Time-phase intervention scoping (gap #3)

[alirezabayatmk]

Tracks capability gap #3 from the refactor audit: the post-refactor Intervene step applies at every forward step with no way to scope the intervention to a phase of generation (prompt vs generated tokens, or a specific token index).

Supersedes #27 (@akatief's original pre-refactor ask against the old Location.token_pos API): closed as not-planned-in-that-form. This issue tracks the same intent against the current Pipeline + Step architecture.

Requirement

Let callers scope an Intervene step to:

• phase="prompt": only during prompt processing
• phase="generation": only on generated tokens
• phase="all": every step (current behaviour, default)
• phase=<int>: a specific generation step index

Core use case: activation patching, where you intervene only during the prompt pass and read the downstream effect on generation.

Design notes

• The Intervene step should stay agnostic of how phases resolve to concrete indices. A new phase keyword must not require changes to Step internals: add phases by extending the resolver, not the Step.
• The phase→indices resolver should always return a concrete List[int], never slice. Concrete lists make logged token positions readable and remove slice/list branching at call sites. Design input from closed PR Refactor/37 list location resolution #43 (@adham137).

Porting sources

• Commit fb766b4 (fix(model): scope interventions to prompt processing). Reference only: the API surface has changed.
• Closed PR Refactor/37 list location resolution #43 (@adham137): List[int]-over-slice design point above.

0.1.0-blocking: one of the 5 capability gaps the refactor is missing.