AI spam classification: default to 2 KiB of text and a +2.0 cap on the model's tag, as calibrated for low-end CPU-only instances

The calibration harness (tests/src/system/ai_calibration.rs, ignored) sends
what the classifier sends to a real local model and scores its answers with
the classifier's parser.

crates/features/src/ai/limits.rs

@@ -31,11 +31,11 @@ pub struct AiLimits {
 impl Default for AiLimits {
     fn default() -> Self {
         AiLimits {
-            spam_max_added: 5.0,
+            spam_max_added: 2.0,
             spam_max_subtracted: 1.0,
             spam_call_ceiling: Duration::from_millis(20_000),
             max_concurrent_calls: 4,
-            max_content_bytes: 16_384,
+            max_content_bytes: 2_048,
             failure_backoff: Duration::from_millis(60_000),
             user_calls_per_hour: 60,
         }

crates/spam-filter/src/analysis/score.rs

@@ -59,7 +59,7 @@ impl SpamFilterAnalyzeScore for Server {
             // inbuxa: AI-13: the model's tag moves the score only so far, and
             // never discards or rejects on its own
             if inbuxa_features::ai::answer::is_llm_tag(tag) {
-                let (max_added, max_subtracted) = ctx.result.llm_bounds.unwrap_or((5.0, 1.0));
+                let (max_added, max_subtracted) = ctx.result.llm_bounds.unwrap_or((2.0, 1.0));
                 let score = match self.core.spam.lists.scores.get(tag) {
                     Some(SpamFilterAction::Allow(score)) => {
                         inbuxa_features::ai::answer::clamp(*score, max_added, max_subtracted)