Evaluate shared reader code without regressing point lookups #797

Follow-up to #772. OnDiskReader<B> repeats read logic for file and memory backends even though lower tree/cache primitives already use dyn StorageBackend. A disposable prototype stores a Box<dyn StorageBackend> containing the existing MutexStorageBackend<B> and makes the reader itself non-generic.

At v3 baseline 3d19abbce5dd8d5244641ff0d7ba8bad5424d07f, Rust 1.99.1, the audit lock and unchanged profiles:

The cost is one extra allocation per reader plus backend trait-object dispatch. Native savings do not transfer directly to single-backend builds.

Runtime evidence is conditional, not a no-regression claim. Four socket-swapped pairs for each of 17 warm-read shapes at 100K facts preserved expected/equal row counts in all 68 comparisons. Most median p50 changes were within 2.3%, but the short point probe initially showed +5.08%. A focused 32-pair rerun of three short workloads gave median changes of +3.91% point, +0.84% attribute/value, and +2.02% as-of; individual short-run noise remains large. The fixture was in bench tmpfs, so this does not measure fsync or cold disk performance.

The prototype passed the two existing reader-unit tests, native/browser/WASI builds, wasm-bindgen packaging, and cargo +1.89.0 check --all-features --locked. That is not sufficient coverage to ship a reader refactor.

Deliver a small evaluated change only if its tradeoffs are acceptable:

• Compare this prototype with extracting shared non-generic read helpers, which may retain specialization without another allocation.
• Resolve the point-lookup slowdown with longer controlled runs and measure reader creation, cold/cache-disabled reads, concurrent queries, cursor reuse and browser/WASI costs.
• Preserve lock lifetimes, cache ownership, error ordering, UUID batching, temporal visibility, snapshot pinning and all public API/format behavior.
• Run the full relevant storage/query/transaction/fault/concurrency suites, platform and coverage gates on the bench before an implementation PR is accepted.
• Keep the existing size cap, coverage floors, MSRV and portability. Reject the change if reliable performance is compromised for the size saving.