v2.5.0 — Correctness & accuracy

Correctness and accuracy release — no source API change; the public API
stays additive over v2.1.0. 75 issues closed (#137, #157–#177, #180,
#183–#212, #216, #223–#242, #264, #274–#276) from one branch
(dev/v2.4.2, 156 commits), each fix with a regression gate shown to
fail on the code before it. Every one of the 27 distributions was swept
against an mpmath oracle on all three fleet ISAs (AVX-512/clang-cl, AVX2
and NEON/AppleClang): 26,882 rows per ISA including the new survival
rows, 190 oracle self-checks, 0 contract violations (v2.4.1: 32 on each
ISA). The correctness run grows 86 → 132 tests, all of them in CI;
CI gains ThreadSanitizer and FMA-contraction legs; the dispatch tables
are re-derived from fresh captures and verified by a loaded A/B.

Fixed — accuracy

• Quantiles: Gaussian/LogNormal and the probit return finite values for
  p < 2^-54 (#158); Student-t in the deep tail and at large ν (#159,
  #233); Gamma/Erlang/ChiSquared at every p and at extreme shapes (#160,
  #223, #226); Beta to the accuracy law at every p — inverse_beta_i
  dropped its 1e-8 stop and clamp and its tiny-shape cancellation (#137,
  #230, #232); Poisson, Binomial and NegativeBinomial/Geometric search
  from a normal guess, unbounded, and return the smallest k with
  F(k) ≥ p, ties included (#170, #171, #204); discrete quantiles are
  non-decreasing across p = ½ and reach Q(1); NaN in gives NaN out in
  every class (#194).
• Incomplete gamma and beta to double precision at large shape (#166);
  the discrete log-pmfs without the lgamma cancellation at large counts
  (#172); stable log-beta normalisers; log1pmx in atanh form; the
  constants SQRT_PI, INV_SQRT_2 and four others correctly rounded.
• Lower-tail CDFs of Weibull, Exponential, Rayleigh and Pareto through
  log1p/expm1; Geometric CDF in closed form (#202); Gamma/Erlang CDF where
  x·rate overflows (#224); the Gamma and InverseGamma rate product
  corrected to second order, also under FMA contraction; Fisher-F CDF
  with a subnormal intermediate (#216) and at DBL_MAX for tiny d2
  (#242); Student-t CDF beyond |t| = √ν; Beta CDF at shapes below 1e-3
  (#229) and above the branch point for b < 1.
• von Mises: CDF and quantile by a fixed-cost quadrature in
  w = κ(1 − cos θ) (closes #111); density at large κ without cancellation;
  entropy with both signs right (#234); the standardized circular
  kurtosis (#276); no hang at κ ≥ 6.5e15 (#231); the normal limit above
  κ = 1000.
• TruncatedNormal: near-bound CDF series and quantile (#201, #225),
  half-infinite windows, far-tail windows with a subnormal normaliser
  (#205), variance/skewness/kurtosis on narrow windows (#236), fit on
  well-posed windows (#239).
• Support edges and limits: Gamma/ChiSquared/Weibull at x = 0 with shape
  < 1 (#161); Rayleigh and Weibull at +∞ (#164); Poisson logpdf −∞ out of
  support (#165); NegativeBinomial/Geometric cdf(1e300) (#206); Gamma,
  Weibull and Uniform contract edges; subnormal and overflowing scale
  parameters in Rayleigh, HalfNormal and Uniform (#203, #240).
• Moments and fits: Gamma entropy from shape ~1e10 (#235); Student-t fit
  for ν ≤ 2 and past ν = 1000 with a Gaussian limit (#237); Weibull fit
  scale-invariant (#238); Gamma fit on degenerate data throws (#207);
  Binomial fit rejects bad observations (#195); the coefficient of
  variation from one parameter state.
• getSurvival is complement-native in every distribution: the eleven
  that computed 1 − F returned exact 0 in the upper tail, and
  getHazard returned +∞ inside the support (#197).
• Samplers: Beta at small shapes in log space (#228); von Mises at tiny
  κ (#231); Fisher-F without spurious +∞ (#242); Uniform when b − a
  overflows (#240); Gaussian without a thread_local Box–Muller spare,
  so seeded streams are reproducible (#188).

Fixed — concurrency and contracts

• One parameter snapshot per call: scalar methods and batch lambdas no
  longer mix two parameter states (Binomial, NegativeBinomial, von
  Mises, Beta; #175); PARALLEL and WORK_STEALING run the
  VECTORIZED kernel over 1024-element slices in every operation (#191),
  with a partial last slice folded into the previous one so every
  strategy returns the same bits; the Student-t PARALLEL CDF is no longer
  a serial loop (#176).
• Locks: setters validate under the lock (#183); move-assignment locks
  both objects (#184); TruncatedNormal fit commits only if the bounds
  are unchanged (#185); vector sample draws from one state (#186);
  owners reject what their delegates reject (#227); lock-free getters
  read the new parameters after copy/move assignment (#192); copy/move
  constructors copy the cache (#163); the WORK_STEALING lambdas wait on
  their own latch, not the global pool (#180); SystemCapabilities
  reads the physical core count from the CPU detector (#190).
• Casts and UB: counts stay in double until validated (#167); the
  lgamma signgam write is avoided in library code (#173; the one
  libstdc++ write inside std::cyl_bessel_i is #267, v3.0.0).
• Streams: operator<< writes full precision in every class (#208);
  Poisson and TruncatedNormal operator>> read their own output (#187,
  #193).
• Batch overloads check sizes on empty input and keep the overlap
  assert (#196); exact special-parameter shortcuts fire only on exact
  equality (#199).

Added

• Oracle and sweep: tools/accuracy_sweep emits both sides of every
  kernel switch and the survival rows; tools/accuracy_vs_mpmath.py
  scores relative error down to DBL_MIN and ULP distance below it,
  carries 190 self-checks with negative controls, and refuses to rewrite
  the doc from a partial sweep.
• Tests: test_concurrency_stress (all 27 distributions: racing
  setters, readers, fits and streams against a one-state oracle, a
  strategy differential, a deadlock watchdog; #210-class),
  test_distribution_identities and test_io_fit_fuzz (#210),
  test_survival_tails, test_contract_gates; the timing cases split
  out of mixed binaries so their 214 correctness cases run in CI (#198,
  #212); the batch-fit speedup gate measures above the parallel
  threshold in steady state (#177).
• CI: a ThreadSanitizer leg (Linux, clang-17) and an FMA-contraction leg
  (clang-17, -mfma); manual dispatch; the macOS stress-test variance
  documented.
• Tooling: scripts/analyze_crossovers.py (the canonical sustained
  V→P crossover; #146), scripts/derive_threshold_rows.py (every
  dispatch table from its bundle under one set of rules),
  tools/bench/table_ab_bench.cpp (AUTO against the forced strategies
  under two headers) and tools/bench/v242_cost_bench.cpp (per-case
  cost against v2.4.1).
• Build: warning-clean under clang-cl, which takes cl.exe's /arch:
  flags and is the Windows performance build (6245e4e, 606607f); a
  .git-blame-ignore-revs for the clang-format 20.1.8 pass; the manual
  GTest fallback links (#274); macOS ad-hoc codesign is idempotent
  (#162).

Performance

• Dispatch tables (dispatch_thresholds.h) re-derived for kAvx512,
  kAvx2, kAvx and kNeon from three-run captures at the release head,
  then verified by a loaded A/B (tools/bench/table_ab_bench.cpp, AUTO
  against the forced strategies under the old and new headers): AUTO rows
  more than 1.25× behind their best forced strategy fell 202 → 24 on NEON
  with the first derivation and 13 → 9 with eight more rows, 208 → 5 on
  AVX-512, and 158 → 2 (AVX2), 146 → 2 (AVX) and 173 → 17 (SSE2) on Kaby
  Lake. The discrete pmfs, the Student-t and von Mises CDFs and Beta
  leave NEVER; most continuous CDFs cross to PARALLEL at 8k–25k instead
  of 25k–2M; the only NEVER cell left is kNeon Uniform PDF.
• Speed work on the accuracy paths: Student-t and Beta densities SIMD at
  every shape; vector_log1p and vector_expm1 where they win; the
  Gamma family's Stirling-form batch density from α = 20; the Binomial
  quantile from a normal guess.
• Costs of the accuracy fixes, against v2.4.1, by parameter band (the
  cost steps at the algorithm switch points; the full table is in
  docs/VALIDATION_HISTORY.md): Gamma and Beta densities 1.0× at small
  shapes, 2.6–3.4× from shape 20 (the Stirling prefactor); Student-t at
  ν ≥ 1e3 CDF 4.4–5.2× and quantile 6.5–13× (the large-ν tail paths);
  the von Mises quantile 5–23×, flat in κ (a fixed-cost Newton on the
  exact CDF replaced a grid lookup), while the von Mises CDF is faster
  above κ = 30 (0.8× at 30, 0.2× at 300); the discrete densities
  1.3–1.8× (the Loader-style log-pmf); Gamma and discrete quantiles and
  Poisson CDF at large λ faster everywhere (0.04–0.7×).

Known limitations

• Thresholds are per distribution, not per shape: Gamma(20), Gamma(1000)
  and Student-t(1000) densities cross to PARALLEL earlier than their
  rows; a v3.0.0 input.
• Deferred to the accuracy patch (milestone #8): #248–#251, #253,
  #257–#265, #273 (TruncatedNormal variance on windows narrower than
  ~1e-154 σ), #277 (the symmetric-window median reads −1.4e-16 on the
  UCRT). To v3.0.0: #247, #252 (correctly rounded I_x, corvus), #266 (the
  shape-20 Stirling switch), #267 (std::cyl_bessel_i's signgam
  write, suppressed under TSan).