
