fix(pipeline): scale DSv4 hybrid indexer loss

[FDecaYed]

What

Teach the pipeline loss-scale selector that the dsv4_hybrid attention variant uses the same DSA indexer-loss auto-scaler as dsa. The parameterized test verifies both names and preserves the caller config.

This two-file correctness hook is deliberately independent: peeling it away keeps the attention implementation PR out of the pipeline-parallel CODEOWNER group.

Provenance

• Supports the DSv4 hybrid attention introduced in [dev] [DeepSeek-v4] Part 1: Hybrid Attention with CSA and HCA #4458.
• Refactors the corresponding pipeline hook captured in [Tracking] DeepSeek-v4 HybridModel merge plan #5795.
• Original attention implementation by the [dev] [DeepSeek-v4] Part 1: Hybrid Attention with CSA and HCA #4458 authors; DSv4 main adaptation by @hxbai.
• Reconstructed against current main; feature cutoff: 2026-07-21.
• GPTModel integration is intentionally excluded.

Testing

• ruff, Black --check, Python compilation, and git diff --check passed.
• Focused pytest execution is blocked because the repository uv environment on this host has no Torch; execution is left to CI.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[FDecaYed]

/ok to test 1b08c42

[FDecaYed]

Closing this thin standalone hook because the indexer-loss scheduling change is absorbed into DSv4 child D2 together with SBHD orchestration and its tests. Tracking PR: #5795.