v2.4.0 — New Distributions (Foundation)

New Distributions (Foundation) — the library grows 19 → 27. Four issues
closed (#54, #55, #56, #57) across ten squash-merged PRs into
dev/v2.4.0 (#130–#135, #139, #140, #142, #143); correctness suite
grows 58 → 74; every dispatch-threshold table is now measurement-backed.

Added

• Logistic and Gumbel distributions (#54): new vector_exp SIMD batch
  pipelines; log1p/expm1 tail stability; Gumbel ships max-stable
  (gumbel_r) only — the min variant is a user-side −X reflection.
• Bernoulli and Erlang distributions (#55): delegation wrappers
  (Binomial n=1, Gamma with integer shape, rate-parameterized); Erlang
  PDF/LogPDF is pure Gamma delegation.
• FisherF and InverseGamma distributions (#56): complement-native
  steered tails via detail::gamma_q/steered beta_i; F is
  closed-form PDF + steered CDF, not a Beta delegation.
• HalfNormal and TruncatedNormal distributions (#57): regime-split
  erfc normalization Z = Φ(β)−Φ(α) in survival form; the constructor
  rejects truncation windows whose Z underflows double (~±37.5σ);
  inverse-CDF sampling.
• All eight are born compliant with the #103 ±inf-limit and #104
  finite-best-effort-quantile contracts, batch NaN propagation (#102
  class), the no-aliasing contract (#112), and carry no #125-class int
  narrowing — asserted by their enhanced tests from birth.
• accuracy_sweep + the mpmath oracle cover all 27 distributions
  (6063 → 9210 rows, oracle self-checks 42 → 72); characterization
  blocks regenerated for AVX-512, AVX2, and NEON.
• ParallelUtils::parallelForSlices / WorkStealingPool:: parallelForSlices: element-denominated serial gates with SIMD-slice
  callbacks (#143).
• LIBSTATS_MAX_SIMD_TIER capped-build calibration recipe (AGENTS.md
  step 6) and the first measured kAvx table via a capped AVX build —
  also the first native AVX-tier validation since the AVX-only machine
  retired.

Fixed

• Sliced batch paths never actually forked: 18 lambda sites in
  beta/gamma/lognormal (and the new fisher_f/inverse_gamma) passed
  slice counts to element-denominated parallelFor gates, so their
  PARALLEL strategy ran serial below ~8.4M elements — and the June
  threshold calibration measured those secretly-serial paths (#143).
  Post-repair, every tier's thresholds re-measured with sustained
  V→P crossovers; Beta PDF/LogPDF dispatches to VECTORIZED everywhere
  (a real fork is a 0.65–0.92× loss on every measured tier).
• Gamma PDF/LogPDF returned NaN at +inf for every α ≥ 1 and the batch
  LogPDF let the −4605 clamp escape at −inf; scalar and batch guarded,
  healing ChiSquared (k ≥ 2) and Erlang (#130).
• Gamma deep-tail quantiles escaped to garbage (absolute-tolerance
  stop accepted a silently-wrong band, e.g. quantile(1e-15) with true
  CDF ~1e-50): log-domain seed + relative-in-p stopping; batch
  CDF(+inf) lanes fixed for the Gamma family (#142).
• fisher_f denormal-band CDF: beta argument reformulated as
  y = x/(x + d2/d1) (#142).
• truncated_normal near-lower-bound CDF: Hermite-series form takes
  0.32–0.45 relative error at lo+1ulp down to ~1e-14 (#142).
• Stale/provisional dispatch rows retired on measured data: Cauchy CDF
  (#109, closed), von Mises CDF → NEVER on kNeon/kAvx2/kAvx (#144
  tracks the owed kAvx512 cell), Laplace kNeon trial value.

Changed

• Dead pre-v2.4.0 f_cdf/inverse_f_cdf stubs deleted; Erlang
  simplified to pure delegation; integration workflow covers all 27
  (#139, #145).
• Enhanced test binaries carry no timing label — the contract gates
  they contain must run in CI's correctness pass (#135).