Media indexing

Media indexing reads file metadata and retains capture time, embedded location, source identity, and technical properties. The original file remains the source for downloads and content conversion.

Original contents are hashed as a stream for common source issues. Content failures keep account acknowledgements across unchanged retries. Empty files finish as informational skips and retain any earlier imported records. Import and metadata changes remain transactional; unexpected service failures keep operational recovery separately from source acceptance.

GPS and media have separate scan/watch policies; their shipped defaults and manual rescan actions are in background processing. The format registry defines recognized image and video extensions. Metadata extraction and content decoding still depend on the format and packaged tools.

Index changes enqueue correlation work. Resized previews and video streams are requested separately from metadata indexing. Content freshness and source revisions prevent obsolete content from becoming current results.

The shared GPS/media file indexer commits each directory discovery in an independent transaction. Failed writes remain for the next scan. Watcher events coalesce on a separate discovery thread. Every source is committed before import dispatch. SCHEDULED is the durable discovered/unprocessed queue; PROCESSING is committed before the consumer starts. Discovery does not wait for imports. Progress counts include the full committed backlog, and the UI shows discovery separately while its total is still growing.

Directory scans compare at most 256 source identities at a time using narrow database projections. Full index entities are loaded only for state changes. Scans retain counts rather than complete changed-file or missing-file lists. After discovery commits, pending work is read by increasing file ID in batches of 128; each batch finishes before the next is dispatched. The database retains the full backlog while scan memory stays bounded. Missing sources are checked in a separate paged pass after a complete directory walk.

Work is serialized per path through completion persistence; a stored generation rejects superseded callbacks. Discovering another revision during active work persists RESCHEDULED without replacing that consumer's source snapshot. The start callback preserves this status; the older completion returns it to SCHEDULED for the next revision instead of marking the file completed. lastSuccessfulGeneration is a processing attempt number, not a timestamp; it allows an unchanged restoration to cancel removal before cleanup starts. File attributes are refreshed when work starts. Consumers check the source revision before committing and roll back and queue the latest revision when it changes. Successful completion follows the domain-data commit; independent completion writes retry failures, and restart/rescan safely replays orphaned work. Committed discovery survives a failed import or a lost dispatch notification.

Missing-source cleanup requires an accessible source root and confirmed absence. Incomplete walks do not dispatch removals. Failed cleanup remains visible as failed work; REMOVED records successful cleanup. Original files are read only.

Location projection, time correction, and activity correlation are distinct stored concerns. A manually assigned location does not rewrite EXIF.

See correlation, selection, and delivery. The media runtime owns native packages and their checks.