exp 168: reject Database resolved-runtime cache

[danReynolds]

Hypothesis

Exp 159 attacked fixed per-round-trip overhead inside Writer._request by
caching _sendPort and removing the await _workerPort.future microtask
hop. Every public Database method (select, selectBytes, execute,
executeBatch, transaction) still does final ... = await _runtime; at
the top — one microtask hop above the writer/reader call, on every Database
method call, because _runtime is already-resolved after open.

If exp 159's residual is really port-wake + event-loop scheduling, the
same pattern at the layer above should chip another hop off every call. At
~1-2 µs per hop the upper bound on sequential-awaited (2000 writes) is
~2-4 ms out of a ~32 ms median: theoretical ~6-12%, comfortably above the
focused-benchmark noise band exp 159 used to accept its win.

Approach

lib/src/database.dart:

• New nullable field _DatabaseRuntime? _resolvedRuntime.
• Inside the existing Future.sync(() async { ... }) body, after spawn
  awaits complete, assign the runtime record to _resolvedRuntime before
  return runtime.
• Hot paths replace final ... = await _runtime; with
  final ... = _resolvedRuntime ?? await _runtime; at 5 call sites.
• close and diagnostics keep await _runtime — cold.
• stream keeps Stream.fromFuture(_runtime).asyncExpand(...) — one-shot
  per stream creation.

No public API change. No new measurement scaffolding.

Results

dart run benchmark/experiments/writer_pipelining.dart, two passes, 7
rounds each, order flipped between passes per the JOURNAL "phase-ordered
A/B" lesson.

pass	shape	baseline	candidate	delta
1	sequential-awaited (2000)	32.169 ms	31.425 ms	−2.3%
1	concurrent-burst (10×200)	24.465 ms	25.818 ms	+5.5%
1	transaction-guardrail (50×10)	4.587 ms	4.244 ms	−7.5%
2	sequential-awaited (2000)	31.794 ms	32.603 ms	+2.5%
2	concurrent-burst (10×200)	24.089 ms	25.138 ms	+4.4%
2	transaction-guardrail (50×10)	4.093 ms	4.342 ms	+6.1%

Sequential-awaited and transaction-guardrail flip sign between passes at
the same magnitude as round-to-round variance. Concurrent-burst trends
small-positive (regression) both passes but inside the per-round spread.

dart analyze lib: clean. dart test test/database_test.dart test/stream_test.dart: 75/75 pass.

Outcome

Rejected — below the focused-benchmark noise floor across two
order-flipped passes. The theoretical 6–12% headroom did not show up in
measured wall time; deltas alternate sign between passes at the same
magnitude as single-phase variance. Same shape as the recent
overhead-removal rejection cluster (exp 145, exp 148, exp 151): once
per-call cost drops below ~2 µs the focused harness can't separate it
from machine jitter.

No runtime code kept. Exp 159's "further reduction needs request batching
across calls (group commit) or a different transport" framing remains the
correct steering for the sequential-write floor. Would reopen if a future
measurement system can resolve sub-1 µs per-call deltas on real
workloads.

Test plan

• dart analyze lib — clean
• dart test test/database_test.dart — 49/49 pass
• dart test test/stream_test.dart test/stream_invalidation_coalescing_test.dart test/stream_cache_hit_reliability_test.dart — pass
• dart test test/stream_dependency_shapes_test.dart test/stream_overflow_fallback_test.dart test/stream_trigger_cascade_test.dart — pass
• dart run benchmark/finalize_experiment.dart --experiment=experiments/168-resolved-runtime-cache.md — green

🤖 Generated with Claude Code