fix(lora-worker): default network_alpha to dim/2 to match proven Sindra runs

comfy-dev cross-check: Sindra v1/v2 used alpha=dim/2 (0.5 LoRA scaling),
which produced the validated likeness; the initial alpha=dim (1.0) was a
stronger, unvalidated default. Align the default to the proven value
(operator/per-request can still override). Tests updated (14 green).

services/lora-training-worker/tests/test_invocation.py

@@ -124,8 +124,8 @@ def test_relative_dataset_dir_rejected():
 
 def test_alpha_scales_with_dim():
     _, _, params = build_command(_req(tier="fast"))
-    # fast -> dim 16; alpha = dim * ratio (default 1.0) = 16
+    # fast -> dim 16; alpha = dim * ratio (0.5, matching Sindra) = 8
     argv, _, _ = build_command(_req(tier="fast"))
     assert _flag_value(argv, "--network_dim") == "16"
-    assert _flag_value(argv, "--network_alpha") == "16"
+    assert _flag_value(argv, "--network_alpha") == "8"
     assert params["dim"] == 16

services/lora-training-worker/worker/config.py

@@ -76,7 +76,10 @@ SDXL_HPARAMS = {
     "lr_warmup_steps": "0",
     "train_batch_size_lean": "1",  # 3090
     "train_batch_size_full": "2",  # A6000
-    "network_alpha_ratio": 1.0,  # alpha = dim * ratio
+    # alpha = dim * ratio. 0.5 (alpha=dim/2) matches the PROVEN Sindra v1/v2 runs (comfy-dev
+    # cross-check 2026-07-06) — it's the scaling that produced the validated likeness. alpha=dim
+    # (1.0) is a valid stronger default but wasn't what Sindra used; operator can override.
+    "network_alpha_ratio": 0.5,
     "min_snr_gamma": "5",
     "noise_offset": "0.1",
     "save_precision": "bf16",