Rejected

Jun 8, 2026 · Rejected · parameter-encoding-and-binding

Experiment 142: Single-row text parameter direct encoding

Date: 2026-06-08

Status: Rejected

Direction:parameter-encoding-and-binding

Benchmark Run: Tracelite A/B retest, exp-142-tracelite-single-row-text

Problem

PR #130 proposed applying the exp 125 / exp 126 direct text payload packing

pattern to the single-row allocateParams path. That path is used by

parameterized reads and single-row writes that do not go through the wide-batch

matrix encoder.

The old focused harness showed strong local wins on small string-heavy shapes,

but it was produced by the pre-Tracelite experiment loop. The reassessment

question was whether the change still looked merge-worthy when retested against

current origin/main with the integrated Tracelite baseline/candidate workflow

and peer guardrails.

Hypothesis

Removing the temporary List<Uint8List?> plus per-string Uint8List allocation

from allocateParams should improve standardized workloads that include

single-row string parameter binding, or at least stay neutral.

Accept only if Tracelite clears the primary resqlite improvement gate on the

closest current lanes and sqlite_async guardrails remain clean. Reject or

defer if the primary lanes are neutral/noisy or trend slower, because the change

adds more custom string-encoding logic to a hot binding path.

Approach

Created two resqlite worktrees from current origin/main at

1979ec3c4069ffa960df465971b23e5e53323768:

Baseline: /Users/dan/.codex/worktrees/retest-exp142-baseline
Candidate: /Users/dan/.codex/worktrees/retest-exp142-candidate

The candidate reapplied only the code and focused harness from PR #130:

lib/src/native/resqlite_bindings.dart
benchmark/experiments/single_row_param_encoding.dart

Candidate patch shape:

pass 1 measures string UTF-8 byte length without allocating temporary encoded

byte lists,

ASCII-only parameter rows use a direct codeUnitAt copy into the existing

native parameter buffer,

mixed/non-ASCII rows use the shared _writeUtf8 encoder from the batch path,
int, double, blob, null, wide-batch, narrow-batch, and public API behavior are

unchanged.

Before the A/B run, both retest worktrees resolved dependencies with ARM64 Dart.

The candidate then passed:

 /Users/dan/Coding/flutter_arm64/bin/cache/dart-sdk/bin/dart analyze \ lib/src/native/resqlite_bindings.dart \ benchmark/experiments/single_row_param_encoding.dart /Users/dan/Coding/flutter_arm64/bin/cache/dart-sdk/bin/dart test \ test/database_test.dart test/transaction_test.dart

Ran the integrated Tracelite A/B workflow with pinned Tracelite

a2bf3648836fcf680d0aceccb18c2b31a2109586 and ARM64 Dart:

 /Users/dan/Coding/flutter_arm64/bin/cache/dart-sdk/bin/dart run \ benchmark/run_tracelite_experiment.dart \ --dart=/Users/dan/Coding/flutter_arm64/bin/cache/dart-sdk/bin/dart \ --tracelite-root=/Users/dan/Coding/tracelite \ --baseline-root=/Users/dan/.codex/worktrees/retest-exp142-baseline \ --candidate-root=/Users/dan/.codex/worktrees/retest-exp142-candidate \ --label=exp-142-tracelite-single-row-text \ --direction=parameter-encoding-and-binding \ --suite-scenarios=chat-sim,narrow-batch-insert \ --policy-scenarios=chat-sim,narrow-batch-insert \ --interfaces=sqlite_async,resqlite \ --guardrail-peers=sqlite_async \ --runs=2 \ --min-repetitions=5 \ --max-repetitions=12 \ --out-dir=build/tracelite-experiments/exp-142-tracelite-single-row-text

Artifacts:

build/tracelite-experiments/exp-142-tracelite-single-row-text/resqlite-tracelite-experiment.json
build/tracelite-experiments/exp-142-tracelite-single-row-text/baseline/history.json
build/tracelite-experiments/exp-142-tracelite-single-row-text/candidate/history.json
build/tracelite-experiments/exp-142-tracelite-single-row-text/decision/decision.json
build/tracelite-experiments/exp-142-tracelite-single-row-text/decision/insights.md

Results

The wrapper collected both sides and wrote the decision artifacts:

step	status
baseline suite history	ok
candidate suite history	ok
decision artifact	inconclusive

Tracelite decision policy:

field	value
expectation	improvement
primary threshold	38.0%
max guardrail regression	28.5%
max CV	28.5%

Primary comparisons:

role	scenario	peer	metric	baseline	candidate	change	max CV	p	status	effect
primary	`chat-sim`	`resqlite`	`measured_elapsed_ns`	15.7 ms	16.7 ms	+6.86%	24.2%	0.716	neutral	inconclusive
primary	`narrow-batch-insert`	`resqlite`	`measured_elapsed_ns`	11.8 ms	13.7 ms	+16.4%	27.5%	0.057	neutral	inconclusive

Guardrail comparisons:

role	scenario	peer	metric	baseline	candidate	change	max CV	p	status	effect
guardrail	`chat-sim`	`sqlite_async`	`measured_elapsed_ns`	27.4 ms	27.8 ms	+1.41%	20.3%	0.460	neutral	pass
guardrail	`narrow-batch-insert`	`sqlite_async`	`measured_elapsed_ns`	12.1 ms	11.6 ms	-3.68%	15.3%	0.800	neutral	pass

Decision insights:

severity	finding	detail
warning	Decision is inconclusive	Evidence is not strong enough for a production decision.
warning	Primary metric did not clear	`chat-sim` changed +6.86% with max CV 24.2%, p=0.716, 95% delta CI -1.02 ms..3.17 ms.
warning	Primary metric did not clear	`narrow-batch-insert` changed +16.4% with max CV 27.5%, p=0.057, 95% delta CI -575 us..4.44 ms.

The baseline and candidate explain artifacts also warned that these short

lanes are partly harness-dominated. That limits how much we can infer about the

exact micro-cost, but it does not create positive evidence for carrying more

custom binding code: the closest standardized lanes did not improve and both

resqlite primary rows moved slower.

The decision graph-data export validated and produced:

dataset	rows
`scenario_series`	2240
`peer_summary`	16
`decision_summary`	1
`decision_comparisons`	4

Decision

Reject. Do not carry the allocateParams direct text encoding code from PR

#130.

The old focused harness remains a useful clue that single-row text parameter

encoding can be made faster in isolation, but the current Tracelite retest did

not convert that clue into production evidence. On the closest available

standard lanes, the candidate was neutral/inconclusive and directionally slower:

+6.86% on chat-sim and +16.4% on narrow-batch-insert.

This reinforces exp 146's parameter-encoding lesson: small and narrow binding

changes should stay on the generic path unless a current workload shows

parameter encoding is material and a Tracelite A/B decision clears the primary

gate. If a future app profile shows high-frequency single-row string parameter

binding as a real bottleneck, add a Tracelite scenario or profile lane that

directly covers that shape before retrying this implementation.

Future Notes

The current Tracelite scenario set does not perfectly isolate the original

focused harness shape from PR #130. That is a reason to avoid overclaiming a

runtime regression, not a reason to merge the code. A future retry should first

add or select a scenario where single-row string parameter encoding is a

material fraction of measured time, then rerun the integrated A/B workflow.

Validation

/Users/dan/Coding/flutter_arm64/bin/cache/dart-sdk/bin/dart pub get in

both retest worktrees

/Users/dan/Coding/flutter_arm64/bin/cache/dart-sdk/bin/dart analyze lib/src/native/resqlite_bindings.dart benchmark/experiments/single_row_param_encoding.dart
/Users/dan/Coding/flutter_arm64/bin/cache/dart-sdk/bin/dart test test/database_test.dart test/transaction_test.dart
/Users/dan/Coding/flutter_arm64/bin/cache/dart-sdk/bin/dart run benchmark/run_tracelite_experiment.dart ...