Follow up grader workflow resolution for mixed runtime modes and multi-worktree sessions

[MacAttak]

Summary

Post-merge follow-up from PR #194.

The resolver changes in evals/framework/grader.py fixed the immediate red CI and stale legacy-bridge issues, but two additional review concerns remain open and should be handled explicitly in a follow-up:

1. Respect configured config.git.runtime.mode before cross-mode fallbacks.

   • Current resolver checks project-visible sessions before legacy fallback regardless of active mode.
   • In mixed-layout repos with stale .specwright-local state, grading could prefer the wrong runtime mode.

2. Resolve the session for the current worktree instead of the first glob match.

   • When worktrees/*/session.json is globbed, the first attached workflow can belong to a different worktree/helper worktree.
   • Grader state/gate checks should prefer the session that matches the current checkout context (for example worktreePath / worktree identity).

Why this is deferred

These are correctness hardening items, but they were not needed to release the workflow-legibility/devex improvement. PR #194 was merged once CI was green and the primary operator-surface/runtime blockers were fixed.

Suggested scope

• tighten resolver order by active runtime mode
• filter project-visible sessions to the current worktree
• add regression coverage for mixed-mode stale state and multi-worktree helper session ordering

Source

• PR fix(sync): guard gone-branch cleanup and operator workflow proof #194 review threads on evals/framework/grader.py