Enable shared native private feedback and general rich replies

[huangruiteng]

Native private Chat disabled shared reactions and flattened authored Markdown, while serial delivery let a slow provider readback hold independent queue/control replies. Network identity probes also disappeared from the configuration snapshot on failure, causing the listener to treat a temporary authorization problem as a durable disconnect. This change enables default Get/OnIt feedback, Markdown posts and quiet ordinary admission, bounded independent delivery, and reliable discovery of configured private Apps.

Private reconciliation handles at most four persisted requests concurrently, without an executor backlog. Per-request locks and provider journals preserve exact-ID recovery and no-resend behavior; Core recovers the selected request under its existing Session fence. Admission reuses its initial observation for source reading, then rechecks authority inside the Core fence. Inbox preparation is reused only within one locked reconciliation; identity, grants and original source remain freshly verified before writes.

Listener snapshots now read one local CLI profile inventory and compare the current App against its Core binding. The actual App determines the existing machine/App consumer lease. An unreadable inventory is unknown rather than removed; a missing/replaced App or explicit disconnect removes the old route. This discovery does not grant admission or delivery authority, and no successful authorization is cached across requests. Project Codex start/exact resume adopts current host model/effort defaults while preserving the original thread and explicit choices. Old plain-text attempts retain their original delivery format.

Validation on exact head: 212 focused Python regressions; Ruff; Mypy for the changed provider; vocabulary advisory; and actual-PR-base premerge with 5 direct + 14 selected checks pass. The source service composition passes 93 provider/runtime tests and builds the packaged Chat frontend. Regressions cover live stream handling of transient verification/inventory failures, explicit disconnect, App retarget/lease identity, independently rejected authorization, source-fence revocation, blocked old readback with independent stop/other-App delivery, and bounded workers. No new shared vocabulary owner is introduced.

Independent measurements qualify specific costs: synthetic blocked-readback feedback improves about 4.04s to 1.00s; the same real two-App discovery method improves 4.18–5.19s to 72–83ms over three repeats. Neither establishes an end-to-end latency SLO. The new source candidate keeps both App listeners and original persisted Sessions/grants. Sustained availability, complete timely control/progress presentation and live queue notice timing remain open, with slow/time-out samples retained.

Stacked on #5555 (20f803b884f34408d70027f1eecd652cc0bec249); premerge uses that actual base and its CLI output comparator. The prior origin/main comparison failure is retained, including its reproduction on the unchanged earlier head. Previous full CI had failing Python shards/pytest/merge-gate; current-head CI must be evaluated separately. Runtime changes remain maintainer-merge only. Formal installation/login recovery, direct browser access, media and permission callbacks remain unqualified.