Support phase-specific model selection

[maxBRT]

Goal

Ship should support selecting different models for different Phases of a Run.

The blog post calls out selecting different models for different phases. Today Ship has a single model setting, and that same model is passed to every Agent invocation.

Why

Implement, Review, and Final have different cost and quality tradeoffs.

Examples:

• Implement might use a strong coding model.
• Review might use a stronger reasoning model.
• Final might use a slower or more careful model for integration review and PR creation.
• Some users may want cheaper models for low-risk tickets and stronger models for Final.

Initial scope

Allow model selection at Phase granularity:

• Implement model
• Review model
• Final model

The current global model setting should probably remain as a default or fallback.

Questions to answer during planning

• What config shape is easiest to understand?
• How should flags override phase-specific YAML config?
• Should --model continue to set all Phases for a single invocation?
• Should there be flags like --implement-model, --review-model, and --final-model?
• How should this interact with different Agent kinds, since model names are provider-specific?

Possible implementation direction

Add config like:

model: "" # default for all phases when phase-specific model is absent
models:
  implement: ""
  review: ""
  final: ""

Then resolve the model in runPhase based on the Phase name.

Acceptance sketch

• A user can configure a different model for Implement, Review, and Final.
• The existing model config keeps working as a fallback.
• CLI flags can override the config in a predictable way.
• Tests verify each Agent request receives the expected model for each Phase.