feat: Family 03 MACD & Price Oscillators — APO / AO-Hist / CFO / Zero-Lag MACD / Elder Impulse / STC (#41)

* feat(apo): add Absolute Price Oscillator

EMA(close, fast) - EMA(close, slow). Like MACD without the signal EMA.
Defaults to (fast = 12, slow = 26); fast must be strictly less than
slow.

Touchpoints: apo.rs + mod.rs + lib.rs re-export, PyApo + __init__.py
+ test_new_indicators SCALAR + test_known_values flat reference,
ApoNode + index.d.ts/index.js + indicators.test.js factory + reference,
WasmApo via scalar macro, scalar-fuzz target, README + CHANGELOG.

* fix(apo): add PyApo + ApoNode + WasmApo bindings missed from ec269d8

The previous APO commit (ec269d8) only registered APO in the Python
__init__.py / Node index.js / Node index.d.ts / fuzz / tests / docs.
The actual PyApo pyclass, ApoNode napi class, and WasmApo wasm class
edits silently no-op'd because the underlying lib.rs files had been
touched by a branch switch between Read and Edit. The bindings were
therefore advertising APO from the Python module / Node package /
WASM module but not actually exposing it.

Fix: insert PyApo block + add_class call in bindings/python/src/lib.rs,
ApoNode block in bindings/node/src/lib.rs, WasmApo macro line in
bindings/wasm/src/lib.rs. cargo test workspace stays at 615 (no new
tests added; the existing test_known_values + indicators.test.js
references would have failed at import once the bindings rebuilt
without these classes).

* feat(ao-histogram): add Awesome Oscillator Histogram

AO - SMA(AO, sma_period). A configurable variant of the existing
AcceleratorOscillator (which fixes fast=5, slow=34, sma=5).
Three parameters; defaults match Bill Williams' Accelerator.

Touchpoints: awesome_oscillator_histogram.rs + mod.rs + lib.rs
re-export, PyAoHist + __init__.py + test_new_indicators CANDLE_SCALAR
+ test_known_values flat reference, AwesomeOscillatorHistogramNode +
index.d.ts/index.js + indicators.test.js factory + reference,
WasmAoHist, candle-fuzz target, README + CHANGELOG.

* feat(cfo): add Chande Forecast Oscillator

100 * (close - LinReg(close, period)) / close. Positive when close
overshoots the linear forecast, negative when it undershoots. Holds
the previous value if the close is zero (percentage form undefined).
Single param period (default 14).

Touchpoints: cfo.rs + mod.rs + lib.rs re-export, PyCfo + __init__.py
+ test_new_indicators SCALAR + test_known_values linear reference,
CfoNode + index.d.ts/index.js + indicators.test.js factory + reference,
WasmCfo via scalar macro, scalar-fuzz target, README + CHANGELOG.

* fix(cfo): add WasmCfo binding missed from 733afd9

* feat(zero-lag-macd): add Zero-Lag MACD

Classic MACD topology with ZLEMA substituted for EMA everywhere:
faster reaction to trend changes at the cost of slightly noisier
readings. Multi-output ZeroLagMacdOutput { macd, signal, histogram }.
Three parameters (fast = 12, slow = 26, signal = 9); fast must be
strictly less than slow.

Touchpoints: zero_lag_macd.rs + mod.rs + lib.rs re-export, PyZeroLagMacd
+ __init__.py + test_new_indicators MULTI + test_known_values flat
reference, ZeroLagMacdNode + ZeroLagMacdValue + index.d.ts/index.js +
indicators.test.js multi factory + reference, WasmZeroLagMacd, scalar
fuzz with hand-rolled drive (multi-output bypasses the f64-only
helper), README + CHANGELOG.

* feat(elder-impulse): add Alexander Elder Impulse System

Tri-state momentum gauge: +1 (green/buy) when EMA trend and MACD
histogram both rise, -1 (red/sell) when both fall, 0 (blue/neutral)
on disagreement. Four parameters (ema_period, macd_fast, macd_slow,
macd_signal); defaults (13, 12, 26, 9) match Elder.

Internally feeds both branches on every input so they warm in parallel;
needs one bar past the slowest branch to seed direction state.

Touchpoints: elder_impulse.rs + mod.rs + lib.rs re-export, PyElderImpulse
+ __init__.py + test_new_indicators SCALAR + test_known_values neutral
reference, ElderImpulseNode + index.d.ts/index.js + indicators.test.js
factory + reference, WasmElderImpulse via scalar macro, scalar-fuzz
target, README + CHANGELOG.

* feat(stc): add Schaff Trend Cycle

Doug Schaff's doubly-Stochastic-smoothed MACD. Bounded [0, 100]
reading that reacts faster than MACD by extracting the percentile of
MACD within a recent window, half-EMA-smoothing it, and re-stochasing
the smoothed series. Four parameters (fast = 23, slow = 50,
schaff_period = 10, factor = 0.5); fast must be strictly less than
slow and factor must lie in (0, 1].

Output clamped to [0, 100] to absorb floating-point rounding. The
stochastic stages clamp to 0 when their rolling range collapses (flat
input or perfectly monotone trend), so a flat series settles
deterministically at 0 after warmup.

Touchpoints: stc.rs + mod.rs + lib.rs re-export, PyStc + __init__.py
+ test_new_indicators SCALAR + test_known_values flat reference,
StcNode + index.d.ts/index.js + indicators.test.js factory + reference,
WasmStc via scalar macro, scalar-fuzz target, README + CHANGELOG.

* fix(stc): rename last_stc -> last_value to satisfy clippy

* ci: Retry setup-node and setup-python on CDN flakes

Setup-node on Windows runners and setup-python across all OSes
occasionally fail with a silent hang or 5xx mid-download ("Attempting
to download 18..." → fail in <1s) — pure upstream CDN flake. The fix
ran on this branch's previous merge commit (24e723f) had to be
re-triggered manually via `gh run rerun --failed`.

Wrap both setup actions with continue-on-error and a follow-up retry
step that waits 30s and re-runs the same setup. The retry only fires
when the first attempt failed (steps.<id>.outcome == 'failure'), so a
green setup costs nothing extra. The retry uses the identical pinned
SHA so we still get supply-chain verification on both attempts.

Applied to ci.yml (Python matrix and Node matrix). release.yml has
the same setup-node / setup-python steps but is rarely re-run, so
the existing manual rerun pattern stays sufficient for now.

* test(zero-lag-macd): Fix MULTI dict shape mismatch + cover warmup_period

ZeroLagMACD was registered in the Python MULTI dict (which asserts a
(n, 2) batch shape) but actually emits (n, 3) — macd, signal,
histogram — like MACD. Moved out into its own standalone test
test_zero_lag_macd_streaming_matches_batch (3-tuple shape), and
included in the lifecycle sweep. Mirrors the existing Alligator
pattern for 3-output candle indicators.

Also adds a unit test for ZeroLagMacd::warmup_period that pins both
the (12, 26, 9) classic case and a small-period config — these four
lines were the codecov/patch miss on PR 41.

bindings/python/tests/test_known_values.py

@@ -234,6 +234,56 @@ def test_vidya_constant_series_holds_seed():
     np.testing.assert_allclose(out[4:], 42.0, atol=1e-12)
 
 
+def test_zero_lag_macd_constant_series_converges_to_zero():
+    # Each inner ZLEMA reproduces a constant, so macd, signal and histogram
+    # are all 0 once the slowest branch warms up.
+    out = ta.ZeroLagMACD(3, 5, 3).batch(np.full(60, 42.0, dtype=np.float64))
+    # Take the last row and verify all three columns are 0.
+    last = out[-1]
+    assert math.isclose(last[0], 0.0, abs_tol=1e-12)
+    assert math.isclose(last[1], 0.0, abs_tol=1e-12)
+    assert math.isclose(last[2], 0.0, abs_tol=1e-12)
+
+
+def test_awesome_oscillator_histogram_flat_series_converges_to_zero():
+    # Flat median price -> AO = 0 -> SMA(AO) = 0 -> AOHist = 0.
+    n = 50
+    high = np.full(n, 11.0)
+    low = np.full(n, 9.0)
+    out = ta.AwesomeOscillatorHistogram(3, 5, 3).batch(high, low)
+    # warmup = slow + sma - 1 = 5 + 3 - 1 = 7.
+    np.testing.assert_allclose(out[6:], 0.0, atol=1e-12)
+
+
+def test_stc_constant_series_yields_zero():
+    # Flat input collapses both stochastic stages to zero -> STC stays at 0.
+    out = ta.STC(3, 5, 4, 0.5).batch(np.full(60, 42.0, dtype=np.float64))
+    ready = out[~np.isnan(out)]
+    assert ready.size > 0
+    np.testing.assert_array_equal(ready[-5:], np.zeros(5))
+
+
+def test_elder_impulse_constant_series_is_neutral():
+    # Flat input -> neither EMA nor MACD histogram moves -> Impulse stays at 0.
+    out = ta.ElderImpulse(13, 12, 26, 9).batch(np.full(120, 42.0, dtype=np.float64))
+    ready = out[~np.isnan(out)]
+    assert ready.size > 0
+    np.testing.assert_array_equal(ready, np.zeros_like(ready))
+
+
+def test_cfo_perfect_linear_series_yields_zero():
+    # LinReg of a perfectly linear series fits exactly, so CFO = 0 after warmup.
+    out = ta.CFO(5).batch(np.arange(1.0, 21.0, dtype=np.float64) * 2.0)
+    np.testing.assert_allclose(out[4:], 0.0, atol=1e-9)
+
+
+def test_apo_constant_series_converges_to_zero():
+    # Both EMAs reproduce a constant exactly, so APO = 0 after warmup.
+    out = ta.APO(3, 5).batch(np.full(30, 42.0, dtype=np.float64))
+    assert np.all(np.isnan(out[:4]))
+    np.testing.assert_allclose(out[4:], 0.0, atol=1e-12)
+
+
 def test_macd_constant_series_converges_to_zero():
     out = ta.MACD().batch(np.full(200, 100.0))
     # Last row's MACD and signal must be ~0.