Incremental index size observations, November 18, 2026 These are deterministic synthetic size/layout experiments, not timing benchmarks. Fedora 44, c7i.4xlarge VM, GCC 16.2.1, 8KB pages, assertions, default btree fillfactor 90, four copies of each key. Before fixes Both fixes Compression off Fixed-length path 8,118,272 5,414,912 5,562,368 Variable-length path 7,929,856 6,365,184 6,356,992 Fixed-length URL 3,153,920 2,670,592 2,867,200 All numbers are pg_relation_size() bytes after inserting 100,000 rows into an initially empty index. The source for these experiments was the pre-publication candidate on upstream 81b2179b492, before the November 18 rebase and amcheck diagnostic fix. I did not remeasure these sizes on the final publication commit. Independent experiments removed the decoded-page reserve rejection, changed rightmost split ranking, and combined both. Off controls did not change. At 12,000 rows, two repetitions matched every recorded metric. The combined change left variable-length paths one page larger than off, with no final prefix directories in that workload. Parent/rootdescend/heapallindexed checks and index-only versus heap value digests passed in all 144 observations. Index-only scan plans showed zero heap fetches. This includes a sweep of fillfactors 10, 90, and 100, descending order, shuffled order, and a single repeated key. At fillfactor 10 some indexes grew substantially; shuffled fixed paths grew one page. Honoring fillfactor is not a universal size improvement. Reproduction outline (use a disposable database): CREATE EXTENSION amcheck; CREATE TABLE bench (k text COLLATE "C") WITH (autovacuum_enabled = false); CREATE INDEX bench_idx ON bench (k) WITH (prefix_compression = on); INSERT INTO bench SELECT '/srv/objects/customer-0042/archive/' || lpad(n::text, 9, '0') || '/' || repeat('a', 96) || '.json' FROM generate_series(1, 100000) g CROSS JOIN LATERAL (SELECT (g - 1) / 4 AS n) s ORDER BY k, g; VACUUM (ANALYZE) bench; SELECT pg_relation_size('bench_idx'); SELECT bt_index_parent_check('bench_idx', true, true); For variable-length paths, replace repeat('a', 96) with: left(repeat(md5(n::text), 8), 1 + (n * 37 % 240)) For URLs, replace the whole key expression with: 'https://shop.example.test/catalog/products/' || lpad(n::text, 9, '0') || '?view=full' Repeat in a fresh table with prefix_compression = off for its control. The outline isolates the size measurement. The original harness also recorded per-page directory, posting-list, density, and TID counters, and compared ordered key digests through independent access paths. It used additional heap columns, but indexed only k for these shapes. These observations do not establish query speedups, zero default-off cost, or behavior under concurrent or cold-cache workloads. Historical metal timing results came from an older candidate and are not current qualification of this patch.