fix(balances): validate the full signature, surface caveats in cog_explain, memoise coverage windows (#25)

Final-review findings F-1, F-2, F-6, F-8 (plus the F-9 @return reword,
which shares R/balances.R).

F-2: .validate_balance_inputs() checked 2 of cog_balances()' 7 arguments.
years = integer(0) leaked a raw DuckDB 'Parser Error ... AND year IN ()'
with the generated SQL echoed back; govid = character(0) and a non-character
category returned 0 rows with no error at all; recipe = c("a","b") threw
'the condition has length > 1' from inside .validate_recipe_id(). Replaced
with a call to the money verbs' own .validate_verb_inputs() (R/spending.R),
which validates the exact superset needed. Deleted the local copy rather
than extending it -- two validators is how they drift. Placed AFTER
.coerce_govid_input(), because .validate_verb_inputs() asserts
is.character(govid) and a data-frame govid is not unwrapped before that.
This is helper reuse of the same kind as .build_verb_sql()/.attach_per_capita();
the verb still does NOT route through .verb_spendrev().

F-1: falls out of F-2 for free -- the recipe/category mutual-exclusivity
guard lives inside .validate_verb_inputs(). Previously recipe silently
discarded category AND overwrote provenance$category with the recipe label,
so a caller asking for Fund Balances got X40/Z77 insurance-trust holdings
with no trace of the dropped filter.

F-6: cog_explain() rendered every provenance caveat block except
balance_caveats. Since .emit_balance_caveats() fires at most once per
session -- and is routinely consumed by a suppressMessages() call or an
unread knitr chunk -- cog_explain() is the only surface left for a caller
who deliberately audits the result. Added a 'Holdings caveats' section
guarded on !is.null(prov$balance_caveats). Also relabels the cosmetic
'Concept: NA' line on balance results as 'not applicable (holdings are a
stock, not a flow)'.

F-8: the coverage-window query has no govid and no year predicate -- its
answer depends only on the mounted corpus -- yet it scanned all of
balance_long on every call (35% of verb runtime on the fixture, and a
per-request throughput ceiling for cog-api#26). Memoised in
.uscogdata_env$balance_coverage_windows, invalidated by cog_close(), the
same pattern as .uscogdata_env$manifest.

man/cog_balances.Rd

@@ -49,14 +49,17 @@ wide era to the modern one.}
 Tibble with columns `year`, `canonical_govid`, `gov_name`,
   `balance_subtype`, `category`, `amt_nominal`, `codes_included`,
   `aggregate_fallback`, plus optional `amt_per_capita_nominal` and
-  `pop_source` (when `per_capita = TRUE`), and optional `amt_real` and
-  `amt_per_capita_real` (when `adjust_to_year` is set). Amounts are full
-  US dollars.
+  `pop_source` (when `per_capita = TRUE`), optional `amt_real` (when
+  `adjust_to_year` is set), and optional `amt_per_capita_real` (only when
+  **both** `per_capita = TRUE` and `adjust_to_year` are set -- there is no
+  nominal per-capita column to deflate otherwise). Amounts are full US
+  dollars.
 
   Carries a `provenance` attribute matching
   `inst/schemas/provenance-v1.json`, whose `balance_caveats` block reports
-  `not_gaap`, `not_gaap_note`, `coverage_window` (measured per-subtype year
-  extents) and `truncated` (subtypes whose coverage falls short of the
+  `not_gaap`, `not_gaap_note`, `coverage_window` (measured year extents for
+  every balance subtype in the mounted corpus, not only the observed ones)
+  and `truncated` (the observed subtypes whose coverage falls short of the
   requested years). `expenditure_concept`/`revenue_concept` are `NA` --
   holdings are a stock, not a flow, so neither concept vocabulary applies.
 }