prime 0.5.5

prime 0.5.4.9001

Compute ridge ranking and BIC on a centered unit-variance response so that numerical guards do not change screening when response units change. Ridge path RSS diagnostics remain in the original response units; BIC diagnostics omit the common response-scale constant.
Compute paired comparison scales in response-variance units to avoid squaring extremely large or small raw risk differences; diagnostics retain their units.
Normalize precomputed games by their positive full-set value and use log-space ratio potentials, returning allocations in the supplied game units.
Apply the relative null tolerance only to interior subset values, preserve both endpoints, and remove positive-value snapping from projection. Setting zero_tolerance = 0 now retains every representable positive projected value.
Validate the null tolerance and reject unrepresentable normalized game ratios clearly. Zero-full games project to zero and return zero allocations.

Reject exact projection edge counts beyond 32-bit index capacity before subset evaluation or projection allocation, without introducing a RAM budget.
Support a single predictor with the existing MASS ridge normalization and preserve screening under response-colliding or non-syntactic predictor names.
Correct singleton-cap and stopping descriptions, and document the existing empty-screen allocation sentinel.
Replace production forward all-row distance sweeps with subset-specific indexed search, fused with querywise local fits. No query-by-n distance or query-by-K neighbor array is materialized in the production risk path.
Add explicit neighbor_eps (zero exact; positive Euclidean distance approximation) and query_block_size controls. All queries and the entire neighbor pool remain in use by default, with the existing K rule.
Separate scalar risk caching from a selection cache retaining at most two query-contribution vectors; allocation no longer retains a vector for every subset.
Native exact equal-distance ties use original row indices. This can differ from historical FNN tie ordering on duplicate or quantized data. Continuous-data and tied-data accuracy are verified separately.
Vendor the unmodified BSD-2-Clause nanoflann 1.9.0 header and license.

Replaced the all-query quadratic forward-distance cache with shared blocks of at most 256 queries. All queries, candidate neighborhoods, and statistical calibration rules remain unchanged; the distance cache now uses at most 256 * n doubles rather than query_count * n doubles.
Preserved per-candidate querywise SSE addition order and boundary-tie flags across blocks, including multithreaded candidate evaluation.

Replaced the fixed ridge threshold with BIC selection along the ridge-ranked prefix path.
Added paired querywise calibration to forward gains and sequential backward deletion losses.
Preserved raw risk ordering when all forward candidate values lie below the zero-signal boundary; calibration still determines whether the leader enters.
Constructed the final proportional allocation from noise-adjusted subset risks before monotone projection.
Removed the obsolete alpha tuning argument.

Initial release of the PRIME screening and allocation workflow.
Added local-linear subset-value estimation, inclusive forward screening, dimension-adjusted backward reduction, and Data Twinning support.
Added structured print, summary, and plot methods for fitted and screening objects.