…RANDOM, lazy Cesium)

Four architectural moves, each gated on ?v=2. v1 stays unchanged.

1. Primary read surface: samples_map_lite.parquet (60 MB) instead of
   wide.parquet (278 MB). The lite file has every column the Explorer
   needs for the list + globe view except description.

2. No ORDER BY RANDOM(). v1 uses RANDOM() which forces a scan across
   row groups; v2 uses bare LIMIT, accepting row-order bias in exchange
   for ~20× query speedup on columnar parquet. (Trade-off acceptable
   for a viz sample; revisit if source clustering becomes visible.)

3. Lazy description fetch. v2 drops description from sampleData and
   adds a lazyDescription cell that queries wide.parquet for just the
   one pid when a sample is clicked. sampleCard falls back to
   lazyDescription when s.description is empty.

4. Lazy Cesium mount. v2 returns null from the viewer cell until
   viewMode === 'globe', so the viewer constructor (~500 ms) doesn't
   run for users who stay in list/table view. v1 mounts eagerly.

whereClause handles column-name drift (v1 uses `n`, v2 uses `source`)
and skips the otype filter for v2 (lite is samples-only). Text search
in v2 is limited to label + place_name (description isn't loaded
eagerly). v1 keeps description search.

Next: measure v2 and compare against the PR isamplesorg#124 baseline.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…RANDOM, lazy Cesium)

Four architectural moves, each gated on ?v=2. v1 stays unchanged.

1. Primary read surface: samples_map_lite.parquet (60 MB) instead of
   wide.parquet (278 MB). The lite file has every column the Explorer
   needs for the list + globe view except description.

2. No ORDER BY RANDOM(). v1 uses RANDOM() which forces a scan across
   row groups; v2 uses bare LIMIT, accepting row-order bias in exchange
   for ~20× query speedup on columnar parquet. (Trade-off acceptable
   for a viz sample; revisit if source clustering becomes visible.)

3. Lazy description fetch. v2 drops description from sampleData and
   adds a lazyDescription cell that queries wide.parquet for just the
   one pid when a sample is clicked. sampleCard falls back to
   lazyDescription when s.description is empty.

4. Lazy Cesium mount. v2 returns null from the viewer cell until
   viewMode === 'globe', so the viewer constructor (~500 ms) doesn't
   run for users who stay in list/table view. v1 mounts eagerly.

whereClause handles column-name drift (v1 uses `n`, v2 uses `source`)
and skips the otype filter for v2 (lite is samples-only). Text search
in v2 is limited to label + place_name (description isn't loaded
eagerly). v1 keeps description search.

Next: measure v2 and compare against the PR isamplesorg#124 baseline.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…RANDOM, lazy Cesium) (#126)

Four architectural moves, each gated on ?v=2. v1 stays unchanged.

1. Primary read surface: samples_map_lite.parquet (60 MB) instead of
   wide.parquet (278 MB). The lite file has every column the Explorer
   needs for the list + globe view except description.

2. No ORDER BY RANDOM(). v1 uses RANDOM() which forces a scan across
   row groups; v2 uses bare LIMIT, accepting row-order bias in exchange
   for ~20× query speedup on columnar parquet. (Trade-off acceptable
   for a viz sample; revisit if source clustering becomes visible.)

3. Lazy description fetch. v2 drops description from sampleData and
   adds a lazyDescription cell that queries wide.parquet for just the
   one pid when a sample is clicked. sampleCard falls back to
   lazyDescription when s.description is empty.

4. Lazy Cesium mount. v2 returns null from the viewer cell until
   viewMode === 'globe', so the viewer constructor (~500 ms) doesn't
   run for users who stay in list/table view. v1 mounts eagerly.

whereClause handles column-name drift (v1 uses `n`, v2 uses `source`)
and skips the otype filter for v2 (lite is samples-only). Text search
in v2 is limited to label + place_name (description isn't loaded
eagerly). v1 keeps description search.

Next: measure v2 and compare against the PR #124 baseline.

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>