Keep shared-cache rollout disabled for now

The current cache slows small and medium warm reads. Large files benefit. A benchmark-only SQLite change identifies a substantial part of the overhead.

16 and 256 KiB: raw source cache with repeated transformation and gzip. 2 MiB: existing transformed-file cache. Two warm batches per setting: 32 serial requests or 64 requests at concurrency eight. CPU and memory include both server and benchmark client.

• Correctness: 2,700 HTTP responses passed byte checks.
• Admin scaling: one pool statistics query reached 78.7 ms p95 at 100,000 synthetic blob records. It blocks the event loop.
• Eviction: the 4 MiB limit held, but cycling over a 6.25 MiB working set caused 300 misses in 300 reads.
• Next steps: reduce metadata commits, cache admin aggregates, and test a representative mixed workload across server processes.

NORMAL was tested only on temporary cache indexes. In WAL mode it can lose recent commits after power loss, so recovery needs validation before adoption. SQLite durability documentation.

Local compiled-code evaluation · Node 24.14.1 · 2 November 2026. Shared host, synthetic content, warm filesystem cache, no TLS or production traffic. These measurements are not production capacity guarantees. Application settings remain unchanged.