`src/megatron/bridge/recipes/common.py` exports five base config
builders that every recipe in the project inherits from:

  - `_pretrain_common`
  - `_sft_common`
  - `_peft_common`
  - `_sft_common_vlm`
  - `_peft_common_vlm`

These were the foundation of every recipe but had **zero direct unit
test coverage**. Recipe-level tests catch model-specific overrides;
they don't catch regressions in the shared base.

Add a focused test file with one class per helper (5 classes,
27 tests total). All 5 helpers are pure config builders with no HF
I/O, so the tests are fast and require no mocking.

Coverage highlights:

  - `_pretrain_common`: documented model=None placeholder, training
    cadence (300k iters / GBS 32 / MBS 2), validation cadence, DDP
    overlap + dist optimizer, HuggingFaceTokenizer placeholder,
    mock-blend dataset, RNG seed 1234, no PEFT, bf16_mixed precision.
  - `_sft_common`: shorter training (1k iters / GBS 128 / MBS 1),
    low LR (5e-6), adam_beta2=0.98, minimal DDP (no overlap by
    default), RNG seed 5678, no PEFT, pretrained_checkpoint slot.
  - `_peft_common`: higher LR (1e-4) than SFT, default LoRA attached
    with the documented standard targets (linear_qkv / linear_proj /
    linear_fc1 / linear_fc2) and dim/alpha 32, training cadence
    matches SFT, RNG seed 5678.
  - `_sft_common_vlm`: overrides LLM SFT — longer training (300k iters)
    with smaller GBS=32 / MBS=2, NullTokenizer (processor-driven),
    HFDatasetConversationProvider with `make_cord_v2_dataset`, DDP
    without overlap, RNG seed 1234, lr=3e-4.
  - `_peft_common_vlm`: inherits LoRA from `_peft_common`,
    NullTokenizer, HF conversation provider, RNG seed 1234.

Tests-only — no production code changes. Locks in the documented
contract so future refactors of the inheritance chain (e.g. adding a
field, changing a default) fail loudly rather than silently changing
the defaults every recipe inherits.

Signed-off-by: lonexreb <[email protected]>