stillwater-sc / universal / 32501448965

85%

master: 84%

push

github

fix(elreal): close the sin/cos round-trip gate -- 234 digits to 321 and scaling (#1369)

sin(asin(x))==x and cos(acos(x))==x were gated off in the LEVEL_4 transcendental
suite because they capped at ~234 digits against a >= 300 bar. They now reach 321
digits at depth 20 and scale linearly with depth -- 256 at 16, 321 at 20, 385 at
24, 512 at 32, 768 at 48, i.e. 16.0 digits per unit of depth against a
theoretical k * log10(2) = 15.95 -- tracking a clean pi/6 digit for digit. That
closes the last unmet acceptance criterion of #931 Phase 7; the full LEVEL_4
suite passes all 32 identities.

Nothing in sin/cos changed. The fix was 386737c5 (#1362), part of v4.9.0's
narrow-host work, which stopped exempting the double host from the rule that an
error-free transform must normalise its OPERANDS. Bisected: at #1361 the
round-trip is still 234 at depth 20 and 24; at #1362 it is 321 and 385.

Three passes of #1076 concluded the loss was a conditioning defect in the sin/cos
Maclaurin recurrence, reasoning that sqrt(0.75) plateaued at 19 blocks at every
depth while staying 308-digit accurate, so the plateau had to be double's
physical 2^-1022 floor -- something no downstream change could move. That middle
step named the symptom as its own cause. The floor was the bug: blocks carried
their scale in the host significand instead of the wide integer<256> exponent, so
the chain bottomed out on double's exponent range. Post-fix sqrt resolves one
block per unit of depth (40 blocks at depth 40) and the series consumes the
deeper argument unchanged.

- transcendentals_highprecision.cpp: ELREAL_SINCOS_ROUNDTRIP_HIGH_PRECISION -> 1,
  and kDepth's "double-host ceiling" comment corrected -- it is a cost knob now.
- asin_sqrt_precision.cpp: the assessment becomes a resolution, recording why the
  original diagnosis was self-confirming. The characterization now locks the
  fixed behaviour, including a depth-SCALING check: the bug's signature was a
  flat line, so clearing 300 a... (continued)

3 of 3 new or added lines in 1 file covered. (100.0%)

43095 of 50466 relevant lines covered (85.39%)

7151924.08 hits per line