Tracking issue for the five extensions parked on branches while the architecture settles: #7, #8, #9, #10, #11.
Why they were parked
Two of the roadmap's phases were fighting each other. Merging QAPglm() and QAPcss() into one skeleton means reconciling every branch point twice, and widening the test suite over duplicated code doubles that cost rather than removing it. Four of those branch points served model types that no test can reach, no runner can exercise, or no user can select.
Nothing is deleted. Each extension is on its own branch at bc96783, and each strip on develop is one commit that git revert reinstates.
What it bought
Suggests fell from eight modelling packages to three:
| Before |
After |
| lme4, nnet, fixest, gmm, MASS, pscl, torch, glmmTMB |
lme4, MASS, pscl |
MASS ships with R, so the marginal install cost is lme4 and pscl. The core — gaussian, binomial and poisson through stats::lm() and stats::glm(), HC3 errors, both permutation schemes, and all the network shapes — needs nothing from Suggests at all.
Proposed order back
| # |
Extension |
When |
Blocked on |
| 1 |
#8 fixest fixed effects |
after the formula grammar is settled |
a decision, not code |
| 2 |
#7 multinomial + comparison |
after the two engines are merged |
rewrite, not revert |
| 3 |
#11 torch GPU |
after the design-matrix performance work |
whether it is still needed |
| 4 |
#10 GMM |
whenever |
its singular covariance is understood |
| 5 |
#9 glmmTMB mixed |
whenever |
whether 62 dependencies are worth two combinations |
#8 comes first because it is a naming decision with no engineering behind it, and because a bar in the formula currently means one thing only, which is a state worth either keeping deliberately or leaving deliberately.
#7 should be rewritten against the merged engine rather than reverted onto it. Its 21 forks are exactly what the merge is trying to remove.
Open question
Two of these are candidates for dropping rather than reinstating. Does anyone fit a multinomial MRQAP (#7)? Is the GPU path worth its weight once the permutation loop stops rebuilding the design matrix (#11)?
Tracking issue for the five extensions parked on branches while the architecture settles: #7, #8, #9, #10, #11.
Why they were parked
Two of the roadmap's phases were fighting each other. Merging
QAPglm()andQAPcss()into one skeleton means reconciling every branch point twice, and widening the test suite over duplicated code doubles that cost rather than removing it. Four of those branch points served model types that no test can reach, no runner can exercise, or no user can select.Nothing is deleted. Each extension is on its own branch at
bc96783, and each strip ondevelopis one commit thatgit revertreinstates.What it bought
Suggestsfell from eight modelling packages to three:MASSships with R, so the marginal install cost islme4andpscl. The core — gaussian, binomial and poisson throughstats::lm()andstats::glm(), HC3 errors, both permutation schemes, and all the network shapes — needs nothing fromSuggestsat all.Proposed order back
#8 comes first because it is a naming decision with no engineering behind it, and because a bar in the formula currently means one thing only, which is a state worth either keeping deliberately or leaving deliberately.
#7 should be rewritten against the merged engine rather than reverted onto it. Its 21 forks are exactly what the merge is trying to remove.
Open question
Two of these are candidates for dropping rather than reinstating. Does anyone fit a multinomial MRQAP (#7)? Is the GPU path worth its weight once the permutation loop stops rebuilding the design matrix (#11)?