Fix/dual noise units and compose cholesky

[0xgny]

fix for issue:#15

What problem this addresses

Two dimensional-analysis bugs in the pure-numpy core. Both were invisible to
the test suite: the only assertion on λ_η was >= 0, which a dimensionally
inverted quantity satisfies.

1. state/noise.py — λ_η was a count, not a rate.
calibrate() computed lambda_eta = N_jumps * dt, the reciprocal of an
intensity. σ_τ had the matching problem: sqrt(BPV/n) is per-BAR volatility.

2. events/operators.py — compose() squared the Cholesky factor wrongly.
Sigma_w is documented as a Cholesky factor and read as S @ S.T everywhere
else, but compose() used S @ S. For systemic_crisis_operator — the one
constructor with off-diagonal noise — S @ S is not PSD, so np.linalg.cholesky
raised LinAlgError. A systemic crisis could not be composed with any other
event at all.

Connection to the theoretical framework

Theorem III.1 ADDS σ_τ² and λ_η·m₂^η, so both must share a time unit. They
didn't, which made the bound depend on your bar size rather than on the market.
Now: T = n·dt, σ_τ² = BPV/T, λ_η = N_jumps/T.

Fix 2 restores the composition law in §IV: Cov = L₁L₁ᵀ + A₁(L₂L₂ᵀ)A₁ᵀ, which
is what makes the monoid/groupoid claim in the module docstring hold.

Results

Ground truth: 3 jumps/day, 2% daily vol, 10 days, two sampling rates.

	true	before (5-min)	before (1-min)	after (5-min)	after (1-min)
λ_η (jumps/day)	3.0	0.38	0.077	3.00	3.00
σ_τ (daily vol)	0.020	0.0022	0.0010	0.0198	0.0199
Cramér-Rao bound	—	9.5e-4	2.0e-4	8.4e-3	8.3e-3

compose(rate_change, systemic_crisis): LinAlgError → works.

Replication: python -m pytest tests/test_noise.py tests/test_events.py -v

Testing

Tests 133 → 144. Each new test was verified to FAIL on the parent commit:
5/6 new noise tests, 4/5 new event tests. The ones that pass on old code are
deliberate — each suite keeps a diagonal-noise case guarding against
over-correction.

Tests pin semantics, not magic numbers: λ is a rate (held jump count fixed
while varying elapsed time — the comparison a count-shaped estimator can't
fake), and covariance stays associative under composition.

Known limitation, not addressed here

λ_η units are now correct, but the Lee–Mykland detector is underpowered on
daily bars (needs ~3.5σ moves). demo/run_egamec.py reports 0.001 jumps/day
against a synthetic truth of 0.01–0.10. Separate problem; flagging it rather
than papering over it.

One downstream effect

notebooks/day03_dual_noise.ipynb prints σ_τ against a hardcoded "true"
string. The estimate is now correct (0.190 vs true annual vol 0.20 — that
notebook's dt is in years), but the hardcoded comparison is a per-bar number
mislabeled "/day", so the line now visibly disagrees. Two-line notebook fix;
happy to include it here or split it out — reviewer's call.

Questions for reviewers

1. Should calibrate() take an explicit unit (per="day") rather than
   inferring from dt? The current contract is "whatever unit dt is in,"
   which is what the notebook tripped on.
2. Is the 1e-9 Cholesky jitter in compose() the right regularizer, or
   should near-singular Σ raise instead of being silently nudged?

PS: More details in changes.md