feat(background): separate execution results from report delivery

[nonoqing]

Summary

Scheduled runs can finish successfully without a report, but a model acknowledgement such as “remaining quiet” previously still appeared as a reply and could create unread work or notifications. Independent failure events also lacked an OS notification subscription. This change separates execution, report delivery and history-read state, preserving the complete original execution.

• Persist a run/generation-bound delivery decision; accept structured report proposals and adjudicate legacy acknowledgements against task instructions with a tool-free compatibility request. Uncertain or failed decisions retain visible delivery.
• Fold quiet runs into expandable, searchable execution records across Web UI/mobile; preserve earlier unread results and report actual failures. Apply the durable decision to Weixin proactive output.
• Recover history-read failures by resynchronizing from the owning host without rerunning the task. Keep stale requests fenced and reject controller-local fallback for detached history.

Type and Areas

Feature / bug fix. Rust Runtime and Core, session persistence, shared event contracts, Web UI, mobile web, Remote Connect, documentation.

Motivation / Impact

A conditional monitor that has no reportable change stays quiet even when its model writes a silence acknowledgement. Failures remain visible. A quiet 11:00 run cannot replace the unread identity of a 09:00 report. Older jobs/events remain readable; unknown policy strings survive round trips, unsupported descriptors stay visible, and peer policy writes negotiate background_result_delivery_v1.

The current producer is the native scheduler. Detached job completion and other producers are not silently opted into suppression.

Verification

AI-assisted implementation; focused policy, persistence, UI and compatibility coverage. Prior implementation validation passed:

• Core background contracts: 4 tests; cron scheduling/tool contracts: 37; Runtime decisions: 4; DTO serialization: 2.
• Session metadata/storage: 35; Remote Connect contracts: 75; event projections: 13; remote-surface registry: 20.
• Web delivery/unread/notification/API/peer tests: 182; search/timeline/receipt: 20 (one overlapping receipt test); source snapshot safety: 7.
• Real Chrome mobile receipt rendering at 390/768/1280 px, raw detail expansion, visible failures/questions, generation fences; 2 host-stream browser tests with simulated host/relay.
• Web/mobile type checks, i18n audit, Appearance contracts, core boundaries, changed-Rust formatting and diff/object-size checks passed.

Rebased onto upstream/main 96557706b. Rebase verification passed 271 tests across 16 Web/notification/next-input-suggestion integration files, Web type checking, 3 real Chrome mobile/host-stream browser tests, i18n audit and core boundaries. All 4 Core background finalization/persistence/upgrade/bot tests also passed after the rebase. No product-code conflicts remained; the one guide conflict retains both verification commands.

Known full-check limitations: check:web encounters 25 pre-existing typography violations in unchanged design-system files. Local theme:color-audit:all passes the color roots but discovers an unrelated pre-existing untracked MiniApp; that directory is excluded from this PR. No baselines were loosened. Real SSH/Docker, a real peer device, live IM delivery and semantic accuracy against a configured model provider were not exercised.

Reviewer Notes

See the design and compatibility boundaries or 中文说明. Additional adjudication usage is recorded separately; raw task output is retained. There are no dependency or packaging changes.

Checklist

• This PR is focused and excludes secrets, temporary prompts, scratch files and unrelated artifacts.
• Relevant verification and remaining coverage limits are recorded above.
• User-facing strings, docs and locales are updated where applicable.