feat(assistant): add background tasks, global notifications and context exports

[TNT-Likely]

Background

Assistant work needed clearer controls, durable background visibility, readable mobile layouts, and a reusable summary when continuing a conversation elsewhere. Notifications were limited to assistant tasks and required opening a heavy modal. Conversation titles and export history also needed to remain useful as research accumulated.

Changes

Assistant task lifecycle and streaming

• Add stop, conservative retry, progress, elapsed time, recovery actions, and restoration of the latest task after returning to a conversation.
• Preserve task and message identity during retry; prevent replay after a tool invocation has begun and serialize approval decisions.
• Discover background work without browser markers and reconcile interrupted tasks on server startup.
• Restore the original application lifespan, prevent duplicate startup and schedulers, and initialize the first MiniRacer context before concurrent market data workers.
• Memoize historical Markdown, batch live tokens, reduce durable SSE polling latency, and follow growing content while preserving the user's manual scroll position.
• Remove display-only trailing whitespace during streaming and translate live tool labels.

Global notification center

• Unify assistant outcomes and approvals, Agent reports, and price alerts into persisted notification events and receipts.
• Keep read state independent of pending approvals. Add source/status filters, stable pagination, observed-ID read boundaries, archive protection, and source navigation.
• Migrate legacy assistant notifications while preserving read/resolved state and suppressing historical toasts.
• Use a lightweight bell preview for unread or pending items. Desktop hover opens the preview; mobile uses click. Hide the redundant global task indicator on the assistant page.

Conversation history and portable exports

• Show running/queued/approval indicators in conversation history. Generate a bounded title after the first successful reply and support manual rename without overwriting it later.
• Generate a Markdown context summary as a durable background job with HTTP 202 submission, status reads, resumable checkpoints, startup recovery, bounded model calls, and completion/failure notifications.
• Export goals, constraints, facts, decisions, progress, unresolved questions, and next steps. Cover saved messages only and clearly identify snapshots captured during an active task.
• Reduce duplicate supplemental facts and source metadata while preserving saved message content, dated evidence, and source links. Pack fragments according to the configured context budget.
• Reuse identical snapshots, retry failed jobs from saved progress, and open the exact export from its notification.
• Keep export history in each conversation's actions menu, with metadata-only pagination, status polling, detail preview, copy, and Markdown download.

Interface improvements

• Use available mobile height and opaque conversation history surfaces.
• Prioritize the assistant navigation entry on desktop and mobile.
• Add a shared soft-face default avatar while preserving custom uploads and falling back when an uploaded image cannot load.

Validation

• Backend: pytest tests/ -x -q -rs, using a temporary SQLite engine/session factory before collection to protect the development database: 885 passed, 3 skipped, with four consecutive full-suite passes after the SSE fixture correction.
• Market data package: pytest packages/marketdata/tests -q: 188 passed.
• Agent runtime package: pytest packages/pan-agent-runtime/tests -q: 43 passed.
• Frontend: cd frontend && pnpm exec vitest run: 177 passed across 41 files.
• Frontend: pnpm run check:i18n: passed for 126 migrated files.
• Frontend: pnpm run build: TypeScript and Vite build passed.
• git diff --check: passed.
• Add 19 backend export cases and 8 frontend export/avatar/API cases covering input contracts, model-output validation, snapshot reuse, metadata history, worker deduplication, exact notification targets, checkpoint retry/recovery, shutdown, deletion, transaction rollback, HTTP 202, polling cleanup, and StrictMode submission.
• Fix an existing log-SSE test fixture that shared one native SQLite connection across a reader thread and a writer. Use separate connections and an explicit baseline-read signal; the targeted 5-test suite passes.
• Previously recorded browser checks covered task controls/background navigation, opaque mobile history, responsive composer height, notification routing, title/rename, and controlled streaming/scroll fixtures. The latest export/history/avatar changes were verified by isolated automated tests and build; no new real-provider or physical-device validation is claimed.

Boundaries and risks

• Task execution and export workers remain single-process asyncio implementations. Assistant retries remain conservative around tool side effects; exports perform read-only model summarization and can resume saved progress.
• Notifications use the existing shared-installation authentication model. They are an in-app inbox, without OS push or per-user/tenant delivery.
• Export content is sent to the configured summary model. Model quality and real-provider latency were not re-evaluated in this merge check. Saved snapshots and history are deleted with their conversation.
• Exports retain sequential rolling summarization, bounded per-call/job timeouts, and a 32-part limit. Large conversations can require multiple model calls.
• Local checks use Python 3.13 and Node 20.20.2; the release workflow uses Python 3.11 and the Docker frontend build uses Node 24.14.0. Local builds pass with the existing Node engine and chunk-size warnings.
• Three local skips cover an unavailable optional OpenTelemetry SDK and two Windows-only checks. Existing Pydantic/Starlette deprecation warnings remain.
• Local .docs/ records and AGENTS.md are excluded from this pull request.

Follow-up

• Verify provider summary quality and latency during user acceptance, and verify keyboard/scroll behavior on physical mobile devices.
• Introduce durable worker leases, idempotent tool contracts, and explicit multiworker ownership before extending task recovery across external side effects or multiple processes.