perf(cci,correl): test the divisor instead of reformulating around it (#395)

[mario4tier]

Follow-up to #397, on the maintainer's call: the reformulation that fix used cost 11% on CCI and 14% on CORREL, and what it bought over simply testing the divisor is precision below ~1e-308 and ~1e-78 — magnitudes outside TA-Lib's operating domain, unreachable by the fuzz generator's own bar geometry, quoted by no instrument. Not worth the regression.

Both functions keep their original arithmetic and add an exact test on the divisor:

CCI     && 0.015*tempReal2 != 0.0      then  tempReal/(0.015*tempReal2)
CORREL  && ssX*ssY > 0.0               then  spXY/sqrt(ssX*ssY)

Exact, not a band, in both — the flatness question is already answered by the tests above them, and these ask only whether the value the division uses exists. An absolute band on either would be the #253 defect at a new address.

The #395 defect is still fixed. Such a window is now degenerate and takes the same 0.0 arm an identical-price window already did — which is what cci.yaml's silence about NaN/Inf and correl.md's "the output is 0 rather than an error or NaN" both already promised. What is given up is the correct CCI value below ~1e-308 and correlations accurate past ~1e-78; both now answer 0.

Value-neutral, measured three ways

• LEGACY passes with CCI's tolerance row removed — bit-exact against frozen v0.6.4 again, and back in the "bit-exact" list ta_test_legacy.c names.
• --fuzz-064 CCI tolerated cases: 1365 → 18, the 18 being the pre-existing TA_CCI need a fix for when checking zero values #7 ones. 1347 cases returned to bit-exact.
• An output hash over periods 2..60 of two corpora is bit-identical to the pre-CCI, CORREL and WILLR divide by a value their guard never tested (the #390 shape, three more sites) #395 library, for both functions.

Performance

Two full independent builds, Zen 4 laptop, min of 14 and 16 alternating rounds × 200/250 reps over 20k bars:

	run 1	run 2
CCI	0.999	0.980
CORREL	0.954	1.017
control spread (SMA, RSI, ATR, STOCH)	0.975–1.001	0.953–1.011

Both sit inside the noise where they were 5× outside it. (An earlier attempt that linked two objects ahead of the archive instead of building two libraries moved RSI by 8% and was discarded.)

Sweep

CORREL clears without an annotation now: the clearing test looks through sqrt/fabs, the same transparency the firing side already used, and it is sound because sqrt(v) is zero exactly when v is. The inline arm's self-test is rebuilt around the shipped shape — it removes guards rather than injecting a divisor — and still pins all three halves: fires on a guard naming only an operand, clears when one names the divisor, silent when nothing bounds either.

WILLR is untouched. Its change moves values on ordinary data, its batch tier came out ~8% faster, and the clamp ruling stands.

Green

Full C suite · generator cargo test + clippy -D warnings · Rust clippy/doc/doctests/tests · --codegen against the frozen oracle (161 passed, 0 failed in each of four backends) · --xlang-hash (four backends bit-identical) · --fuzz-064 (0 failures) · regen-check · source-list/cargo-lock/stream-retcode checks.

Both value legs re-sabotage-proven against a build of the pre-#395 arithmetic: CCI came back ±Inf on all 39 bars, CORREL NaN, -1, +1.

https://claude.ai/code/session_01GesMuivQUwX4xHr4hEtnS8