fix(data-io): bind lazy datasets into the namespace at load

[lbm364dl]

What was wrong

Confirmed on origin/main with a normally installed package (R CMD INSTALL
into a temp lib, fresh Rscript --vanilla, no library(whep)):

rows in whep::polities: 753
resolve_polity_label: object 'polity_label_aliases' not found
add_polity_code: object 'polity_area_crosswalk' not found
get_polity_geometries: object 'polities' not found
polities in namespace: FALSE

Mechanism, exactly as diagnosed in #641: with LazyData: true, loadNamespace()
lazy-loads data/Rdata.rdb into ns$.__NAMESPACE__.$lazydata, not into the
namespace. Package code is evaluated in the namespace, whose parents are
imports → base → globalenv, so a bare polities can only be found via the
search path — which holds the datasets only once library(whep) has attached
the package env. whep::polities works because :: falls back to the lazydata
environment. All 100 lazy-loaded objects (56 topics) are invisible to package
code in the loaded-but-not-attached state, and 55 of the 56 datasets are
referenced by bare name in R/.

What changed

New R/zzz.R (25 lines, no call sites touched — options 1 and 3 in the issue
would have rewritten ~667 references across 67 files):

• .onLoad() binds each lazydata name that the namespace does not already
  define into the namespace with delayedAssign(). The namespace is still
  unlocked while .onLoad() runs, so this is a legal assignment; sealing
  afterwards just locks the promises.
• The bindings stay lazy — verified with
  rlang::env_binding_are_lazy(asNamespace("whep"), c("polities", "biomass_coefs"))
  → TRUE TRUE right after loadNamespace("whep"), so nothing is eagerly
  loaded and memory behaviour is unchanged.
• Existing namespace objects are never shadowed (setdiff against ls(ns)),
  and identical(whep::polities, get("polities", asNamespace("whep"))) is
  TRUE.
• library(whep) and pkgload::load_all() are unaffected (under load_all()
  the datasets are already in the namespace, so the binder is a no-op).

Classification: mechanical — a packaging/visibility fix. No published value
changes.

Verification

tests/testthat/test_zzz.R: four unit tests of the binder (binds, keeps lazy,
never shadows, no-op without a lazydata env) plus an invariant that every object
in the package's lazydata environment is reachable from the namespace — the
guard that would have caught this.

Before the fix, running the new test file against the installed
(unpatched) package, i.e. the R CMD check state:

── Failure ('test_zzz.R:64:3'): every shipped dataset is reachable from the namespace ──
`missing` (`actual`) not equal to character() (`expected`).
     actual                | expected
 [1] "amg_h_by_input_type" -
 [2] "animals_codes"       -
 [3] "biomass_coefs"       -
 ... and 90 more ...
[ FAIL 5 | WARN 0 | SKIP 0 | PASS 0 ]

After the fix, same installed-package run:

[ FAIL 0 | WARN 0 | SKIP 0 | PASS 9 ] Done!

and the issue's reproduction script:

rows in whep::polities: 753
resolve_polity_label: OK
add_polity_code: OK
get_polity_geometries: OK
polities in namespace: TRUE

Also run:

• devtools::test(filter="zzz") → FAIL 0 | PASS 9
• devtools::test(filter="polities|gapfilling|datasets") → FAIL 0 | SKIP 1 | PASS 1080
• rcmdcheck::rcmdcheck(build_args="--no-build-vignettes", args=c("--no-tests","--ignore-vignettes"))
  → 0 errors | 0 warnings | 0 notes (with _R_CHECK_FORCE_SUGGESTS_=false;
  archive and RSQLite are not installed on this machine)
• air format ., lintr::lint() on both changed files → no lints
• No roxygen, no new exported object, no new NSE symbol, so man/,
  _pkgdown.yml and utils::globalVariables() are untouched.

Closes #641.

🤖 Generated with Claude Code