[ENH] Speed up Arsenal classifier - #3655

Cast X to float32 before _apply_kernels: the kernel weights and output
features are already float32, so float64 input only added per-element
promotion in the hot loop. Matches the original ROCKET reference
implementation, which is float32 throughout. Measured 1.11-1.12x on
univariate transform (interleaved A/B, min of 5), neutral on
multivariate; outputs shift by at most 1.9e-05 across univariate,
multivariate and unnormalised reference configurations (tolerance-level
change agreed for this case, not bit-identical). All rocket-family
tests pass unchanged.

Co-Authored-By: Claude Fable 5 <[email protected]>

# Conflicts:
#	aeon/transformations/collection/convolution_based/_rocket.py

…tests

Nothing passed keep_transformed_data=True after the train-estimate path
was fused into _fit_ensemble_estimator, so the parameter, the Xt
accumulation and the list-wrapped Xt/idx indirection in
_train_probas_for_estimator are removed. Outputs verified identical to
the pre-cleanup branch, which is itself equivalent to main within 3e-08
(the float32 rocket tolerance).

Tests: the OOB-indices docstring now matches what is asserted,
fit_predict coverage checks valid probability rows rather than only a
shape, and a new test asserts n_jobs=1 and n_jobs=2 produce identical
weights, predictions and train estimates.

Co-Authored-By: Claude Fable 5 <[email protected]>

RidgeClassifierCV with the default scoring=None sets best_score_ to the
negative LOO mean squared error, so ensemble members were weighted by
|LOO error|: all weights were negative and worse members received
larger absolute weight, inverting the CV-accuracy weighting described
in the docstring and the HC2 paper. Passing scoring=accuracy at both
sites makes best_score_ the LOO CV accuracy; this also switches alpha
selection to accuracy, a deliberate choice.

Measured impact at default parameters is negligible (identical test and
OOB accuracy to 4 d.p. on unit_test, italy_power and basic_motions over
3 seeds) because member LOO errors cluster within ~4%, but small
heterogeneous ensembles such as the results_comparison test config
shift visibly. Expected classifier results for Arsenal regenerated with
the writer recipe; the old stored values contained the -0.0
probabilities characteristic of the negative weights. A new test guards
that weights lie in [0, 1].

Co-Authored-By: Claude Fable 5 <[email protected]>

- class_weight docstring no longer offers balanced_subsample, which
  RidgeClassifierCV does not accept and which raised at fit; it now
  documents dict/balanced/None, and a test checks the parameter
  reaches every member ridge.
- _fit_ensemble_estimator returns the member weight, removing the
  steps[2][1] re-digging; the scaler no longer receives a meaningless y
  and uses one fit_transform pass.
- Contract fitting clamps each batch to the remaining budget so
  contract_max_n_estimators is never exceeded when n_jobs > 1; the
  contract test asserts the cap.
- An all-in-bag bootstrap now reports weight 0.0 (no evidence) instead
  of a fake perfect accuracy; the value was and is unused in
  aggregation because the OOB index set is empty.
- predict_proba no longer rounds to 8 decimals; the rounding lost
  information and could manufacture exact ties for the random
  tie-breaker. Stored expected results are unaffected at 4 decimals
  (verified by regeneration).

Co-Authored-By: Claude Fable 5 <[email protected]>

test_arsenal was five near-identical fit-and-check blocks; it becomes
one test parametrised over the three rocket transformers and one/four
channels, with named configuration reused in the assertions, plus a
separate invalid-input test. Docstrings now state exactly what is
asserted (the class-vote and OOB helper tests no longer claim memory
or vectorisation properties they do not test), implementation-detail
names are dropped, seeds are fixed throughout, and spellings follow
British English.

Co-Authored-By: Claude Fable 5 <[email protected]>

Set in _fit and never read anywhere; the base class already enforces
equal-length transform input, so the attribute implied a length check
that does not exist. The fit-time length is now a local used only to
cap kernel dilation.

Co-Authored-By: Claude Fable 5 <[email protected]>

_transform_kernels(X) now means the same thing on Rocket, MiniRocket
and MultiRocket: apply the fitted kernels, the caller owns any
normalisation. MiniRocket applies no normalisation so it delegates to
_transform; MultiRocket extracts the kernels-only part of _transform
exactly as Rocket does. Arsenal drops both of its per-transformer
branches and calls the contract uniformly in fit and predict.

Bit-identical outputs verified for the minirocket and multirocket
Arsenal paths (predict_proba and fit_predict_proba); a new
parametrised test asserts _transform_kernels equals transform for all
three transformers when no normalisation is configured.

Co-Authored-By: Claude Fable 5 <[email protected]>

The batch method was never exercised (the transformer calls forward
directly): a new test asserts both of its branches return exactly the
single-pass result. A single-group test covers the divisor == 1 path
in forward, which every existing test skipped by using eight or more
groups, and pins the feature-count invariant
2 * n_kernels * n_groups * num_dilations.

Co-Authored-By: Claude Fable 5 <[email protected]>

Arsenal fits and predicts its members in joblib threads, and each
member transform entered a numba parallel=True region. Numba's default
workqueue threading layer terminates the process on concurrent entry
("Concurrent access has been detected"), which surfaced on CI as
crashed pytest-xdist workers in the Arsenal n_jobs tests; reproduced
deterministically with NUMBA_THREADING_LAYER=workqueue.

Each rocket transform kernel is now compiled twice from one body:
the existing parallel dispatcher for standalone n_jobs > 1 use, and a
serial nogil dispatcher (prange degrades to range) used when
n_jobs == 1. Serial kernels never enter the threading layer, so
ensemble members transform genuinely concurrently. The n_jobs > 1
parallel path additionally holds a shared lock around the global
thread-count swap and launch.

The workqueue repro survives 20 Arsenal n_jobs=2 fits where the old
code aborted immediately. Serial timing equals the previous
parallel-at-one-thread path within noise. MiniRocket and MultiRocket
ensemble outputs are bit-identical; Rocket outputs shift by at most
2e-06 because fastmath compiles the serial and parallel pipelines
with different reassociation, an order below the accepted float32
tolerance. All 82 rocket-family and Arsenal tests pass.

Co-Authored-By: Claude Fable 5 <[email protected]>

The dual serial/parallel kernel compilations introduced for thread
safety could produce different results because fastmath lets the two
compilations reassociate float32 arithmetic differently; the estimator
multithreading check caught MultiRocket giving different transform
results at n_jobs=1 versus n_jobs>1.

Per transformer, chosen by measurement:
- Rocket: the outer dispatchers are compiled without fastmath (measured
  free); the per-kernel helper functions keep fastmath because they are
  separately compiled shared units used by both dispatchers, so results
  are identical by construction. Full performance retained (interleaved
  A/B equal to the previous fastmath build).
- MultiRocket: both compilations drop fastmath (~5 percent, within
  noise, of transform time) since its arithmetic is inline in the
  compiled bodies.
- MiniRocket: fastmath is worth 1.4x on its transform, so it reverts to
  a single parallel compilation used for every n_jobs value under the
  shared numba lock; identity across thread counts holds because the
  same machine code runs, and the lock keeps ensemble threading safe.

Verified: transform outputs identical for n_jobs=1 vs 2 across all
three transformers over multiple shapes and seeds, the estimator
multithreading checks pass, all 82 rocket-family and Arsenal tests
pass, and the workqueue crash repro still survives.

Co-Authored-By: Claude Fable 5 <[email protected]>

Addresses Matthew's review point that the one-hot predict_proba rewrite was
not different from the default and did not belong in this PR.

The branch had vectorised BaseClassifier._predict_proba and then added
byte-identical copies of that body to HydraClassifier and
MultiRocketHydraClassifier, plus the same block inline in the Rocket,
MiniRocket and MultiRocket classifiers: the same four lines in six places.
None of it is used by Arsenal, which overrides _predict_proba itself, so
all of it is reverted to main.

Arsenal keeps its own internal use of np.searchsorted for mapping member
predictions to class indices when aggregating votes, which is unrelated.

Co-Authored-By: Claude Opus 5 <[email protected]>

Timing the branch against main showed the transform changes were a net loss
for two of the three rocket variants. Paired runs, alternating order, fresh
process and numba cache per measurement, n_kernels=2000, n_estimators=10:

  rocket        n_jobs=1   +10.0% fit  +11.2% predict  +10.1% total
  rocket        n_jobs=4   + 9.5% fit  +12.1% predict  +10.3% total
  minirocket    n_jobs=4   -11.1% fit  -119%  predict  -19.7% total
  multirocket   n_jobs=1   -39.5% fit  -44.6% predict  -41.5% total

MultiRocket's loss is the fastmath removal, which 13f878d described as
"~5 percent, within noise"; it is not. MiniRocket's loss was not the numba
lock: restoring the dual serial/parallel compilation it used to have did not
recover it, though that experiment did show the two compilations are
bit-identical over 216 n_jobs pairings, so the invariance concern that
removed them was unfounded.

Both files now match main exactly. Rocket keeps its changes, being the only
variant with a measured win and the one Arsenal uses by default.

Arsenal now picks the transform entry point via _transform_with: Rocket is
the only rocket transform that normalises by default, so only it needs the
kernels-only path once the ensemble has normalised. MiniRocket never
normalises and MultiRocket defaults to normalise=False, so _transform is
already kernels-only for them.

Co-Authored-By: Claude Opus 5 <[email protected]>

…to ajb/arsenal