Set the record store insert overwrite setting (allowOverwrite) to false for all non-clustered collections on attached storage, so a replicated-rid insert colliding with an existing key fails with WT_DUPLICATE_KEY instead of silently overwriting a different document.

Details

• Today we use allowOverwrite=false only for clustered and recordIdsReplicated collections and true for everything else (see getRecordStoreOptions).
• Under the set-wide design, a collection's effective rid-mode changes with node membership, so the setting cannot remain derived from a per-collection property.
• The setting is runtime-only (RecordStore cursor params), not baked into the WiredTiger table config, so the flip requires no on-disk migration and applies to existing collections.
• The blind-write optimization that motivates overwrite=true is disagg-only (AttachedPersistenceProvider::shouldUseBlindWriteWhenSafe returns false for ASC).

Performance

Measured with paired sys-perf patches (change vs comment-only baseline at the same commit): flat on all realistic insert workloads; residual 1-2.5% regression only under extreme single-collection insert contention (parallel_insert_replica, 32-1024 threads). See the allowOverwrite decision tab in the technical design.

Acceptance criteria

• All non-clustered collections on attached storage are created/opened with allowOverwrite=false.
• No behaviour change for Disagg.
• Paired sys-perf validation included.