feat(supabase): add findings_* tables for paper-figure summaries (formula-code/fc-eval#19)

[atharvas]

Summary

Adds the database half of formula-code/fc-eval#19: eight public-read tables backing the seven paper-figure summaries the website (formula-code/formula-code.github.io) currently ships as static JSON under `src/data/findings/f{1..7}_*.json`.

Once this lands and the exporter populates rows, the website's `findings.js` shim can swap from static JSON to live PostgREST fetches at `https://api.formulacode.org/rest/v1/findings_*\` — no API gateway changes required (same anon-key surface as `repositories` / `pull_requests` / `harbor_runs`).

Tables added

Table	Paper artifact
`findings_global_leaderboard`	Table 1
`findings_stratified_advantage`	Figure 3
`findings_tag_advantage`	Table 2
`findings_repo_quintiles`	Table 3
`findings_cost_pareto`	Figure 4 / Table 10
`findings_workload_tradeoff`	Figure 5
`findings_temporal_generalization`	Table 4 / Figure 1
`findings_metadata`	sidecar (`_source`, `_generated_at`, `axis_metadata`, `row_count`)

Schema choices

• Column names match the website scaffolds verbatim — `rp_rank` (not `rp_rank_adv`), `advantage_norm` (not `normalized_advantage`), `speedup_geomean` (not `mean_speedup`), `worst_workload_speedup`, `cost_usd_per_task` + `advantage_weighted`. This means the website's `findings.js` can swap fetch sources with no field renaming.
• Per-task rows use `(owner, repo, issue_number)` — aligns with `pull_requests` PK and `harbor_runs` columns. `findings_workload_tradeoff` carries an FK to `pull_requests` so the website can join back for PR metadata in one PostgREST call.
• RLS + anon grants follow the existing pattern — enable RLS, `public_read` SELECT policy for `anon`, then explicit `GRANT SELECT` (since `00015_revoke_anon_select.sql` revoked the schema-wide default).
• Migration sequencing — branches off `main` (currently at 00015). #23 lands 00016/00017/00019/00020; this PR's 00021 sits cleanly on top after that merges.

Test plan

Local apply + RLS sanity:

```bash
make supabase-up
docker exec supabase_db_datasmith_new psql -U postgres -d postgres \
-f /dev/stdin < supabase/migrations/00021_findings_tables.sql

RLS gating is in place:

docker exec supabase_db_datasmith_new psql -U postgres -d postgres -c \
"SELECT tablename, policyname FROM pg_policies WHERE tablename LIKE 'findings_%';"

→ 8 rows, all 'public_read'

Anon SELECT returns 200 OK (empty before exporter runs):

curl -s -H "apikey: $SUPABASE_ANON_KEY" \
http://127.0.0.1:54321/rest/v1/findings_global_leaderboard?select=*
```

Schema verification against canonical contract: each table's columns are a 1:1 match for the website scaffolds at `formula-code/formula-code.github.io:src/data/findings/f{1..7}_*.json` (verified during plan phase).

Follow-up

• Exporter PR — `analysis/export_website_findings.py` exists locally in fc-eval and populates these tables via `--upsert-supabase`, but the `analysis/` tree there is currently uncommitted; it'll ship in a separate PR once the dependency tree (`nb_utils.py` / `task.py` / `ranking.py` / configs) is consolidated.
• Prod deploy — once this PR lands, the migration needs to be applied to the prod Supabase instance, then the exporter run with `--upsert-supabase` to populate initial rows.
• Refresh cadence — manual for v1; a scheduled refresh (GitHub Action / cron) is a follow-up.

🤖 Generated with Claude Code