Align checkpoints/eval to 50/100/150 -> 25/50/75/100, keep all 4

max_steps 150->100, ckpt.interval 50->25 with keep_last=4 (all four
checkpoints retained instead of pruning to the last 3), orchestrator.eval
interval matched to 25 so val40 always runs at each retained checkpoint.
Full 500-task test-set eval at each checkpoint runs standalone after the
weights save, not wired into the live RL loop -- keeps training fast and
keeps the 'test set touched once for the final report' framing honest.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: JayZenith

configs/arm_a_rl.toml

@@ -1,5 +1,5 @@
 # Arm A: reactive agent GRPO over 212 disjoint MBPP RL train tasks.
-max_steps = 150
+max_steps = 100
 seq_len = 4096
 output_dir = "outputs/arm_a_rl"
 
@@ -41,7 +41,7 @@ type = "zero_advantage"
 enforce = true
 
 [orchestrator.eval]
-interval = 50
+interval = 25
 num_examples = 40
 group_size = 1
 
@@ -65,8 +65,8 @@ lr = 1e-6
 weight_decay = 0.01
 
 [ckpt]
-interval = 50
-keep_last = 3
+interval = 25
+keep_last = 4
 
 [orchestrator.ckpt]
 wait_for_weights_timeout = 60

configs/arm_b_rl.toml

@@ -1,5 +1,5 @@
 # Arm B: matched GRPO plus verified-label prediction CE.
-max_steps = 150
+max_steps = 100
 seq_len = 4096
 output_dir = "outputs/arm_b_rl"
 
@@ -46,7 +46,7 @@ type = "zero_advantage"
 enforce = true
 
 [orchestrator.eval]
-interval = 50
+interval = 25
 num_examples = 40
 group_size = 1
 
@@ -70,8 +70,8 @@ lr = 1e-6
 weight_decay = 0.01
 
 [ckpt]
-interval = 50
-keep_last = 3
+interval = 25
+keep_last = 4
 
 [orchestrator.ckpt]
 wait_for_weights_timeout = 60