Add log-asymmetric-triangular (ltriangle3) distribution

[joethorley]

Builds on #169. Adds ltriangle3, a log-triangular distribution with a skewness parameter, following the ltriangle wiring.

Parameterization

If log(Y) has a triangular distribution with mode locationlog, half-width scalelog and skewness skew in [-1, 1], the support is [locationlog - (1 - skew) scalelog, locationlog + (1 + skew) scalelog]. The total width is 2 scalelog regardless of skew, skew = 0 is ltriangle, positive values lengthen the upper limb and skew = +/-1 is a right triangle with the mode at a support limit.

skew is estimated directly with L-BFGS-B bounds at -1 and 1 (bound = TRUE in dist_data), so right-triangular fits are flagged by ssd_at_boundary() and excluded by ssd_fit_dists() unless at_boundary_ok = TRUE.

Changes

• R/ltriangle3.R: ssd_pltriangle3()/qltriangle3/rltriangle3/eltriangle3, sltriangle3() starting values, bltriangle3() bounds, and the triangle3_ssd / ltriangle3_ssd helpers used by pdist() and model averaging.
• src/TMB/ll_ltriangle3.hpp: negative log-likelihood with the same soft-log barrier as ll_ltriangle.hpp, applied per limb; registered in ssdtools_TMBExports.cpp. Verified against the analytic density in R.
• dist_data row (bcanz FALSE, tails FALSE, npars 3, valid TRUE, bound TRUE); ssd_pmulti/qmulti/rmulti and params.R extended.
• tests/testthat/test-ltriangle3.R: test_dist, skew = 0 equals ltriangle, support limits and quantile inversion for skew in {-1, -0.5, 0, 0.5, 1}, skew recovery from simulated data, right-triangular boundary flagging, censored and outlier cases, scale invariance, model averaging support limits. Affected snapshots regenerated.
• distributions.Rmd: new section with pdf/cdf and a figure.

Findings worth knowing

• Right-triangular optima are common. Across the 61 ssddata datasets, about 10 have their maximum likelihood fit at skew = +/-1 with the mode at the smallest or largest observation (boron is one). This is a property of the triangular likelihood (the density is positive at the support edge only for a right triangle), not of the implementation, and is why the boron tests use at_boundary_ok = TRUE.
• Starting values matter. The log-likelihood has a kink at every observation. A moment-based start (support just beyond the data, mode from the mean) reached a lower objective than the symmetric start on 27 datasets and a higher one on 16, with the large gains on the largest datasets, so it is the default. On a few datasets L-BFGS-B still stops at a kink short of the optimum.
• Skew bound margins were tested (skew within +/-(1 - d) for d in 0, 0.001, 0.01, 0.02, 0.05). Exact bounds of +/-1 gave the best convergence record (0 line-search failures on ssddata versus 1 to 12 for the inset margins), so they are kept.
• Scale-invariance tolerances in the tests are looser than for ltriangle (1e-2) because the optimizer stops within a looser neighbourhood of a kinked optimum.

🤖 Generated with Claude Code