Coalesce cold-eye prefetch and bound foreground fetch latency

[philsippl]

Summary

Fixes the cold-eye tail latency exposed in prod on 2026-08-26 (batch 201, 21:50:51Z): a broad-matching uniqueness request discovered ~3.7k second-eye candidates and took 22.8 s of compute, of which ~15.5 s were the two discovered-candidate stages — almost entirely un-metered foreground Postgres reads (two server-side slow-statement alerts bracket exactly those gaps in the log timeline).

Root-cause chain and fixes:

1. Prefetch starvation. Discovery enqueues one tiny Fetch command per full-scan chunk; thousands of commands overflowed the 64-deep queue (queue_skips) and the surviving ones became 1–2-row round trips, so almost nothing was actually prefetched. The prefetch worker now drains and merges everything queued into one reservation pass and one batched read per database round trip, and the queue holds 4096 commands. Barrier ordering is preserved: a barrier drained mid-merge completes after the read that covers every fetch enqueued before it.
2. Foreground fetch latency. Candidates that missed prefetch stalled the scan on a single id = ANY(...) statement over ~144 MiB of TOASTed rows. Large cold fetches — foreground and prefetch alike — are now split into 512-row sub-batches issued concurrently (parallelism 4) via a shared helper.
3. Observability gap. The foreground read had no metric, which is why dashboards showed "DB time is short" while Postgres logged multi-second statements. Added linear_scan_cold_foreground_db_duration / _batch_size / _records_total, plus linear_scan_cold_prefetch_coalesced_records on the worker.

Additionally the cold-eye LFU capacity (default and stage/prod manifests) goes from 4096 to 12288 records (~460 MiB at ~38.4 KiB/record — well inside the deployment's memory headroom) so recurring broad-matcher candidates stay resident across requests instead of being refetched.

Validation

• cargo clippy --all-targets --features db_dependent -- -D warnings
• iris-mpc-cpu lib suite: 396 passed.
• db-backed cold-eye tests incl. new cold_eye_coalesces_prefetch_commands_and_batches_large_fetches (1200 rows; 700 prefetched via ~234 small commands; single fetch across both sub-batch paths; byte-exact round-trip).
• e2e_linear_scan_test, e2e_linear_scan_uniqueness_test (release).

Expected effect on the incident profile: the ~3.7k-candidate reads coalesce into a handful of concurrent batched statements that largely complete during the 5.6 s resident scan, and any residual foreground wait becomes visible on its own dashboard series.

Stack

Based on #2351 (cold-eye cache layer); #2348 should merge this branch after review.