Corpus: exclude own org + repo-level pair-hunting mode

[stackedsax]

What

Two changes to the corpus scraper (scripts/corpus/fetch_corpus.py):

1. Exclude our own org from all scrapes

Every mode now skips repos owned by InsightSoftmax (EXCLUDED_OWNERS), so we
stop re-ingesting BAMMM's own conversions/ and examples back into the corpus
(self-contamination). Applies to both the per-scheduler scrape and the new pairs
mode.

2. New pairs mode — repo-level, multi-format detection

The single-scheduler corpora are ~99% disjoint by repo (only 5 of ~319 repos
appear in 2+ corpora, all docs/example repos), so a PZ-style semantic join across
them for cross-scheduler equivalent pairs would be an expensive near-no-op.

Instead, hunt for repos that themselves contain multiple scheduler formats:

• Phase 1 — scan every scheduler's existing searches at repo granularity
  (search only, no content fetch): build repo -> {scheduler: [files]}.
• Phase 2 — keep repos spanning >= --min-schedulers formats (default 2).
• Phase 3 — fetch those files, run each scheduler's existing accept()
  quality predicate, and save co-located specs under
  testdata/corpus/candidate-pairs/<owner__repo>/<scheduler>/ with a committed
  manifest.json.

Co-located specs from different schedulers are candidate ground-truth conversion
pairs (build-plan stage 2). Specs stay gitignored (third-party licensing); only
the manifest is committed, consistent with the existing corpus policy.

uv run scripts/corpus/fetch_corpus.py pairs
uv run scripts/corpus/fetch_corpus.py pairs --min-schedulers 2 --search-limit 200
make corpus SCHED=pairs

Testing

• --list shows the new pairs mode.
• Offline test with stubbed GitHub API verified: multi-format grouping, own-org
  exclusion (a 2-format InsightSoftmax/* repo is dropped), single-format repos
  dropped, and correct on-disk layout + manifest.
• git check-ignore confirms candidate-pair specs are ignored and the manifest
  is tracked.

🤖 Generated with Claude Code