NEWS

geosam 0.1.2

sam_explore() gains a Detection Quality control (Fast / Balanced / Accurate) tucked under an "Advanced" disclosure. The default is "balanced" (chunked with moderate overlap) and can be overridden with the new quality argument. Fast mode skips chunking entirely for a more responsive feel on small areas; Accurate uses smaller chunks with more overlap.
Esri satellite view in sam_explore() and sam_view() reorders the OpenFreeMap overlay so road geometry sits beneath the satellite and only symbol/label layers ride above. Road name labels are restyled (light text, dark halo) for legibility on dark imagery.
When Fast mode runs multiple text prompts on the same view, vision embeddings are computed once and reused across prompts (via a new detect_text_multi() Python helper modeled on HuggingFace's multi-prompt SAM3 pattern).
A text-embedding cache (LRU bounded to 64 entries) is available for experimental opt-in reuse across detect_text() calls with the same prompt. Disabled by default — the cached path passes pooler_output where the model documents a different shape and may produce different detections than the combined-processor call. The chunked detection path remains on the unchanged combined-processor code path.

Chunked detection now overlaps neighboring chunks and keeps detections whose centroids fall inside the chunk's core region. This reduces missed detections and duplicate detections at tile boundaries.
sam_detect() exposes chunk_size and chunk_overlap for tuning.
sam_filter() now filters chunked results (sf_result) — previously a no-op when only the merged sf was present.

Tiles downloaded from Mapbox/Esri/MapTiler are now cached on disk under rappdirs::user_cache_dir("geosam")/tiles/.... geosam_clear_cache() removes them. Cache keys for paid providers use a hash of the API token rather than a fragment of the token itself.
get_imagery() warns when missing tiles are filled with blank pixels.

geosam_status() distinguishes Sam3Model from Sam3TrackerModel availability and (with check_access = TRUE) verifies that the configured HuggingFace token can reach the gated facebook/sam3 metadata. Failures are classified (missing_token, gated_not_accepted, invalid_token, network_error).
Installation and documentation updated to reflect that SAM3 weights are gated — users must accept the model's access terms in addition to logging in.

References to a "google" imagery source corrected to "maptiler" in docs and arg checks.
First test suite added (chunked filter + interactive quality policy).