[Phase 3] Shared trust-aware session preparation #607

Parent

• Part of Built-in local inference: /local with llama.cpp as the default backend #602
• Accepted plan: plans/602-built-in-local-inference.md, Phase 3
• Integration target: feat/602-built-in-local-inference (never main)

Scope

Unify TUI/print startup, resume, replacement, and candidate-first model switching around one staged preparation/adoption path.

Follow the complete Phase 3 work, boundaries, deterministic tests, security constraints, documentation requirements, and non-goals in the accepted plan. Historical PRs #352 and #417 are evidence only; do not reuse their branches or runtime assumptions.

Dependencies

• Phase 0 child (created in this portfolio run)
• Phase 1 child (created in this portfolio run)
• Phase 2 child (created in this portfolio run)

Acceptance / exit criteria

• Dynamic startup works in TUI/print; project extensions never load before trust; provider-aware atomic history, idle-only transitions, rollback, ownership, and cleanup are deterministic.
• Preserve static/single-model behavior and all architecture/security/non-goal constraints in Built-in local inference: /local with llama.cpp as the default backend #602.
• Add focused tests and required installed/published docs or development notes for this phase.
• Phase-focused uv run pytest, uv run ruff check ., uv run ruff format --check ., and uv run mypy pass.
• Draft PR targets only feat/602-built-in-local-inference.

Authorization

Created under the maintainer's recorded standing authorization for accepted #602 phases 0–7. This does not authorize merge to main, release publication, closing #602, branch deletion, or material scope expansion.