fix: Make WebSocket event ingest idempotent

[Adarsh9977]

Summary

• Ignore duplicate event_id inserts during event ingest so SDK retry frames do not close the WebSocket.
• Add a WebSocket retry regression test that sends the same paired event twice and verifies the connection still receives verdicts.

Test plan

• go test ./internal/ws

Checklist

• CLA signed (see CLA.md)
• Tests pass locally
• Docs updated where needed
• British English; no em-dashes; no marketing fluff

[shahtajkhalid]

Thanks @Adarsh9977 for the clean fix, merging it.

One follow-up: the path is idempotent for the events row but not end to end. On a retry the classifier re-runs and writes a second row for the same event_id, so replays quietly duplicate verdicts and repeat inference. If you want to take it, have InsertEvent report whether it actually inserted (RowsAffected() == 0 means duplicate), and on a duplicate skip classify plus the verdict insert and re-publish the existing verdict. Same tests should pass.

Also, while you're at it: ON CONFLICT (id) DO NOTHING is more precise than INSERT OR IGNORE, which also swallows NOT NULL and FK violations rather than just the duplicate key.

Merging now, follow-up is welcome but not required.

[Adarsh9977]

Thanks @Adarsh9977 for the clean fix, merging it.

One follow-up: the path is idempotent for the events row but not end to end. On a retry the classifier re-runs and writes a second row for the same event_id, so replays quietly duplicate verdicts and repeat inference. If you want to take it, have InsertEvent report whether it actually inserted (RowsAffected() == 0 means duplicate), and on a duplicate skip classify plus the verdict insert and re-publish the existing verdict. Same tests should pass.

Also, while you're at it: ON CONFLICT (id) DO NOTHING is more precise than INSERT OR IGNORE, which also swallows NOT NULL and FK violations rather than just the duplicate key.

Merging now, follow-up is welcome but not required.

Thanks @shahtajkhalid, I'll update it and open a PR shortly