Add backtest checkpoints and backtest_combine(); require newer depend…

DESCRIPTION

@@ -1,7 +1,7 @@
 Package: tbl.now
 Type: Package
 Title: Tidy Data and Workflow Layer for Epidemic Nowcasting
-Version: 1.1.0
+Version: 1.1.1
 Authors@R: 
     c(person("Rodrigo", "Zepeda-Tello", , "rzepeda17@gmail.com", role = c("aut", "cre"),
            comment = c(ORCID = "0000-0003-4471-5270")),
@@ -60,19 +60,19 @@ Imports:
     cli,
     generics,
     grid,
-    dplyr,
+    dplyr (>= 1.1.0),
     ggplot2,
     lifecycle,
     lubridate,
     methods,
     pillar,
-    rlang,
+    rlang (>= 1.0.0),
     S7,
     scales,
     stats,
     tibble,
     tidyr,
-    tidyselect,
+    tidyselect (>= 1.2.1),
     utils
 URL: https://rodrigozepeda.github.io/tbl.now/,
     https://github.com/RodrigoZepeda/tbl.now

DEVELOPMENT_SKILL.md

@@ -181,6 +181,46 @@ S7 class names can contain `::`, which S3 method names cannot represent. Use
 the existing `.onLoad()` registration pattern where required. Re-export
 `generics::tidy`; do not create a competing generic.
 
+## Backtests, checkpoints and combining
+
+`nowcast_backtest()` fits every (engine, date) pair, dates outer and engines
+inner. Two features are built on that order and must keep working when the
+backtest changes:
+
+- **`backtest_combine()`** joins backtests run separately -- other engines on
+  the same dates, or the same engines on new dates. Combining what one call
+  would have produced must return that call's object: same rows in the same
+  order, same `methods`, `now_dates`, `timings` and `truth`, differing only in
+  `elapsed_seconds`. Test that equality, with strata and a grouped `tbl_now`.
+  A (method, `now`) fit in two backtests aborts, a success replaces a failure of
+  the same fit, and backtests that differ in event date, strata, `truth_axis`,
+  `truth_type`, `keep_draws` or quantile levels are refused. If you add a field
+  or table to the `nowcast_backtest` object, teach `backtest_combine()` to
+  combine it.
+- **`checkpoint_file`** lets an interrupted backtest resume. Only the main R
+  session touches the file; `future` workers return their fits and never see
+  the path, so every plan is safe. Never write from a worker, and never write
+  in place: write a temporary file beside it and rename. The file stores a
+  fingerprint of everything a fit depends on (data, `seed`, `keep_draws`, truth
+  settings, each engine by label); a mismatch aborts rather than mixing
+  backtests. A new argument that changes what a fit computes must join the
+  fingerprint. Failed fits are stored but retried on resume. A resumed backtest
+  must equal an uninterrupted one.
+
+Test both under `parallel = TRUE` with a sequential plan and, where the
+installed package is the code under test, real `multisession` workers.
+
+## Dependencies and the environment
+
+Test on the versions CI uses (the current R release and current CRAN
+packages), not an old container. A failure that only an old R or old dependency
+shows is a missing minimum: declare it with `>=` in `DESCRIPTION` (`Depends`,
+`Imports` or `Suggests`) rather than working around it in the code, and say so
+in the commit message. Give a minimum to every package whose newer function or
+argument the code uses (`rlang::hash()` needs rlang 1.0.0, `.by` in dplyr needs
+1.1.0). Compiled packages built under an older R do not load in a newer one;
+reinstall them after upgrading R.
+
 ## Summaries and diagnostics
 
 `validate_tbl_now()` and `diagnose()` share `.tbl_now_findings()`. Add structural

NAMESPACE

@@ -114,6 +114,7 @@ export(align_weeks)
 export(as_tbl_now)
 export(as_tibble)
 export(autoplot)
+export(backtest_combine)
 export(cases_per_date)
 export(censor_reporting_delays)
 export(censor_reporting_delays_above)

NEWS.md

@@ -1,3 +1,32 @@
+# tbl.now 1.1.1
+
+## `simulate_batch()` works on older R versions
+
+`simulate_batch()` aborted with "Every report date is closed" on older R
+versions (seen on R 4.3.3) whatever `closed_dates` was, because `setdiff()`
+returns plain numbers there and `match()` cannot compare them with dates. The
+open dates are now recovered from their numeric values, which is correct on
+every R version.
+
+## `nowcast_backtest()` can resume, and backtests can be combined
+
+`nowcast_backtest()` gains `checkpoint_file` (#90). Each finished
+(engine, date) fit is saved to that file; running the same call again refits
+only the fits that are missing, and a fit that failed is retried. The file
+records a fingerprint of the data, `seed`, `keep_draws`, the truth settings and
+each engine, and a call that differs from it aborts rather than mixing fits from
+two backtests; new `now_dates` and new engines are added to it. Only the main R
+session writes the file (through a temporary file and a rename), so every
+`future::plan()` is safe with `parallel = TRUE`, which then runs and saves the
+fits in waves of `future::nbrOfWorkers()`.
+
+The new `backtest_combine()` joins backtests that were run separately (#91):
+different models on the same dates, the same model on new dates, or both. Combining
+what one call would have produced gives the same object as that call. It aborts
+when a (method, date) fit appears twice or when the backtests differ in strata,
+`truth_axis`, `truth_type`, `keep_draws` or quantile levels, and
+`only_common_dates = TRUE` keeps only the dates every method has.
+
 # tbl.now 1.1.0
 
 ## Breaking: backtest methods are compared on the targets all of them scored