Remove internal note references from train_config.py

Comments only: references to internal notes are removed or reworded. No configuration value, weight or config.json changes.

attnconv_b7_qwen3_4b/step_2616/train_config.py

@@ -115,14 +115,14 @@ The consequence for reading the result: the step-1 paired gate licenses
 the four corrections bought ON TOP OF the attention head -- is still answerable, just not
 from the init: it comes from the held-out paired eval, which compares FINISHED checkpoints
 and never touches initialisation.  That is the ruler the verdict rests on anyway; the
-shared init is worth exactly one step (see [[additive-module-init-trap]]).
+shared init is worth exactly one step.
 
 EPOCH COUNT IS THE OPEN DECISION.  Written at 1 epoch so it reproduces the ladder's unit
 and is directly paired against every arm already on it.  A longer run CONTAINS this one:
 checkpoint at 2616 and the 1-epoch verdict comes out of it for free, which is why there is
 no reason to run a separate 1-epoch arm first.  If the epoch count changes, change
-`max_train_steps` AND `lr_schedule_steps` together -- see [[lr-schedule-confound]] for the
-arm that was confounded by exactly that field being left to a fallback.
+`max_train_steps` AND `lr_schedule_steps` together -- an earlier
+arm was confounded by exactly that field being left to a fallback.
 """
 import os
 

attnhead_b7_qwen3_4b/step_2616/train_config.py

@@ -45,7 +45,7 @@ because o_proj=0 independently starves the attention branch (14 tensors) for any
 A null result could not then have been read as "the head does not help" rather than "the
 gate never opened".  Nothing is zero-initialised now; instead out_norm's weight is set so
 the step-0 bias matches the magnitude of the vanilla markov bias it replaces (~0.064
-measured against 0.069), which is a scale choice and not a gate.  Read a 1-epoch result with that in mind: [[oneepoch-external-anchor]] measured
+measured against 0.069), which is a scale choice and not a gate.  Read a 1-epoch result with that in mind: an external 10-epoch reproduction showed
 that a change which only pays off late reads as null at one epoch.
 
 SERVING COST (measured, refine_exp/bench_regrounding_*.py, H100, CUDA graph):
@@ -132,7 +132,7 @@ train = dict(
     # WRITTEN OUT, not left to a default.  An arm on this ladder was once compared
     # against a baseline whose cosine schedule annealed to zero while its own stopped at
     # 99.6% of peak, because one config omitted this field and fell back to a different
-    # value.  See [[lr-schedule-confound]].
+    # value.
     lr_schedule_steps=2616,
     max_grad_norm=1.0,
     sharding_strategy="no_shard",