Numeric and date ranges on both engines: bleve folds them from doc values, OpenSearch uses range and date_range aggregations. The proto gains BucketRange and bucketDefinition.ranges, the aggregation package the range parser (bounds are numbers or RFC3339 dates, one kind per aggregation, anything else is rejected), the kind of an option and the validation of ranges against the field type. Every requested range is answered in request order, an empty one with a count of zero, with no space answering too; size does not cut ranges, keyAsNumber sorts them by their lower bound. Range keys read from..to. AGG-04, 05 and 07 to 09 in the parity suite.
Term buckets on both engines: bleve folds them from doc values (its facets read numbers as prefix-coded terms), OpenSearch uses terms aggregations. A request gets every bucket, up to 65535 per space (the OpenSearch default for search.max_buckets, held on bleve as well); beyond that it is refused. The proto gains AggregationOption, BucketDefinition (sort, minimum count), AggregationResult and Bucket. The new aggregation package merges the per-space results by position, so several aggregations on one field stay apart, and shapes them per BucketDefinition (minimum count, sort, size). It also validates aggregations against the index mapping (a field the index knows and a client may name, of a type whose values are buckets; the mapping overrides mark the internal fields): the search service relies on it, the graph endpoint asks it first to save the round trip, next to its own checks against the spec, and with the field names of the request so an error reads like the request. A request names a field like the driveItem property, exactly so. A number is keyed as it is, a bool as true or false on both engines, an empty value is no bucket. Pinned in the parity suite as AGG-01 to 03, 17, 34, 37 and 39; the paging rows AGG-41 to 43 pin the first commit through the same matrix machinery, which also shortens long answer lists in the README.
convert.copyFacet rebuilt every facet by hand on the way into the index; mapping.ToProto is the generic counterpart of the FromProto the graph endpoint uses, so both directions share one bridge.
MS Graph style search endpoint: searchRequest with KQL queryString, from/size pagination (values outside the spec's bounds are rejected), entityTypes validated to driveItem, hits as driveItems with webUrl, parentReference, remoteItem for shares, the facets and @libre.graph.permissions.actions.allowedValues, which the search service now projects onto the Entity from the same permission set as the WebDAV report. Properties the endpoint does not evaluate yet (aggregations, aggregationFilters, sortProperties) answer 501 instead of being ignored, an unknown $expand answers 400, and a failing search service keeps its status. Hits reuse what the drive item listing has, which moves a little: the thumbnails helper is named for the drive item it fills (setDriveItemThumbnailsByID, it served shares only), the web URL of an id gets a helper (webURLForID), and the facets come out of the proto through one generic mapping.FromProto instead of a copy per facet.
The search service takes a `from` offset next to the page size and applies it after the cross-space merge; every space answers the full prefix up to from+size, since an offset cannot be distributed. Pages are stable: both engines and the merge break score ties by id, and OpenSearch counts every match (track_total_hits). No engine pages beyond the first 10000 matches, the OpenSearch result window, so a from and size reaching beyond it are rejected on both engines. `page_size` becomes optional on the wire: absent is the default of 200, 0 asks for no matches, -1 for all; WebDAV and the tags listing set it accordingly, and the WebDAV report's `<oc:offset>`, parsed since the report exists and never read, now pages the merged list. The gRPC wrapper passes the request through and keys its cache by the whole request.
Spaces that are disabled while reindexing are skipped. They are now
remembered in a NATS KV bucket (including whether a forced rescan was
requested) and reindexed when the SpaceEnabled event comes in. The
entry is removed when the space gets deleted.
Base ISO is a Canon maker note tag holding Canon's raw encoded value, not the iso of the shot: other cameras got no iso, Canon photos a wrong one (256 for an ISO 800 shot). exif:IsoSpeedRatings is there on tika 3 and tika 4.
Tika 4 prefixes its metadata keys (mp4:, img:), so the live photo facet was never built there. Live photo support needs tika 4: tika 3 got the video keys only in 3.3.2 and never the still image time. The content key is read through one lookup, marked tika 4 and tika 3 legacy.
The byte-based truncation cut through the two-byte UTF-8 sequence of
Motörhead's ö, committing invalid UTF-8 into the parity README. Shared
rune-based ellipsize helper; matrix regenerated.
Dominik Schmidt