[fix]: Rceattle converter publishes wrong composition, age and uncertainty values #405

Found while tracing the Rceattle branch for #400. These are pre-existing lines that the repair in #401 makes execute for the first time, so none is a regression — but the first three publish wrong numbers rather than missing ones. Measured on a GOAatf fit (1 species, 2 sexes, nages = 21, 1961–2050, 59 composition rows).

1. Predicted composition is in the age dimension for length comps. Predictions come from quantities$age_hat, predicted catch at true age, and are used for every row of comp_data. 14 rows are age comps (42 bins) and 45 are length comps (52 bins), so 2,340 of the 2,928 observed values get age-dimensioned predictions. quantities$comp_hat is the aligned quantity — each observation's own dimension, one-to-one with comp_data, rows summing to 1 — and its observed counterpart is quantities$comp_obs. Note filter(estimate != 0) hides the mismatch, because age_hat[, 43:52] is exactly 0 on a length row. The two series are also incomparable in scale even on the age rows: observed row sums run 194–3.6e9 against predicted 1.5e6–2.2e6.

2. Composition rows publish an age the model does not have. Every comp row is joint-sex (Sex = 3), so a comp vector is sex 1's bins followed by sex 2's. Both go into age as bin ordinals, giving age up to 52 for a 21-age model on 2,928 of 5,406 rows. Because age0_length1 is dropped, nothing in the output records which rows are length bins or which sex a bin belongs to, so it cannot be repaired downstream.

3. uncertainty_label = "log_sd" is not in the standard vocabulary. 214 rows carry it, and filter_data() branches on grepl("sd", ...), so it takes the sd branch and builds estimate ± uncertainty on the natural scale. An index observation of 1,112,215 with a log-scale sd of 0.064 gets a band of ±0.06 instead of 981,047–1,260,920. The SS3 vocabulary is sd/se/std/stddev. Separately the predicted row inherits the observation's log_sd, so index_predicted publishes observation error as estimation error.

4. plot_index() resolves to the series with no years. The output carries log_index_hat (module sdrep, 17 values, all year = NA) and index_observed/index_predicted (module data_list, correct years). The module disambiguator picks sdrep, so the figure gets 17 points at year = NA. The existing guard label %notin% c("index_hat", "catch_h") does not list log_index_hat.

5. Projection years publish a catch of 0. catch_hat is only computed on hindcast rows, so the 32 projection years carry catch_predicted = 0 against catch = NA — a landings figure shows the fishery stopping in 2019.

6. A held-out observation publishes a negative year. Rceattle's convention for carrying an observation but excluding it from the likelihood is a negative Year, which converts to year = -1984 with era = "time", putting a point at −1984 on any index figure.

Related: only sdrep and data_list reach the output. The converter emits those two modules, so quantities never reaches the standardized output — that is where n_at_age, biomass_at_age, m_at_age, f_at_age, sel, b0, flimit and ftarget live, ~1.6M rows and 92 labels on this fit. Consequently plot_landings(), plot_catch_comp() and plot_abundance_at_age() fail on converted Rceattle output (landings not found., abundance not found.) while plot_biomass(), plot_spawning_biomass(), plot_recruitment(), plot_index() and plot_recruitment_deviations() return figures. It is also why 31 of the 35 unit cells corrected in #401 are currently dormant, and why the catch_hat → landings mapping in rceattle_var_names.csv never fires.

Happy to take any of these on — I held off in #401 because each is a decision about what the standard means rather than a repair.