Correct the dcast comment about which CBS buckets duplicate

[lbm364dl]

What was wrong

.select_best_source() in R/build_cbs.R justifies its fun.aggregate
argument with a measurement from #429 — 31,642 duplicated (key, source)
combinations at full range "in areas 206 and 999", plus a dump of bucket 999 in
2010 for wheat holding four distinct territories. That is the recorded evidence
for why sum is the right aggregate, and it is also what a reader uses to
reason about the Rest-of-World bucket. Neither bucket duplicates any more:

• 999: all 62 areas with fabio_code == 999 resolve to the single aggregate
  polity ROW-1850-2025, so even with the fold applied the bucket emits one
  row; and since the un-fold became the default (.unfold_rest_of_world(),
  "all") the members are not folded into it at all.
• 206: The polity-level ROW un-fold splits one bucket into several area labels, so .aggregate_to_polities() stops summing it #563 removed polity_name from .aggregate_to_polities()'s
  grouping keys (b2526238), so a bucket folding several live territories now
  emits exactly one row per key. The issue measured 206 at 60fc6610, which is
  not a descendant of that fix — verified with git merge-base --is-ancestor — so the 206 half of the claim was already stale when filed
  too, not only the 999 half.

Measured on this branch, off polity_area_crosswalk alone:

real crosswalk    -> buckets folding >1 live polity in some year 1850-2023: 206
                     (2 polities, from 2011: Sudan, South Sudan)
fabio_code == 999 -> polity_code: "ROW-1850-2025" (single value, 62 areas)

What changed

Comment only, no behaviour change. The measurement is kept but marked as
history, with the two reasons it ended, and the argument for fun.aggregate
restated so it no longer depends on a live duplicate: dcast()'s fallback is
length() applied to every cell, so any future duplicate anywhere — a
bucket, a re-mapped item code, one key reported in two units, since key_cols
excludes unit — would silently turn the whole table into row counts.
fun.aggregate is untouched.

Two tests pin the property instead of a measurement:

• test_build_cbs.R — a duplicated (key, source) pair comes back summed,
  and an unduplicated key in the same table keeps its tonnes.
• test_polity_folds.R — 206 is the only polity_area_code folding more than
  one live polity in any year, and every fabio_code == 999 area resolves to
  ROW-1850-2025. Crosswalk-derived: no pins, no network.

How it was verified

A comment fix cannot have a test that fails before it, so both tests are shown
to be sensitive by mutation instead.

fun.aggregate deleted from the dcast() call:

Failure ('test_build_cbs.R'): a duplicated (key, source) pair is summed, not counted
Failure ('test_build_cbs.R'): a duplicated (key, source) pair is summed, not counted
[ FAIL 2 | WARN 9 | SKIP 0 | PASS 154 ]

(the wheat cell becomes the count 2 instead of 3,347,000, and the unduplicated
maize cell becomes 1 instead of 500 — the global-fallback bug itself)

Crosswalk assertion, with one ROW member given back a live national polity
inside bucket 999:

real crosswalk:    206
mutated crosswalk: 206 999

Restored, both files green:

devtools::test(filter = "build_cbs|polity_folds")
[ FAIL 0 | WARN 19 | SKIP 0 | PASS 274 ]

air format . run; lintr::lint() clean on all three changed files (the repo
.lintr keeps commented_code_linter on, which the first draft of both test
comments tripped). No roxygen, no exports, no NSE symbols, so no
document() / _pkgdown.yml / globalVariables() change. No NEWS.md entry:
no user-visible behaviour changes.

Classification: mechanical — documentation accuracy plus two assertions; no
number moves.

Closes #557.

🤖 Generated with Claude Code