fix: accumulate domain loss predictions instead of overwriting

[zhpjunfei]

What

修复 PEPNet._select_domain_task_output 中 domain 预测值累积的 bug。

Why

new_predictions 初始化为 defaultdict(list)，但旧代码使用 = 赋值单元素列表，导致每次循环覆盖之前的值。最终只有最后一个 (domain_index, value) 参与排序和 stack，使得多 domain 场景下的 torch.gather 选择结果错误。

How

将 new_predictions[tower_loss_name] = [...] 改为 .append(...)，正确累积所有 domain 的预测值。

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

张峻飞 seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[github-actions]

Good catch on the overwrite. One thing worth handling in the same PR: now that the list actually accumulates across domains, the sorted(..., key=lambda x: x[0]) a few lines below (line 196) sorts on domain_index as a string (it comes from rsplit("_", 1)[1]). For ≥10 domains this gives lexicographic order — "10" sorts before "2" — which then misaligns the stacked tensor with the integer index from batch.labels[self._domain_input_name] passed to torch.gather. Before this fix the list always had one element so the ordering never mattered; this fix exposes it.

Suggest casting the index to int when appending so the subsequent sort is numeric:

Suggested change

	new_predictions[tower_loss_name].append(
	(domain_index, tower_domain_loss_predict_value)
	]
	)
	new_predictions[tower_loss_name].append(
	(int(domain_index), tower_domain_loss_predict_value)
	)

[github-actions]

Nice, targeted fix — the defaultdict(list) + = assignment was indeed silently dropping all but the last domain.

A couple of notes worth considering:

• Test gap: _select_domain_task_output is only reached through loss() / update_metric(), and the existing test_pepnet in tzrec/models/pepnet_test.py only calls the forward pass and asserts output shape. The buggy code path was unexercised, and even if it were, a shape-only assertion wouldn't have caught the overwrite. A small unit test that builds a synthetic predictions dict with ≥2 towers × ≥2 domains and verifies the gathered values match the per-row domainf label would lock this in.
• Latent sort bug exposed by this fix: see inline comment on the changed lines — domain_index is sorted as a string, so it breaks at ≥10 domains. Worth folding into this PR since the fix is what makes the sort path actually matter.