fix(convert_output): repair the Rceattle branch and correct its units

[grantdadams]

What is the feature?

• Fixes fix: RCeattle convert_output #400 — convert_output(model = "rceattle") errored on every Rceattle fit I could try, in both input modes.
• Second commit: all 35 unit-carrying rows in rceattle_var_names.csv were kg or numbers, where CEATTLE reports metric tons and thousands of fish. Groundwork for [feat] Call units from converter, not arguments: B, discard, F #388–[feat] Call units from converter, not arguments: index, landings, projections, total catch #392, since once a figure reads its unit from the sheet a wrong one becomes a wrong axis label. It also unswaps the Flimit/Ftarget labels, which had each other's name.

How have you implemented the solution?

• Wrong element names (index_log_q, catch_hat, deprecated Q_block), positional module selection, file[[ instead of dat[[, comp_data overwritten with catch data, and a cross_join() on the predictions. Per-defect detail with the measured before/after is in the commit messages.
• Verified on GOAatf, bundled with Rceattle: 35 columns, 21,768 rows (was 150,372), including 11,340 selectivity and 3,780 natural mortality rows
• Tests pin the composition and index alignment by value against the fit's own arrays. devtools::test() has 0 failures, with the same 67 warnings as main.
• New 40 KB fixture from the public GOAatf, with prepare_rceattle_output_file.R alongside it so the recipe survives the merge.

Does the PR impact any other area of the project, maybe another repo?

• No other model's path is touched; every code change is inside the model == "rceattle" branch.
• data/example_data.rda regenerates bit-identically, so the update-example-data trigger has nothing to rebuild.
• [fix]: Rceattle converter publishes wrong composition, age and uncertainty values #405 is the companion — six pre-existing defects this repair makes reachable, three of which publish wrong numbers rather than missing ones. Worth reading before merging; I left them alone because each is a decision about what the standard means.
• Only 4 of the 35 corrected unit cells reach the output today (the sheet has no data_list block); the rest are dormant until quantities is processed. Also in [fix]: Rceattle converter publishes wrong composition, age and uncertainty values #405.
• The checks here are at action_required — first PR from a fork, so they need approval to run.

[grantdadams]

I reviewed this before anyone had to and force-pushed a corrected version. What changed, in case you read the first one:

• Three things I'd called defects were not defects. rbind() works once the renames in this PR pad every frame, so I reverted to it — it also catches a renamed Fleet_code, which bind_rows() turned into a silent 36th column. Setting era on the catch frame is cosmetic, since the mutate at the end of the branch already recomputes it. And you already refuse a multispecies fit on the index data, so mine is a second guard, not the first.
• One defect I'd introduced. Selecting modules by name meant a getsd = FALSE fit returned a table with no biomass, SSB or recruitment and still printed "Conversion finished!". It now warns.
• Tests. Mine asserted row counts but no values, leaving the composition fix's central premise unpinned; it's now checked against the fit's own array.
• Also trimmed this description by about 95% and moved the findings to [fix]: Rceattle converter publishes wrong composition, age and uncertainty values #405.

[Schiano-NOAA]

@grantdadams thank you so much Grant! I appreciate you tackling a first review. I plan to do a review tomorrow, but my main checks will be to verify that this works in convert_output and that all of our figures and tables look reasonable as well.

This is a huge help with the changes to Rceattle! I will lyk if I have any questions, but I think I can take it from here.

[Schiano-NOAA]

Is there a condition where sdrep or data_list would not be in names(dat)?

[grantdadams]

sdrep, yes. fit_mod() leaves it out when getsd = FALSE, when estimateMode = 3 (no estimation), for a projection-only fit with a constant-F control rule, and when the Hessian is not positive definite. The output then has no biomass, SSB, recruitment or parameter estimates, so the converter warns. The input data, selectivity and M still come through.

data_list is always on a fit returned by fit_mod(). The check is for a fit that was trimmed or built by hand. Without it the converter can't attach years or read the data, so it stops with a clear message.

[Schiano-NOAA]

@grantdadams This generally looks good to me. A couple questions before I merge:

(1) is there selectivity somewhere in the output file?
(2) is there error for recruitment deviations?
(3) is natural mortality reported anywhere? is it a single point or a time series?

Please lmk where they are located if these values are in the output file. Thanks!

[grantdadams]

@Schiano-NOAA thanks! All 3 are in the Rceattle fit, and my commit adds them to the converted output.

1. Selectivity. It's read from fit$quantities$sel_at_age (fleet x sex x age x year) as selectivity_age, so plot_selectivity() works. The fishery curve for GOAatf peaks around 1.7. That is the model's actual output, because Rceattle normalizes non-parametric selectivity to a mean of 1, not a maximum of 1.
2. Recruitment deviation error. The deviations now have a standard error, taken from the diagonal of sdrep$cov.fixed, and plot_recruitment_deviations() draws the bars. They're also kept when recruitment is fitted as a random effect (random_rec = TRUE), in which case they come from sdrep$par.random.
3. Natural mortality. It's read from fit$quantities$M_at_age as natural_mortality, by sex, age and year. For GOAatf it's fixed at 0.2 for females and 0.35 for males, the same at every age and year. Rceattle can estimate M and let it vary by year, so in general it's a time series. plot_natural_mortality() works with it.
4. Selectivity at length (quantities$sel_at_length) is not converted yet. It is only filled for fleets with Selectivity_dimension = "Length" (all zeros on GOAatf), and the fit stores length bin ordinals, not lengths, unless pop_lengths was supplied. Converting it needs a decision on what goes in length_bins, and a separate module_name so plot_selectivity() does not silently pick age or length when both are present.

Ages on these rows are the actual ages: the array's age dimension is offset by minage. The GOAatf conversion now has 21,768 rows.