fix: use the correct index max_local threshold, not table's

[iheitlager]

Summary

• 006-btree Req 7 documented a real coverage gap: no fixture exercised an overflowing index key. Adding one (overflow_index_key.db, an ~8000-byte indexed TEXT key against a 4096-byte page) immediately hit PayloadTooShort — the gap was hiding an actual bug, not just missing coverage.
• local_payload_size computed max_local as usable_size - 35 unconditionally, but SQLite defines a smaller max_local for index cells (leaf AND interior) than for table leaf cells: (usable_size - 12) * 64 / 255 - 23 vs usable_size - 35. Every index cell whose payload landed between the two thresholds was read with a local_size far larger than what SQLite actually reserved on the page.
• Fixing the threshold surfaced a second latent bug: index entry delete (src/btree/index/delete.rs) never freed a removed entry's overflow chain at all (table delete already does). Index entries essentially never overflowed under the old, too-generous threshold, so the gap was never exercised. Added free_overflow_chain, wired into both delete_from_leaf and the interior-match outright-delete path (remove_entry_by_child).

Test plan

• New corpus test overflowing_index_key_reassembles_byte_identical_to_oracle against the new fixture
• New regression test deleting_an_entry_with_overflow_frees_its_overflow_chain
• Pre-existing deleting_all_entries_orphans_no_page (which started failing once the threshold was corrected, exposing the orphaned-overflow-page bug) now passes
• cargo test --lib (925 tests), full corpus suite (373 tests, including fixture-regeneration reproducibility), clippy, make assurance (006-btree 18/18, no dead links) all clean

spend: ~2x the trivial fixture-generation estimate — the fixture work surfaced two real correctness bugs (wrong max_local, leaked index overflow chains on delete) that needed fixing, not just a missing fixture.

🤖 Generated with Claude Code