perf: defer transcript search snippets and retain reader scope (#201566)

* perf: remove transcript search fan-out bottlenecks

Defer SQLite FTS snippets until the filtered result cursor consumes hits,
retain filtered readers in the existing worker read scope, and bound only
transcript search to six CPU-limited isolates. Keep generic reader caps and
writer ownership unchanged.

Synthetic 100,000-message fixtures reduce median solo search time from
5.27 s to 0.74 s and six-call completion from 14.00 s to 9.75 s on a shared
host; maximum lane queue wait falls from 12.07 s to 0.11 ms. Document the
memory budget and shared-host timing tradeoff. Replace redundant test proof
with concurrency and filtered-reader lifecycle regressions at zero net test
line growth.

* perf: keep transcript search within the existing reader budget

Keep the measured lazy-snippet and reader-scope fixes while restoring the
original two-worker lane. Synthetic six-call batches finish in about 2.8s
with two workers versus 6.8s with six, using less memory.

Replace the removed six-worker policy test with a native SQL work-count
regression: four consumed candidates evaluate four snippets instead of
160. Preserve cold writer reentry and independent reader lane coverage
through a shared fixture, with net-negative test growth.

Remove the manual benchmark from the maintained tree; retain its timing
split and two-worker ablation in PR #201566.