feat(hf): daily snapshot publisher to Hugging Face dataset

[Flotapponnier]

What

Publishes a daily Parquet snapshot of every public OCB benchmark to
huggingface.co/datasets/OpenChainBench/benchmarks.

Goal: backlink Tier 1 (HF dataset pages indexed by Google + scraped by
ChatGPT/Claude/Perplexity), plus a queryable archive for journalists
and academic users.

How it works

Daily cron at 00:00 UTC runs scripts/hf_publisher/publish.py:

1. Fetch /api/citable and /api/stat/<slug> from prod
2. Quorum guard: refuse to publish if <50% benches are live or count <8
3. Build 3 Hive-partitioned Parquet tables:
   • headlines/ (26 rows/day, ~15KB): 1 row per (slug, day) with leader + value
   • providers/ (254 rows/day, ~20KB): 1 row per (slug, provider, day) with p50/p90/p99
   • timeseries/ (1838 rows/day, ~15KB): 1 row per 24h sparkline point
4. Stage README, CITATION.cff, LICENSE, JSON schemas, examples (pandas/polars/duckdb)
5. Push to HF via huggingface_hub
6. Slack alert on abort/crash

Total ~50KB per snapshot, ~18MB/year. Compressed zstd.

Validation

Dry-run executed against live prod:

```
quorum ok: 26/26 live
wrote headlines part-0.parquet rows=26 size=15KB
wrote providers part-0.parquet rows=254 size=20KB (26 leaders)
wrote timeseries part-0.parquet rows=1838 size=15KB
```

Spot-checked content: aggregator-head-lag leader=Mobula 681s, bridge-quote-latency leader=Mobula 342ms, bridge-fee leader=Near Intents, all coherent with /api/citable.

Tests: 15/15 pass offline (quorum guard, row builders, partitioning, template substitution).

Durability features

• Schema versioning: `schema_version` int column in every row, additive-only contract documented in README
• Quorum guard: bad source feed = no publish, previous snapshot stays as truth
• Idempotent: same snapshot_date overwrites itself
• Tests gate every PR: schema regression caught in CI before HF push
• Slack alerts: `SLACK_WEBHOOK_URL` secret pinged on abort/crash
• HF git history: each snapshot = 1 commit, rollback via `git revert` on HF

Required GitHub secrets

• `HF_TOKEN` (write scope on dataset repo)
• `SLACK_WEBHOOK_URL` (optional, for ops alerts)

Test plan

After merge to dev:

• Manually trigger via Actions tab with `dry_run: true` to validate the env wires up
• Re-trigger with `dry_run: false` to ship the first real snapshot
• Verify dataset page at huggingface.co/datasets/OpenChainBench/benchmarks shows preview
• Let cron run once at 00:00 UTC tomorrow, verify a second partition lands
• After 7 days, merge dev -> main (the workflow file lives in CI, not in the runtime path, so the merge is trivial)