fix(tui): make edit-last-turn boundaries authoritative

[qiuYliangM]

Background

Op::EditLastTurn previously located the cut with a raw role == "user" scan even though tool results and runtime-owned control messages use that role for provider compatibility. A tool-using turn could therefore retain the old prompt and tool round-trip. The original fix separated editable prompts but still skipped unsupported real-user content and emitted only an advisory error on preflight rejection, which could leave embedded hosts busy or persisting an optimistic edit.

Changes

• Centralizes edit target selection as editable, unsupported real-user content, or missing.
• Excludes tool results and structurally trusted non-authoritative runtime messages, including current and legacy turn-metadata layouts, without trusting user-authored runtime lookalikes.
• Stops at unsupported real-user content instead of falling through to an older text prompt.
• Emits typed nonrecoverable edit_last_turn_* errors followed by authoritative TurnComplete(Failed) terminals for every edit preflight rejection.
• Keeps the compatibility is_user_turn_prompt export and corrects the prompt_zones::truncate_to comment.
• Adds behavior coverage for tool round-trips, current and future provenance values, legacy metadata, user-authored lookalikes, unsupported content, missing prompts, unchanged history, and zero provider dispatch.

Verification

• cargo test -p codewhale-tui --lib --locked runtime_handoff: 12 passed.
• cargo test -p codewhale-tui --lib --locked edit_last_turn: 3 passed.
• cargo test -p codewhale-tui --tests --locked -- --skip runtime_api::tests::skill_lifecycle_uninstall_removes_installed_skill: library 10,098 passed; Cucumber 14 passed; integration 263 passed; PTY 65 passed, 3 ignored.
• The unfiltered suite has one pre-existing environment-dependent failure in skill_lifecycle_uninstall_removes_installed_skill; its exact rerun fails on the unchanged target as well.
• cargo fmt --all -- --check and git diff --check: passed.
• cargo clippy --workspace --all-targets --locked -- -D warnings was run and is blocked by 27 pre-existing target-branch diagnostics, including private-interface and unused-mut warnings; none points to this pull request diff.

Known risks

• Embedded hosts should consume the failed terminal as the lifecycle authority and use the typed error prefix to roll back optimistic edits. Companion integration: fix(chat): preserve authoritative edit-last-turn boundaries pinvou-agent#320.
• Unsupported real-user content is rejected in place and remains unchanged; this is a deliberate safety boundary rather than an attempt to edit non-text payloads.

No-Issue: this fixes an existing edit replay boundary and lifecycle bug; the background, behavior, and verification are documented above.

[github-actions]

Thanks @qiuYliangM for taking the time to contribute.

This repository is observing a maintainer-managed PR intake gate in dry-run mode, so this pull request is staying open. This note helps maintainers prepare the allowlist before any enforcement is considered.

Please read CONTRIBUTING.md for the expected contribution shape. A maintainer can grant recurring PR access by commenting /lgtm on a pull request.