fix(counties): make the cluster and the entry panel agree on a US county

An operator asked why K1SEI showed "Middlesex" in Info (F2) and "Lower
Connecticut River Valley" in the cluster. Two sources: the entry panel has
a callbook answer, a spot carries only a callsign so the cluster derives one
from the FCC licence ZIP through GeoNames — and GeoNames has followed the
Census in replacing Connecticut's counties with the 2022 planning regions.
No award, callbook or log uses those, so every CT station matched nothing:
new county for ever, and counting toward nothing.

Measured against a full ULS import (1 556 444 US callsigns), 23 223 resolved
to a name the USA-CA reference does not contain. Three causes, three fixes:

  - Spelling. "City and County of San Francisco", "Baltimore (city)",
    "Nome (CA)", plus counties renamed since the award list was drawn
    (Kusilvak, Oglala Lakota, the Valdez-Cordova split) and Alaska's four
    "X City and Borough", whose reference codes read "JUNEAUCITYAND" because
    the county-type suffix strip eats the wrong end. Normalised in
    award.USCountyKey, which both sides already go through. 7 515 callsigns,
    no re-download needed.

  - Doña Ana, NM shipped into the reference as "NM/DO̱AANA" — mangled by a
    non-UTF-8 CSV line, a code nothing could ever produce, so that county was
    unwinnable and silent about it. Row repaired, cntygen now refuses such a
    line, and a test makes every one of the 3 102 references reproduce its own
    code from its own name.

  - Connecticut. A planning region is drawn from towns in several counties, so
    no name maps to a name — only the ZIP can resolve it. cmd/ctzipgen builds
    the table from the Census 2020 crosswalk, filling PO-box-only ZIPs from the
    nearest resolved centroid; all 11 ZIPs GeoNames still labels with a real
    county agree with the result. 15 037 callsigns, applied at import, so the
    store now carries a rules version and Settings says when a re-download is
    needed.

Alignment itself is the last piece: a spot now shows the county the station is
logged with when we have one, and falls back to the ZIP-derived county only for
stations never worked.

cmd/cntygen/main.go

@@ -9,6 +9,7 @@ import (
 	"os"
 	"sort"
 	"strings"
+	"unicode/utf8"
 
 	"hamlog/internal/award"
 )
@@ -58,6 +59,15 @@ func main() {
 	sc.Scan() // header
 	for sc.Scan() {
 		line := sc.Text()
+		// The FIPS CSV is not reliably UTF-8, and the one county whose name is
+		// not ASCII (Doña Ana, NM) came through mangled once already — it landed
+		// in the reference as "NM/DO̱AANA", a code no log could ever match,
+		// silently costing that county for every operator. Refuse the row rather
+		// than emit a broken one.
+		if !utf8.ValidString(line) {
+			fmt.Fprintf(os.Stderr, "cntygen: skipping non-UTF-8 line: %q\n", line)
+			continue
+		}
 		parts := strings.SplitN(line, ",", 3)
 		if len(parts) < 3 {
 			continue