fix(tui_gateway): candidate-inclusive display on warm/live + child-watch resume (#65919 fallout)

[OutThisLife]

Problem

#65919 ("don't wipe messages before final message") made a durable, unconditional change: verification candidates (finish_reason=verification_required / verify_hook_continue) are persisted to state.db, but repair_message_sequence collapses them out of the in-memory model history. That leaves two projections of one session:

• display lineage (verbatim, candidate present) — served by the eager session.resume path and REST /messages.
• model history (candidate collapsed) — what two resume payloads were serving for the user-visible messages.

So switching to a still-live session (or opening a delegated-child watch window) dropped the substantive verification answer that a cold resume of the same session showed. The payloads disagreed about one session — the cross-session "substantive text vanishes on switch" class, and the direct sibling of the resume-duplication regression already fixed in #68149. The #65919 reviewer note called this out pre-merge: the config flag gates only the display/transport callback; the agent-core persistence + collapse ships unconditionally, so every consumer that mixes the two projections must reconcile them — two didn't.

Fix

Reconcile the projections in one place, and apply it to both offending payload builders.

• _reconcile_display_with_live(db_display, in_memory) — pure, DI-testable. Candidate-inclusive DB display is the base; anchors on the last persisted row's (role, text); appends only the in-memory tail the DB doesn't yet cover; trusts the DB display when the tail can't be anchored (DB ahead / diverged).
• _live_visible_history (warm/live reuse) reads the same get_messages_as_conversation(..., include_ancestors=True) the eager resume + REST paths use, reconciled with the fresh in-memory tail — so all surfaces agree by construction while a not-yet-flushed live turn still shows.
• Child-watch (lazy) resume now builds its user-visible messages from the verbatim child-only display projection (repair_alternation=False) instead of the repaired history; the repaired history still feeds live replay. Falls back to the repaired history if the display read fails.

No schema change, no new config, no mid-conversation cache mutation, no desktop/TS change.

Testing

scripts/run_tests.sh tests/test_tui_gateway_server.py tests/tui_gateway/ tests/test_hermes_state.py → 1166 passed, 0 failed.

Coverage (incl. real-SessionDB E2E, not mocks):

• cross-builder consistency: a persisted candidate is collapsed out of the model projection, kept in the display projection, and _live_visible_history equals the eager session.resume display projection (candidate present).
• combined candidate + fully-flushed second turn both appear once.
• lazy child-watch handler test: candidate survives in resp["result"]["messages"].
• unit fallbacks: DB/session_key missing, DB empty (unflushed first turn), DB raising.
• the pre-existing freshness guard test_session_resume_live_payload_uses_current_history_with_ancestors stays green — the reconciliation satisfies both invariants at once.

Scope

Fixes the confirmable backend fallout of #65919 across both remaining offending paths. The other confirmed regression (#68149) is already merged. #68231 (desktop live-render) does not reproduce on current main — apps/desktop/.../interim-sealing.test.tsx:100 ("preserves interim text that the final response does not include") already covers that scenario and is green; that report matches a pre-merge desktop build. This PR additionally guarantees the persisted substantive answer recovers consistently on any refresh/switch.

Relates to #65919, #68149. sweeper:risk-session-state.

[github-actions]

૮ >ﻌ< ა ci review

_{running on 77855ce}

CI timings

CI timings · View job

Wall time 7m54s vs 10m10s (-22.3%). 9 job(s) slower, 9 faster, 2 unchanged.

• Build&Test Docker image / build (arm64, ubuntu-24.04-arm, linux/arm64, type=gha,scope=docker-arm64, type=gha,mode=max,scope...: -151.0s
• Build&Test Docker image / build (amd64, ubuntu-latest, linux/amd64, type=gha,scope=docker-amd64, type=gha,mode=max,scope=do...: -99.0s
• Python tests / e2e: +55.0s
• Python tests / Run tests slice 5/8: -40.0s
• Python tests / Run tests slice 4/8: -35.0s