crtl strtod/printf-%g precision: cJSON floats drift by 1 ulp (tstate red)

• Type: bug (crtl numeric). Track B (library) / Track C shared surface.
• Found: 2026-07-07 by the Track T watcher — test-cjson went red the moment the cjson corpus tree was first installed (STILL-RED across SHAs, not a compiler regression; first report 20260707T191712Z-87ce0f6-borg).

Symptom

-0.0625 (exactly representable) round-trips as -0.062500000000000008: -{"delta":-0.0625, ...} +{"delta":-0.062500000000000008, ...} cJSON's print_number tries %1.15g, reparses with strtod, and falls back to %1.17g when the reparse mismatches. A 1-ulp-inexact crtl strtod (or %g formatting) forces the %1.17g path and the long digits leak out.

Same family, second sighting

tcc-by-pxx vs tcc-by-gcc: one 8-byte .data.ro constant differs when compiling tccpp.c (feature-c-corpus-tcc residual notes, 2026-07-07) — a host strtod difference embedded in emitted rodata. One fix should clear both.

Repro

tools/testmgr.py --tier full --job 'test-cjson#00'

or standalone: crtl strtod("-0.0625") vs glibc — compare bit patterns (compare as int64, not printf).

Direction

crtl strtod likely accumulates via repeated multiplication (decimal mantissa * 10^-k) instead of exact integer scaling — check lib/crtl double parsing; correctly-rounded shortest-path for values with few digits is cheap (integer mantissa + exact power-of-two/five split). %g formatting may have the mirror issue.

Gate

test-cjson green (tstate confirms), tcc .data.ro residual re-checked, lua/zlib/sqlite stay green, bXXX repro test.

Log

• 2026-07-08 — resolved, commit d3342735.