Run the tests with cgo off and without the race detector (closes #538)

sneak ruled the race detector out
(#538 (comment)), and on
Linux it was the only reason the tests needed cgo. The test phase and the
browser test binary now build with CGO_ENABLED=0 and without -race, so no
test binary compiles mattn's C SQLite. The test phase keeps the builder
stage's Debian Go image. Its comments, the README and the code comments that
said the suite runs under -race now say what is true, and the README notes
the exception to the vendored REPO_POLICIES.md.

Model: opus-5-5

internal/logfield/logfield_test.go

@@ -21,8 +21,9 @@ const budget = 64
 
 // batchRunes is how many consecutive code points the charge test logs
 // in one value from U+1000 up. Logging each of those on its own line
-// is too slow for the suite under the race detector; 4,096 at a time
-// is 271 batches, each logged on two lines, so 542 lines per handler.
+// was too slow for the suite when it ran under the race detector;
+// 4,096 at a time is 271 batches, each logged on two lines, so 542
+// lines per handler.
 const batchRunes = 4096
 
 // newHandlers are the two handlers internal/logger can install. Time
@@ -152,10 +153,10 @@ func TestEncodedBytes_ChargesAtLeastWhatTheHandlersEmit(t *testing.T) {
 		batches = append(batches, batch.String())
 	}
 
-	// What EncodedBytes charges for each batch. Under -race -cover this
-	// takes longer than logging the batches, so it is worked out once,
-	// by whichever handler finishes logging first, while the other is
-	// still logging.
+	// What EncodedBytes charges for each batch. When the suite ran
+	// under -race -cover this took longer than logging the batches, so
+	// it is worked out once, by whichever handler finishes logging
+	// first, while the other is still logging.
 	charged := sync.OnceValue(func() []int {
 		costs := make([]int, len(batches))