fix(stoch,kama): divide by the value the guard tests, and bound the efficiency ratio (#390)

[mario4tier]

Closes #390.

Two functions, one defect shape: a guard that does not establish what the arithmetic below it needs.

STOCH / STOCHF

%K guarded highest-lowest and divided by diff = (highest-lowest)/100.0. A denormal range underflows diff to exactly 0.0 while the band still answers "not flat", so the division runs on zero — inf/NaN returned under TA_SUCCESS, which then poisons every later bar through the %K/%D smoothing.

Rather than add a second guard, this divides by the range itself and scales after. That hoist is a 2002 speed optimization and no other library has it — Tulip, ta-rs, cinar/indicator, finta, technicalindicators and yata all compute (c-l)/(h-l)*100. Matching them removes the mismatch structurally, deletes diff from the body and from the streaming handle, and makes the endpoints exact: a close on the window high is x/x, so %K is exactly 100.0.

Measured over a magnitude sweep including denormals: 0 non-finite outputs for well-formed bars where the old form gives 24, and 270,563 close-at-high cases all returning exactly 100.0.

KAMA / ER

The efficiency ratio |periodROC|/sumROC1 cannot exceed 1 — periodROC telescopes over exactly the changes sumROC1 sums. But sumROC1 is kept by add-then-subtract, so absorption leaves it below the window it stands for and the raw ratio runs unbounded: measured 1.1e18, driving the squared smoothing constant to 4.4e35 and KAMA's output 43 decades outside the hull of its own inputs. Clamping restores the bound, so the constant stays in (0,1) and the recurrence is a convex combination again.

This overturns er.c's standing "do NOT fix this with fabs". That ruling was about value stability and was made without the unbounded case in evidence.

Cost against v0.6.4

v0.6.4 has neither defect; #253 opened both while making the bands scale-relative — which it was right about, only the quantity the STOCH band tests moved off the divisor, unremarked.

• STOCH/STOCHF need no new tolerance. Their existing FUZZ_064_TOL rows absorb the change with four orders to spare (max observed 2.27e-12 against 1e-9).
• KAMA needs one skip, gated on the absorption condition and computed from the inputs alone: kama-skipped 1089 → 1139 of 166,852 comparisons. Every case it drops had v0.6.4 emitting a ratio outside [0,1] on the EXTREME shape — in the worst one, a negative output from an all-positive input, from a function flagged TA_FUNC_FLG_OVERLAP.

Verification

Full C suite · generator cargo test · --xlang-hash (52,000 cases, four backends bit-identical) · --codegen against the frozen oracle (6 passed / 0 failed in each backend) · fuzz-064 · regen-check on a clean tree.

LEGACY_TOL re-measured — STOCH 8.53e-14, STOCHF 4.26e-14 — and both rows moved to cause (b), whose prose the change made false where they were.

Two legs pin the clamp, each sabotage-proven by deleting it and watching them redden: a KAMA sign-mirror (KAMA is exactly sign-symmetric but for the signed comparison the clamp neutralizes) and an exact-1.0 predicate on ER's zero-denominator leg. The divisor sweep's self-test (#392) now reconstructs the defect rather than pointing at STOCH, and pins all three of its mechanisms.

Performance

Shipped-build A/B (libta-lib.a, separate TUs, no LTO), Zen 4 laptop, one function per process, min of 12 alternating rounds:

	ratio
STOCH	0.990
STOCHF	0.958
KAMA	1.050
RSI / SMA / EMA / ATR / WILLR (controls)	1.005 / 0.995 / 1.000 / 1.000 / 1.006

Controls inside ±0.6%. STOCH and STOCHF get slightly faster; the per-bar clamp costs KAMA about 5%. ta_bench --language=c was tried first and rejected: it reported STOCH +7.4% and STOCHF −18.0% from the identical edit, with spreads of 30–51% — the single-TU LTO inlining instability the ta-bench skill documents.

Not fixed here

CCI, CORREL and WILLR carry the same guard/divisor mismatch and are out of scope. Reproduced through the shipped API: CCI returns 39 of 67 outputs inf under TA_SUCCESS at denormal quote scale; CORREL returns NaN; WILLR divides residue into the full range on the machine-flat window where STOCHF now returns zeros, despite being documented as bounded in [-100, 0]. Follow-up issue to come.

https://claude.ai/code/session_01Mxrh2CPbhZnLvtzepLeCwt