Improve range planning and immutable write scheduling

CHANGELOG.md

@@ -6,6 +6,11 @@ The format follows Keep a Changelog and the package uses semantic versioning.
 
 ## Unreleased
 
+- Applied every supported lower and upper bound from a conjunctive range query
+  before loading secondary-index candidates.
+- Bounded parallel secondary-index uploads to three operations, overlapped
+  independent immutable Snapshot or Trie writes, and kept HEAD as the final
+  awaited compare-and-swap.
 - Started the mutable-HEAD freshness interval when a successful fetch or
   conditional revalidation completes, preventing slow 304 responses from
   arriving already expired.

docs/ARCHITECTURE.md

@@ -110,9 +110,11 @@ private.
    gzip-compressed when useful, and encrypted.
 5. Every configured secondary index and declared covering projection is
    updated or rebuilt.
-6. New immutable document and index objects are created.
-7. HEAD publishes the document root and all active index references with one
-   ETag compare-and-swap.
+6. New immutable document and index objects are created. Secondary-index
+   uploads use a concurrency bound of three and overlap independent Snapshot
+   or Trie object writes.
+7. After every immutable upload completes, HEAD publishes the document root
+   and all active index references with one ETag compare-and-swap.
 8. The response includes the new HEAD, changed immutable objects, and
    document.
 9. The writing tab updates its cache and broadcasts the bundle to other tabs.

docs/BENCHMARKS.md

@@ -183,10 +183,13 @@ The range returned 25 documents from covering index fields.
 | Medium | 473.98 ms | 1,612.96 ms | 510.99 ms | 1,919.19 ms | 1,500 |
 | Large | 1,044.55 ms | 2,115.30 ms | 1,026.00 ms | 2,018.80 ms | 7,500 |
 
-Current range planning selected every value above the lower bound before
-applying the upper bound, so the 25-document large result evaluated 7,500
-index candidates. Covering fields still kept the operation to two network
-reads and avoided loading document pages.
+At the benchmarked commit, range planning selected every value above the
+lower bound before applying the upper bound, so the 25-document large result
+evaluated 7,500 index candidates. The current planner applies both bounds and
+the equivalent regression test selects 25 candidates. The latency table
+remains the historical measurement and has not been relabelled as a
+post-correction benchmark. Covering fields kept the measured operation to two
+network reads and avoided loading document pages.
 
 ## Full scans
 

docs/QUERIES-INDEXES.md

@@ -189,7 +189,9 @@ defineIndex<Note>(
 ```
 
 Range indexes contain exactly one field. Equality indexes can contain up to
-four fields.
+four fields. For conjunctive range queries, every supported `eq`, `lt`, `lte`,
+`gt`, and `gte` comparison on the indexed field is applied before document
+candidates are loaded.
 
 Only scalar string, number, boolean, or null values are indexed. Arrays and
 objects remain available to bounded local filtering.