Function capture_projection 

pub async fn capture_projection<S: BuildHasher>(
    log: Arc<dyn MetadataEventLog>,
    topics: &HashMap<Uuid, (String, i32), S>,
    captured_at_ms: u64,
) -> Result<DisklessWalCapture, String>

Replay the committed keyed index state up to each partition’s high watermark and freeze it.

The capture only reads. Each partition’s cutoff is its high watermark (ListOffsets LATEST) when the capture starts, and the capture reads from the partition’s low watermark up to that cutoff. It applies each record as it arrives rather than buffer the range, so an index larger than memory still captures. An empty partition is complete at once. Replay fences that brokers publish into the index are not index state, so the capture skips them. Kafka refuses a Produce to an internal topic from any client but __admin_client, so a capture that wrote its own fence would fail against a Kafka-compatible broker.

Errors

Returns an error when the index watermarks cannot be read, when the read fails, when the watermark lookups and the read together take longer than 30 seconds, or when a record cannot be decoded or projected.