Resolve #23: expose benchmark_codes and standardize task identity on (owner, repo, issue_number)

[atharvas]

Closes #23.

Summary

Lands everything npm run sync needs from api.formulacode.org, and folds in a cleanup the issue's discovery surfaced: task_id is constructed two incompatible ways across the codebase (owner__repo-N in the publish/HF path vs owner_repo_N in harbor_adapter), with neither matching the format the website assumed. Per the user directive, the canonical row identity is now (owner, repo, issue_number) everywhere, and task_id becomes a single-column integer mirror of issue_number — kept for joins, no longer a derived string.

What the website now gets (issue acceptance criteria)

Ask	Where it lands
benchmark_codes public table	supabase/migrations/00017_benchmark_codes.sql + new stage 9
Queryable task_id on candidate_containers	supabase/migrations/00019_candidate_containers_task_id.sql (generated column = issue_number)
Level-aggregation statistic documented	CLAUDE.md "Level aggregation" subsection — geometric mean, 4 levels, with code citation
Optional benchmark_type	supabase/migrations/00020_benchmark_information_type.sql (generated, derived from ASV time_*/timeraw_*/mem_*/peakmem_*/track_*)

New pipeline stage 9 scrape_benchmark_source

Walks candidate_containers, uses the existing prepare_repo_checkout to materialize each repo at its SHA, then AST-parses every benchmarks/*.py (or whatever asv.conf.json declares as benchmark_dir) to extract per-function source + colocated setup / setup_cache. Idempotent upsert into benchmark_codes. Tunables: DATASMITH_BENCH_SCRAPE_MAX_FILE_BYTES, DATASMITH_BENCH_SCRAPE_DIRS.

task_id collapse — touched files

github/models.py, publish/records.py, publish/huggingface.py, publish/pipeline.py, harbor_adapter/{records,adapter,utils}.py, runners/harbor_healthcheck.py, harbor_adapter/template/{setup.sh,test.sh,upload.py,parser.py}. Harbor's directory layout uses a derived task_dir_name = f"{owner}__{repo}__{issue_number}" (still derived but contained to the adapter — Harbor's flat-dataset discovery requires it). All tests updated and passing.

One-shot migrations for Harbor's Supabase (NOT in supabase/migrations/)

Harbor's tasks table and snapshot bucket live in a different Supabase project — see HARBOR_SUPABASE_URL. Two scripts handle the cross-project migration:

• scripts/harbor_tasks_migration.sql — adds owner/repo/issue_number, backfills from legacy task_id (handles both Format A and Format B), swaps the PK to the triple, retains task_id as a deprecated integer mirror.
• scripts/migrate_snapshot_keys.py — walks Supabase Storage and moves snapshots/{old}/oracle.tar.gz → snapshots/{owner}/{repo}/{issue_number}/oracle.tar.gz. Defaults to --dry-run.

Apply post-merge, in that order.

Breaking changes

• Harbor tasks PK is now (owner, repo, issue_number). Any external consumer keying on legacy task_id strings must read from the new columns. The task_id column survives as an integer mirror so reads using select=task_id continue to work, just typed differently.
• Snapshot tarballs move under a deeper prefix; the script above renames existing ones.
• Existing built Harbor containers still ship the old TASK_ID-only template — they'll write to the new path correctly but query the old PK shape. Rebuild on next harbor_healthcheck run.

Level-aggregation statistic (the website's open question)

Geometric mean. Code at src/datasmith/harbor_adapter/template/parser.py:158-204. Four levels: per-benchmark, per-class, per-module, overall. The website's CSV historically encodes five levels; datasmith's benchmark_information.benchmark_name is already param-stripped, so the website's 1-Params row is identical to datasmith's level1. Full description is in CLAUDE.md under "Level aggregation."

Test plan

• make check clean (ruff, mypy, deptry)
• uv run pytest tests/scrape tests/publish tests/github — 108 passing, including new AST-extraction tests
• Migrations 00016 / 00017 / 00019 / 00020 applied locally; generated columns verified (35,152 benchmark_information rows → time)
• Reviewer should: run scripts/harbor_tasks_migration.sql --dry-run against Harbor's Supabase and confirm the regex backfill parses every legacy task_id cleanly
• Reviewer should: run scripts/migrate_snapshot_keys.py (dry run) and confirm the planned moves match the live tasks count
• Reviewer should: fc-data --stage 9 --start-date ... smoke run; confirm benchmark_codes populates and one row's source matches the upstream file

🤖 Generated with Claude Code