Operator backfill for L2 dataset decoders (PoX-4 / sBTC / BNS)

[ryanwaits]

Context

Phase 2 datasets (PoX-4, sBTC, BNS — tracked in C1-C3) each get a new L2 decoder + dedicated tables. The default behavior is forward-only: decoders run from their checkpoint onward, dataset shows "data from cursor X forward."

Per the May 6 planning conversation, this is acceptable for v0 launch — it lets us ship the datasets faster without a heavy historical backfill pass for each.

What's deferred

A one-shot operator job that walks each new decoder backward from block 0 and populates the historical rows in its dedicated table. Mirrors the existing genesis-backfill pattern documented in docker/docs/OPERATIONS.md §8 for the parquet publishers.

For each new decoder:

• A --from-block / --to-block CLI mode (or a flag on the existing decoder consume loop)
• A wrapper script that iterates 10K-block windows from genesis up to the decoder's current checkpoint
• Idempotency: re-running an already-processed window is a no-op or fails by default
• Documentation alongside the publisher backfill in OPERATIONS.md §8

When to run

After a Foundation grant + customer demand makes "the full historical archive in dataset form" a real ask. Until then, the forward-only path is fine — customers who need historical data hit /v1/streams/events directly.

Out of scope

• Automatic backfill on first run (deliberate — the operator decides cost).
• Parallelization across windows (sequential is fine; the indexer DB is the bottleneck, not the decoder).