fix(agent): stop reference-only compaction handoff from becoming the active turn (salvage #80696, fixes #80622)

[kshitijk4poor]

Summary

A reference-only [CONTEXT COMPACTION — REFERENCE ONLY] handoff can no longer become the active user turn. Fixes #80622: after an assistant response completed with finish_reason=stop, compaction inserted a standalone role=user handoff carrying a Historical Task Snapshot, no new human ask followed, and the agent started a fresh tool-calling turn that resumed already-completed work.

Salvage of #80696 by @HexLab98 — both commits cherry-picked with authorship preserved, plus one follow-up commit finishing the bug class (review findings).

Who hits this

Long tool-calling sessions on any surface (reported from the Linux Desktop app): the user watches the agent silently restart work it already finished, with no new request — and /retry//undo could target the synthetic handoff instead of the real last exchange.

Changes (contributor commits)

• agent/context_compressor.py — reference_handoff_would_drive_next_model_call() guard predicate (mid-tool-loop exempt); is_user_originated_turn() for dispatchers; SUMMARY_PREFIX gains an explicit "if no user message appears after this summary, do nothing" clause with the prior generation byte-frozen in _HISTORICAL_SUMMARY_PREFIXES (verified byte-identical to upstream/main's live prefix — old-summary renormalization safe).
• agent/conversation_loop.py — 3 post-compaction continue sites skip the model call when only a handoff would drive it; a pending real user ask is restored first.
• agent/turn_context.py — reanchor fallback refuses non-user-originated rows.
• run_agent.py — standalone handoffs always project display_kind=hidden.
• tui_gateway undo/retry walkers use the new predicate; 184-line regression test file.

Follow-up commits (finishing the bug class + review folds)

Commit 2 (bug class):

• 5 more sibling sites converted to is_user_originated_turn: rollback.restore truncation (whose comment claimed predicate parity that was no longer true), gateway /retry, CLI /retry + /undo N, and both CLI resume turn counters — legacy sessions persisted pre-fix carry handoffs with no display_kind, so predicate-level exclusion is required (issue's suggested regression Fix terminal interactivity #4, dispatcher-wide).
• Refund ordering at guard site 1: the api_call_count decrement + iteration-budget refund now happen before the break (previously each skipped turn leaked one budget unit for the agent's lifetime).
• Guard-before-reanchor at site 2 so a restored user ask is what the anchor lands on.
• Prompt carve-out: SUMMARY_PREFIX's new clause now states the mid-tool-loop exception the code-side guard already implements, so a literal-minded model doesn't halt an in-flight exchange.
• Honest skip response: the skipped turn returns a short compaction status instead of replaying the previous turn's answer (finalize_turn's Gateway can deliver assistant response without persisting assistant row #43849 chokepoint would have appended the old prose as a fresh assistant row — duplicated in the transcript and re-delivered to the user).

Commit 3 (simplify-pass folds, three parallel reviewers):

• DB picker parity: list_recent_user_messages (hermes_state_search.py) now drops legacy standalone handoffs in its decode loop — SQL can't see them (role=user, no display_kind). Closes the /undo N pairing skew where the in-memory count (new predicate) and the DB soft-delete pick (old predicate) targeted different turns on legacy sessions. Fetches with headroom so the requested limit is honored; 3 new tests, mutation-checked.
• Single scan: the skip guard now runs the drive-check once (was twice — inside the restore helper and again after) and the restore helper's return value decides the verdict.
• _final_response_from_messages → _HANDOFF_SKIP_FINAL_RESPONSE constant (the parameter was unused).
• _handoff_carries_live_user_content delegates to the canonical _strip_context_summary_handoff_message — also fixes the edge where a merged-shaped row with an EMPTY preserved prior tail was wrongly treated as carrying live content.
• Site-level guard test: rollback.restore with a legacy handoff row.

Validation

Check	Result
PR + compression + tui_gateway suites (post-fold)	448 passed
Full tests/agent/ vs upstream/main baseline	identical failure sets (104 pre-existing env failures on both, 0 new)
E2E probe (sole-handoff skip, honest status, restore, mid-tool-loop, predicate, carve-out live-only, strip delegation)	all pass
SUMMARY_PREFIX frozen-generation byte-identity vs upstream/main	verified programmatically
Mutation checks (revert agent/ to pre-fix; no-op the DB-picker skip)	guard tests fail on base, green on stack
ruff (all touched files)	clean

Consecutive-user-row concern from review: the restore append follows the existing _insert_real_user_anchor precedent — repair_message_sequence merges adjacent user rows before every API call, keeping role alternation valid.

Fixes #80622. Based on #80696 by @HexLab98 — commits cherry-picked to preserve authorship.