fix(ta): return na, not an invented constant, on three degenerate arms

[luisleo526]

The bug

Three oscillators returned an invented constant where the published formula's
denominator is exactly zero, instead of propagating na:

indicator	degenerate condition	returned
ta.cmo	up-sum + down-sum bitwise 0	0.0
ta.tsi	smoothed abs-change denominator 0	0.0
ta.cog	window sum 0	0.0

These are author's-convenience values, not derivations — the sibling ta.wpr
already returns na on the structurally identical hh == ll. At the degenerate
point numerator and denominator are both exactly 0, and Pine surfaces IEEE NaN as
na.

Why only these three

A clean-room TradingView probe was exported and independently adjudicated. It ran
21 legs on a canonical chart: per oscillator, one leg gated on not na(...) and one
gated on equality with the engine's constant, plus a liveness control, a
positive control, and two forced-na controls.

The probe contradicted the obvious inference. "Pine yields na for 0/0, so all
six invented constants are defects" is false for two of them: TradingView
returned exactly 0.0 for ta.cci and exactly 100.0 for ta.mfi (both legs
firing 21/21). Shipping a uniform na change would have broken two arms that are
not defective in that construction.

So this PR fixes only the three arms where TradingView demonstrably returns na
(CMO_NUM, TSI_NUM, COG_NUM each 0/21) and deliberately leaves untouched:

• ta.stoch — its 50.0 is indefensible, but na is wrong too: on a flat window of
  a varying source TradingView returned a number that is neither. The only
  mechanism consistent with both probe families is carry-forward, which is inferred,
  not measured. Not shipped.
• ta.cci — split result; the two surviving hypotheses demand opposite fixes.
• ta.mfi — agrees in one construction but untested at a real flat window;
  ta.cci is the standing proof that construction can mislead.

Probe self-test passed: the liveness controls fired and both forced-na controls
stayed silent, independently confirming that plain Pine 0.0 / 0.0 is na.

Falsified test evidence, corrected

Four pieces of test evidence asserted the old behaviour and are corrected here:

• tests/test_ta_indicators_extras.cpp:268 / :293 / :340 each required the
  invented 0.0 for CMO / TSI / COG. Inverted to is_na, each citing the probe.
• tests/test_ta_osc_edge.cpp:23-24 claimed these constants were "Pine-correct,
  derived by hand from the documented behaviour". That claim is false for these
  arms; it is replaced with a per-arm evidence status block.

A negative control now pins ta.stoch 50.0, ta.cci 0.0 and ta.mfi
100.0 (compute and recompute), so a future change that "finishes the job" on an
unlicensed arm fails loudly. It passes both before and after this change — it is
anchored to current behaviour, not to this fix.

Gates

• RED first: 8 failing checks at the base commit — compute and recompute for each
  of the three arms. Each fixed arm also carries a positive control asserting a finite
  non-degenerate value, so an is_na assertion cannot pass merely from warmup.
• CTest 121/121, zero delta from baseline.
• Corpus gate unchanged (252 ran / ok=252 / 247 excellent / 4 strong / 1 anomaly)
  with all 252/252 tapes byte-identical.
• Full 417-strategy validation sweep on an independent host: board identical to
  base.

Honest limits

These arms are unreachable on the current validation feed — no bar has
high == low, and the longest identical-value runs (2-4 bars) fall short of every
length in use (5-25). So this is a spec-correctness fix with no expected
measurable effect; the byte-identical artifacts are the acceptance criterion, not
an absence of testing. A single changed tape would have falsified that analysis.

For ta.cmo the na is correct wherever the engine's own guard fires; if
TradingView's sums drift at a real flat run, engine and TV could still diverge there.
That is a separate numerics question, not made worse by this change.

A fourth arm in the same family, ta.correlation (src/ta_misc.cpp:174 compute,
:294 recompute, den == 0 -> 0.0), is left untouched because the probe carried no
leg for it. It should not be changed on the strength of the pattern — this family
demonstrably does not behave uniformly.

[Copilot]

Pull request overview

This PR aligns three TA oscillators (ta.cmo, ta.tsi, ta.cog) with Pine/TradingView behavior at exact zero-denominator singularities by returning na (IEEE NaN) instead of previously hardcoded constants, and updates tests to pin the corrected behavior while explicitly preserving known-uncertain “degenerate constant” arms.

Changes:

• Update CMO, TSI, and COG to return na<double>() when their denominators are exactly zero (including COG::recompute).
• Expand edge-focused oscillator tests to cover compute + recompute behavior for the three fixed degenerate arms, and add a negative-control test that intentionally pins stoch/cci/mfi current constants.
• Update indicator extra tests to expect na for the corrected degenerate cases.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
src/ta_oscillators.cpp	Switch CMO/TSI/COG (compute + COG recompute) zero-denominator behavior from invented constants to na.
tests/test_ta_osc_edge.cpp	Add dedicated degenerate-arm coverage for CMO/TSI/COG and a negative-control pin for unresolved arms.
tests/test_ta_indicators_extras.cpp	Update existing oscillator assertions to match na behavior for the three corrected degenerate branches.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.