[Fix] neural: default learn_from_controller off (preserve split learning)

Controller learn tasks (/learnspam, /learnham) historically trained only the
classifiers; neural learned from its own scan-time paths (autotrain, ANN-Train,
autolearn). Default train.learn_from_controller to false so that split-learning
behaviour is preserved out of the box: no NEURAL_LEARN_<rule> learn symbol is
registered by default, the learn task stays the lean classifiers-only pipeline,
and learn_needs_check / the check-pass mask remain dormant. Operators opt a rule
into unified learning per rule with train.learn_from_controller = true.

lualib/plugins/neural.lua

@@ -68,13 +68,15 @@ local default_options = {
     -- means "default to disable_symbols_input" (resolved per-rule at init); set to
     -- false to opt out and keep running the full pipeline for forced learns.
     forced_learn_minimal_scan = nil,
-    -- learn_from_controller: when true, a controller learn task (e.g. /learnspam,
-    -- /learnham) also trains this neural rule -- one corpus push feeds bayes and
-    -- neural. Only symbols-independent rules (disable_symbols_input) are eligible
-    -- (a learn task runs no rule symbol). nil means "default to
-    -- disable_symbols_input" (resolved per-rule at init); set to false to keep a
-    -- fusion rule out of the controller learn path.
-    learn_from_controller = nil,
+    -- learn_from_controller: opt-in (default false) to preserve the historical
+    -- split-learning behaviour -- by default a controller learn task (/learnspam,
+    -- /learnham) trains ONLY the classifiers, and neural keeps training from its
+    -- own scan-time paths (autotrain, ANN-Train, autolearn). Set it to true to
+    -- also train this rule from controller learn tasks, so one corpus push feeds
+    -- bayes and neural. A disable_symbols_input rule then trains from the lean
+    -- learn pipeline; a symbol-dependent rule makes its learn task run a full
+    -- check pass first (its vector reads symbol scores).
+    learn_from_controller = false,
   },
   watch_interval = 60.0,
   lock_expire = 600,

src/plugins/lua/neural.lua

@@ -1789,14 +1789,6 @@ for k, r in pairs(rules) do
     rule_elt.train.forced_learn_minimal_scan = rule_elt.disable_symbols_input and true or false
   end
 
-  -- learn_from_controller defaults ON for symbols-independent rules (so a
-  -- /learnspam corpus push trains them alongside bayes) and OFF otherwise.
-  -- Operators can set it to false to keep a fusion rule out of the controller
-  -- learn path.
-  if rule_elt.train.learn_from_controller == nil then
-    rule_elt.train.learn_from_controller = rule_elt.disable_symbols_input and true or false
-  end
-
   if not rule_elt.profile then
     rule_elt.profile = {}
   end

test/functional/configs/neural_learn_task.conf

@@ -53,6 +53,9 @@ neural {
               max_trains = 1;
               max_iterations = 250;
               classes_bias = 0.0;
+              # opt into controller learning (off by default to preserve the
+              # historical split-learning behaviour).
+              learn_from_controller = true;
           }
           symbol_spam = "NEURAL_SPAM_SHORT";
           symbol_ham = "NEURAL_HAM_SHORT";