Jul 16, 2026 · Rejected · result-transfer-shape

Experiment 232: Reject exact quarter-step REAL JSON fast path

Date: 2026-07-16

Status: Rejected

Direction:result-transfer-shape

Benchmark Run: none — focused

The exact tested harness and runtime are preserved at

archive/exp-232,

with order-flipped measurements recorded in

benchmark/results/2026-07-16T11-07-07Z-exp232-dyadic-real-fastpath.md.

No release-suite run was used because no release lane isolates exact

quarter-step REAL formatting.

Archive:archive/exp-232

Problem

Exp 194 removed snprintf("%.17g") from

selectBytes() for REAL cells that are exactly integral. Its future note was

deliberately stricter about fractional work: pursue another specialization only

if a real production profile shows fractional REAL formatting dominates.

This run skipped that prerequisite. It reasoned that exact .25, .5, and

.75 values occur in domains such as ratings and prices, then moved directly

to an all-quarter focused benchmark. That made the mechanism easy to isolate,

but it did not establish how often the value lattice occurs in representative

selectBytes() traffic.

Hypothesis

For a finite, non-integral REAL value with abs(value) < 1e15, if value * 4

is exactly integral and not divisible by four, the historical

snprintf("%.17g") spelling can be emitted by reusing the native i64 formatter

and appending .25, .5, or .75.

The original acceptance gate required at least 20% improvement on synthetic

10k × 8 and 10k × 20 quarter-step lanes in both orderings, at least 5% on a

synthetic 50%-quarter mixed row, and integral/general-fractional controls inside

±3%. That gate could prove a fast formatter. It was insufficient to prove that

permanent specialization of the generic REAL path was worth shipping because

it omitted activation prevalence, expected aggregate impact, and a persistent

complexity budget.

Approach

The archived prototype placed exp 194's zero and exact-integral checks first,

then:

1. multiplied other REAL values by four and admitted exact non-integral quarter

units inside a strict abs(value) < 1e15 bound;

1. called an out-of-line fixed-point writer that reused the i64 formatter and

appended the exact suffix;

1. sent every miss to the historical snprintf("%.17g") fallback.

Correctness required special handling for negative sub-unit values, the strict

magnitude boundary, a portable no-inline wrapper, and two exported native test

hooks, including one exposing production-path admission. The candidate added

114 lines across native/ and the build hook, plus quarter-specific tests and

benchmark lanes. The behavior was byte-identical, but the maintenance surface

was permanent and every non-integral REAL miss still evaluated the new

recognizer.

The focused harness used 100%-quarter target lanes and a mixed row with four

quarter cells out of eight. Those rows measure the ceiling cleanly; they are not

representative-incidence evidence. During final review, the candidate was

reassessed against the product-value question rather than only its preset

mechanism gate.

Results

Medians in microseconds per query; full raw tables and shared-host control

traces are in the linked result file.

The quarter formatter itself is roughly 4.5× to 7.8× faster. Correctness

coverage also proved exact byte identity across dense signed quarters,

bound-adjacent magnitudes, and 100k deterministic random quarter values.

That is not the same as product value. All three summaries of the non-quarter

fractional control lean candidate-slower: +1.35%, +1.95%, and +0.65% for the

repeat-sequence median. They remain below the noise floor, but the direction is

mechanistically plausible because every miss performs extra classification.

Using the 20-column target/control deltas as a rough linear estimate, exact

quarters would need to comprise about 1.6% to 2.6% of non-integral REAL

cells merely to offset that miss tax. No production or representative

distribution established even that incidence, and no evidence showed

fractional REAL formatting materially affects end-to-end application wall time.

Decision

Rejected. Keep exp 194's exact-integral specialization and remove the

quarter-step runtime path.

The mechanism cleared its synthetic speed and correctness gates, but the run

did not prove a common workload, representative activation rate, or aggregate

user benefit large enough to repay a value-specific branch in the generic REAL

formatter. The exact candidate is preserved at archive/exp-232; runtime,

quarter-specific test exports, and quarter-only harness lanes are removed from

the publication branch. The only retained harness change is a generic untimed

assertion that exp 194's integral and fractional fixtures remain SQLite REAL

and exercise their intended formatter paths.

Reopen only with a production/downstream trace or representative application

showing both that fractional REAL formatting is material and that exact quarter

steps occur often enough to produce a meaningful aggregate win. A smaller

implementation or a broad fractional mechanism would lower the evidence burden.

Do not generalize this result to eighths or another synthetic value lattice.

Selection lesson

The failure happened before benchmarking: the runner optimized an easy-to-test

subdomain without first ranking its expected product value against other live

directions. The exploit-selection rubric now requires representative incidence,

expected aggregate value after miss tax, an explicit persistent-complexity

budget, and a cross-direction shortlist. Moonshots may still probe meaningful

ceilings, but a narrow runtime result needs the product-value evidence before it

ships. Synthetic 100%-eligible lanes are mechanism evidence only.

Validation

• Archived prototype: strict Clang C11 warning build, full analysis, direct

native differential 9/9, selectBytes 9/9, benchmark pipeline 20/20, and full

serial suite 328/328.

• Final publication branch: runtime and quarter-specific test/harness changes

removed; generic REAL fixture invariant retained and exercised by the focused

harness.

• dart run benchmark/finalize_experiment.dart --experiment=experiments/232-dyadic-real-fastpath.md.
• dart run benchmark/check_experiment_dispositions.dart.
• dart analyze --fatal-infos.
• dart test test/benchmark_pipeline_test.dart.
• Direct native differential 7/7 and full serial suite 326/326.
• JSON validation and git diff --check.