fix(tui): preserve chat roles after compaction

[zhuowp]

Summary

When compaction follows a tool result, the saved summary is a user message after the tool exchange. Chat Completions then sends user → assistant(tool_calls) → tool → user, which strict paired chat templates reject. Move only generated summaries before the retained prompt on the outbound wire, place Agent topology checkpoints before the active tool chain, and keep tool-call IDs and persisted history intact. On session restore, retain the position of a generated checkpoint so later turns stay later.

A second compaction must not mistake the runtime-owned topology checkpoint for a real user turn. Keep the actual user prompt and tool pair inside the protected round even when the prompt exceeds the minimum verbatim retention budget; runtime text cannot stand in for the user's prompt in the coverage check.

Testing

• Focused TUI lib tests: 14 Chat wire, 16 runtime handoff, 13 last-round compaction, 4 retained-user, 1 checkpoint restore, and 1 SyncSession restore passed (49 total).
• cargo fmt --all -- --check and git diff --check passed.
• Local Rust 1.96 compilation required a temporary crate-level allow for unrelated existing unfulfilled_lint_expectations; the allow was removed before both commits. Full workspace checks and a live provider call were not run.

No-Issue: reproduced downstream bug on upstream main; this PR provides the report and fix.

[github-actions]

Thanks @zhuowp 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.

[zhuowp]

Independent review, round 1: changes requested.

P1 — crates/tui/src/compaction/last_round.rs still treats the new runtime-owned topology checkpoint as a plain user turn. After the first compaction, a history shaped as [user(question), topology, assistant(tool_use), user(tool_result), summary] makes last_round_start anchor on the topology instead of the real question. The configured verbatim retention budget is clamped to at least 2,000 tokens; a question longer than that budget can therefore be truncated on the second compaction while the survival check passes. Exclude runtime-owned user messages from last-round prompt/coverage selection and add a second-compaction regression using the minimum effective budget and a sufficiently long question. A direct zero-budget unit case may also isolate the algorithm, but is not a production configuration.

No other blocking issue found in this round. The new wire ordering, provenance check, and restore-position behavior otherwise match the intended fix.

[zhuowp]

Independent review, round 2: PASS.

Reviewed the complete diff against upstream main, including the follow-up for the long-user-prompt recompaction case. The protected round now anchors on the real user prompt, runtime-owned topology cannot consume verbatim retention budget or satisfy user-prompt coverage, and mixed genuine user text/tool results remain eligible. Independently ran the focused long-prompt, runtime-text coverage, and existing retention tests; each passed.

No remaining blocking code-review findings. A live strict-provider call and migration of pre-fix sessions remain outside this PR's verification and scope.