Stop → restart → resume → verify

This repository demo runs entirely locally. It uses six fake customers, a JSONL receipt store and an idempotent file update. It does not connect to your production database.

git clone https://github.com/8dazo/captor.git
cd captor
pnpm install
pnpm --filter @captar/types build
pnpm --filter @captar/config build
pnpm --filter @captar/utils build
pnpm demo:backfill

The runner launches separate Node processes for three phases:

1. Stop: reserve two writes per batch, commit four writes, checkpoint offset 4, then block the third batch at the four-write ceiling.
2. Resume: reopen the store in a new process, skip the first four items and process the last two with a new two-write budget.
3. Verify: independently read application state, assert all six customers are present, and confirm the failed and succeeded receipts.

Finally the runner executes captor runs and captor inspect against the same receipt file. Each invocation creates a unique .captor/recovery-demo-* directory and prints its path.

This demonstrates a controlled stop and restart, not exactly-once delivery across a process crash. There is a gap between an application side effect and its checkpoint save. Interrupted batches may be replayed. The demo's update is idempotent; a real integration needs equivalent protection and stable input ordering.

See backfill boundaries before using this pattern on production data.