feat(phase8 #80 D8.6 chunk-2): wire UIMessageStore.write live + drop snapshot_assembler / agent_artifact

[earayu]

Summary

Task #80 D8.6 hard-cut chunk-2 per PM Option 2 lock (msg=d916b44a) — A+B in one PR:

• A Read-side artifact-fallback removal: delete snapshot_assembler.py + tests, rip out assemble_parts_from_artifacts / extract_error_text from services.py:get_turn_snapshot and chat_service:_build_v3_chat_history.
• B Write-side live wire-in: build UIMessageStore.write into the agent runtime end-of-turn (compose UIMessage(parts=[TextPart, *(SourceUrlPart+DataCitationPart per ref)])), delete ArtifactService / agent_artifact table / /agent/artifacts/{id} route / AgentArtifactEnvelope / AgentArtifactType and migrate 4 readers (HistoryWriter, chat_completion_service, evaluation/worker, snapshot endpoint).

agent_timeline_event deletion is chunk-3 (out of scope per PM lock).

Diff

 16 files changed, 472 insertions(+), 968 deletions(-)
 -site-net delete: 496 LOC

PM 3 review-face gates (msg=d916b44a + msg=c84e9fa4)

1. UIMessageStore.write is wired into live runtime (not test-only)

• aperag/domains/agent_runtime/runtime.py:436+: end-of-turn now calls await self.uimessage_store.write(turn_id=..., chat_id=..., message=UIMessage(...)) with parts composed by _compose_assistant_parts(turn_id, answer_text, references).
• AgentRuntimeTaskManager.__init__ constructs the singleton via UIMessageDbOps(session_factory=get_async_session) + AgentRuntimeRedisStore(); same instance is passed to TurnService, HistoryWriter, and PydanticAIRuntime.
• runtime_manager.uimessage_store is exposed as a public attribute so chat_completion_service and evaluation/worker can read the canonical store directly.

2. Read path still complete after artifact-fallback delete

• Snapshot endpoint (services.py:get_turn_snapshot) reads await self.uimessage_store.read(turn_id) — answer text + citations come back as TextPart + SourceUrlPart + DataCitationPart (D8 §2 byte-equal). FAILED error_text falls back to the AgentTurn row's error_message.
• Chat history endpoint (chat_service.py:_build_v3_chat_history) reads canonical UIMessage per-turn, lazy-builds the store with the same factory pattern as chat_title_service (chunk-1).
• OpenAI-compat completion (chat_completion_service._build_completion_content) projects TextPart + DataCitationPart into the existing answer + DOC_QA_REFERENCES + json envelope shape.
• Evaluation worker reads the same canonical store.

3. Write set strictly disjoint from chunk-3

• Zero touch to agent_timeline_event / EventService.append_event / redis_store.append_event / get_events_after. SSE streaming + reload mechanism unchanged.
• All 18 emit() call-sites in runtime.py retained (turn lifecycle + tool lifecycle + state.changed + text.delta).

Caller sweep

Full grep across aperag/ + tests/ for AgentArtifact* / artifact_service / create_artifact / query_agent_artifact* / snapshot_assembler / extract_error_text / assemble_parts_from_artifacts / answer_artifact_id / reference_bundle_artifact_id:

• Zero remaining Python call-sites in production code or tests.
• Remaining hits are docstring narratives explaining the removal + 1 bash comment in tests/e2e_http/scripts/run_chat_collection_flow.sh:164 (historical context).

Migration

New alembic head d8e6c2b4f1a9 (revises c8f2d34a51e7):

def upgrade():
    drop_index("ix_agent_turn_reference_bundle_artifact_id", "agent_turn")
    drop_index("ix_agent_turn_answer_artifact_id", "agent_turn")
    drop_column("agent_turn", "reference_bundle_artifact_id")
    drop_column("agent_turn", "answer_artifact_id")
    drop_index("ix_agent_artifact_turn_id", "agent_artifact")
    drop_index("ix_agent_artifact_artifact_type", "agent_artifact")
    drop_index("idx_agent_artifact_turn_type", "agent_artifact")
    drop_table("agent_artifact")

Downgrade reconstructs the table + columns + indices for rollback symmetry.

Tests

• Rewrote test_agent_runtime_v3.py (3 snapshot tests + history writer test + view test) using _FakeUIMessageStore.
• Rewrote test_chat_service.py (3 history tests) using _FakeUIMessageStore.
• Rewrote test_chat_completion_service.py (3 OpenAI-compat tests) using _FakeUIMessageStore.
• Rewrote test_agent_runtime_openapi_contract.py to assert /agent/artifacts/{id} route + AgentArtifactEnvelope schema are absent from FastAPI OpenAPI spec.
• Added test_agent_runtime_views_no_artifact_route regression guard.
• Full unit suite: 989 passed, 29 skipped, ruff + format clean.

Out of scope (chunk-3, separate PR)

• agent_timeline_event table removal
• Replay/reload semantic change (Redis-only vs UIMessage replay)
• EventService.append_event DB write removal

Test plan

• uv run pytest tests/unit_test/ -q — 989 passed
• uvx ruff check + format — clean
• Caller sweep for legacy artifact symbols → zero Python call-sites
• Migration upgrade/downgrade symmetry verified