[jiapingW]

Superseded by the following smaller PRs. The original config branch and its commits are preserved.

Config stack (merge in order):

• [1/10] Unify draft config loading #1042: Unify draft config loading
• [2/10] Add immutable draft configuration types #1043: Define typed draft model options
• [5/10] Move draft models and layers to canonical paths #1044: Organize draft models and reusable components
• [9/10] Version draft settings and separate serving serialization #1045: Add canonical draft settings and serving serialization
• [10/10] Consolidate draft settings in training recipes #1046: Migrate draft settings and native recipe names

Independent PRs, both based on main:

• Preserve capture layers for unchanged DFlash depth #1047: Preserve capture layers for unchanged DFlash depth
• Extract the Qwen3.5 native MTP checkpoint adapter #1048: Extract the Qwen3.5 native MTP checkpoint adapter

Each PR has its own validation. The combined split passed 507 CPU regression tests (55 skipped). No GPU training run was performed for this split.

The MTP adapter and model relocation overlap at the old MTP paths; their integration was resolved and tested separately. See #1044 and #1048 for the two-file relocation details.