Improve Weighted Moving Average (WMA) indicator parity with Cython

[nicolad]

Restores feature-parity and behavioural equivalence between the Rust WeightedMovingAverage implementation and the canonical Python/Cython reference.

---

1 · Why this matters — Context & Motivation

Theme	Why it matters
Correctness ⇄ Consistency	Mixed Python (quant research) + Rust/WASM (execution) stacks must emit identical WMA values; silent divergence leaks P ∿ L.
Robust edge-case handling	Zero-weight, sub-ε weight-sum or NaN inputs previously produced UB or silent drift.
Simpler mental model	Removes the redundant has_inputs flag and double-initialisation logic.
Test-driven culture	Full-parity rstest suite protects against future regressions and enables confident refactors.

---

2 · What changed

Concern	Rust before	Rust after (this PR)
Validation rules	Sparse, panicked on some edge cases only	✨ Full validation in new_checked: period > 0, weights.len() == period, ∑wᵢ > ε
Redundant state	has_inputs boolean tracked manually	Removed — has_inputs() now derives from !inputs.is_empty()
Weighted sum loop	Manual index + reverse_weights vec	Direct zip(self.inputs.iter().rev(), self.weights.iter().rev()) (zero alloc)
Initialisation flag	Two competing assignments	Single canonical rule: initialized = count() ≥ period
Error messages	Generic	Descriptive, actionable ("period must equal weights.len()")
Unit tests	8 cases	30 cases / 100 % line & branch coverage
Docs	Minimal	Elaborate /// comments on invariants & panics
Performance note	–	No observable regression (same O(1) update); future SIMD possible ✈️

---

3 · Implementation notes

• Validation constant EPS = f64::EPSILON ensures weight-sum check does not break on FP rounding.
• weighted_average() is safe because self.weights.len() == self.period is proven by construction.
• Negative weights are tolerated iff the total is positive (common in certain signal schemes); covered by tests.
• NaN inputs poison the output until reset() — matches Python behaviour.
• WeightedMovingAverage is still !Send + !Sync; wrap in Arc<Mutex<…>> when sharing between threads.

---

4 · Tests added / updated

Test name	Purpose
new_panics_on_zero_period / new_checked_err_on_zero_period	Guards period > 0
new_panics_on_zero_weight_sum / weight_sum_below_epsilon variants	Enforces ∑wᵢ > ε
initialized_flag_transitions	Canonical initialisation after period samples
count_matches_inputs_and_has_inputs	Verifies redundant flag removal
weighted_average_with_non_uniform_weights	Numerical accuracy with 1:2:3 weights
negative_weights_positive_sum	Mixed-sign weights sanity
nan_input_propagates	NaN poisoning rule
single_period_returns_latest_input	period == 1 degenerates to last price
value_with_sparse_weights	Sparse weight-vector edge case
+ 20 more robustness cases → 30 total / 100 % coverage.

---

Related #2507

[nicolad]

needs some attention from you @cjdsellers 👇

#	Weight rule / behaviour	Rust WMA (new_checked)	Python WMA	Parity
1	weights.len() == period	✔ Enforced (check_predicate_true)	✔ Enforced	✅
2	Σ weights > ε (f64::EPSILON / np.finfo(float).eps)	✔ Enforced	✔ Enforced	✅
3	Negative weights allowed if Σ > ε	✔ Allowed	✔ Allowed	✅
4	Weights optional (None) → SMA fallback	✖ Not supported (must pass Vec<f64>)	✔ Supported (weights=None)	⚠️
5	Warm-up uses trailing subset of weights (weights[-n:])	✔ Yes (weights[self.period-n..])	✔ Yes	✅
6	Guard that warm-up Σ weights > ε	✖ Open	✖ Open	⚠️
7	Invalid weights ⇒ runtime error	panic! (in new), Err (checked)	ValueError	✅
8	Degenerate case period=1, weights=[1.0] → passthrough	✔ Works	✔ Works	✅
9	NaN inside weights detected / rejected	✖ No explicit guard (Σ check catches)	✖ Same	✅

[cjdsellers]

Great stuff, and thanks for all the tests!

Removing has_inputs is a nice refinement.

On the suggestions, I don't think we need to support optional weights at this stage?

[cjdsellers]

If we're using a VecDeque with a fixed capacity then this might not be necessary?

[nicolad]

VecDeque::with_capacity(n) only pre-allocates space; it doesn’t enforce a hard cap.
If you keep pushing after len == n, the deque re-allocates and grows, so you must still pop_front() (or otherwise drop the oldest element) to keep the window at exactly period items.

[cjdsellers]

thanks @nicolad 👌