feat(agent-org): add authoritative team lifecycle

[ShiboSheng]

Problem

Fixes #756.

The existing Agent Org lifecycle does not provide an authoritative, restart-safe boundary for a long-lived Team. Initial member creation can be resumed only through loosely coupled session state, finality is inferred by multiple paths, Run View and background recovery do not have a sufficiently strict read/write boundary, and the frontend retains fallback polling for states that should be quiet.

That model cannot safely support a Team that finishes one batch of work, becomes reusable and idle, and later accepts more work. It also makes identity duplication, stale quiescence decisions, accidental read-side writes, and retained background activity difficult to reject by construction.

Solution

This PR implements the first lifecycle slice from the long-lived Team design:

• Persist Starting together with canonical member materialization receipts and the initial-input intent. Restart recovery continues the same (member_id, agent_id) identities instead of regenerating a Team.
• Centralize formal-work quiescence facts and decisions across Turn, Task, Inbox, materialization, final-answer, generation, and work_revision state.
• Allow the finalizer and the Working Watchdog to perform Working -> Idle only through a generation-and-revision conditional transition. Stale observations become no-ops.
• Keep Run View bounded and pure-read. It projects the shared quiescence facts but cannot transition lifecycle state, claim Inbox rows, Wake a Member, perform recovery, or update timestamps.
• Replace the unbounded/mixed watchdog work with one 60-second, running-only scan, LIMIT 100, and one 250 ms cooperative deadline shared by the entire tick.
• Replace frontend polling of every non-terminal state with an explicit starting/running allowlist. Hidden windows and the last pollable subscriber destroy the shared timer; visibility recovery performs one immediate bounded revalidation.
• Remove startup's blanket Running-to-Paused mutation. Starting recovery and plan-artifact repair remain separate one-shot owners rather than periodic Watchdog duties.
• Put the entire redesign behind one fail-closed frontend/backend rollout gate, ORGII_AGENT_ORG_REDESIGN, which remains disabled by default.

The current compatibility wire value for Working remains running.

Potential risks

• The lifecycle changes are intentionally cross-layer: launch, persistence, Turn finalization, Task/Inbox work revisions, Tauri projection, startup recovery, Watchdog inspection, and frontend subscription ownership must stay aligned. Unit, integration, hook, and packaged-runtime checks cover these boundaries, but future changes must continue using the centralized lifecycle APIs.
• The redesign is an incomplete internal stack after PR1. If a release pipeline accidentally sets ORGII_AGENT_ORG_REDESIGN=1, users can reach Idle but cannot yet submit a new Root/Group/direct Turn or mutate the Task graph from Idle. The production default must remain disabled until the later stack is complete.
• The Watchdog intentionally provides no fairness cursor. When more than 100 Working Teams exist, remaining Teams wait for a later tick; correctness continues to belong to event/finalizer owners rather than the Watchdog.
• Schema changes are additive and initialization is idempotent. Rollback is to disable the rollout gate and revert the code; the additive receipt/progress rows can remain without affecting the gated-off product path. No destructive migration or automatic historical cleanup is included.
• Running -> Idle uses a serialized generation/work_revision check. A concurrent work mutation wins by changing the revision, so the idle attempt becomes a no-op; later PRs must not introduce a second lifecycle owner or bypass that revision boundary.

Current implementation state

• Starting owns first materialization and restart recovery. Successful construction ends in running when initial formal work exists or idle when it does not; unrecoverable construction ends in failed with structured diagnostics.
• Idle means the current formal work is quiescent and the Team is retained. It is not a terminal Completed state.
• Quiescence fails closed for unknown or inconsistent formal-work facts and validates the exact presented/observed work revision before idling.
• Run View reads the same canonical facts in one bounded read transaction and has no mutation or Wake authority.
• The global Watchdog scans only Working (status='running') Teams. Starting recovery is a one-shot launch/startup responsibility.
• Frontend fallback polling has one shared owner and exists only while a visible subscriber still represents a starting or running Team.
• The rollout gate is absent from Team settings, prompts, model tools, and persisted Team snapshots. Missing, malformed, 0, or true values remain disabled; only 1 enables the internal slice.

Scope boundaries and next-PR handoff

PR1 does not implement Pause/Resume redesign, Archive/Delete, the final Task FSM, additional Writers, UserDirectedWork, direct Member work, Group multi-mention, peer side quests, or the final Group read surface.

The enabled PR1-only intermediate state deliberately rejects new work from Idle:

• Root/direct Turn submission returns team_idle.
• Group Chat persistence rejects idle before writing an Inbox row.
• Task graph mutation accepts only running and returns agent_org_run_not_mutable for Idle.
• No Provider, Member Wake, or implicit Idle -> Working transition occurs.

PR7 is expected to add the atomic first Task graph plus Idle -> Working activation. PR8 opens direct Member UserDirectedWork, PR9 opens Group targeting and peer side quests, and PR10 completes the bounded Group read surface and default rollout. Earlier follow-up PRs must extend the frozen launch roster, generation, revision, and lifecycle transition APIs established here rather than creating another state owner.

Verification

Automated checks:

• cargo test -p agent_core — passed on the frozen PR1 implementation before packaged acceptance.
• pnpm typecheck — passed before packaged acceptance.
• cargo clippy --all-targets -- -D warnings — passed again after fast-forwarding to the latest develop.
• Commit hooks passed: lint-staged, scoped TypeScript checking, scoped Clippy for agent_core, org2, and session_persistence, ESLint 0, circular dependency count 0.
• git diff --check upstream/develop..HEAD — passed.
• The final PR diff contains no documentation files, local paths, credentials, temporary harnesses, caches, or build artifacts.

Packaged Tauri acceptance used a temporary packaged .app launched through macOS open, an isolated ORGII_HOME/SQLite database, isolated ports, a mock provider, and the redesign gate enabled in both frontend and backend:

• Visible Running Team for five minutes: 6 successful Run View requests (one immediate plus five 60-second intervals), with one shared poll owner.
• Hidden for 5 minutes 15 seconds: 0 Run View requests.
• Visibility return: exactly 1 immediate revalidation, followed by the normal 60-second cadence.
• Closing the last pollable Team subscriber and waiting more than 70 seconds: 0 residual Run View requests and no Timer/Provider/Wake activity.
• RSS changed by 16 KB during the foreground window and by 112 KB during the hidden sample; hidden CPU samples were 0.0%.
• Product refresh used Command+R. The before/after Run, progress, Task, Inbox, Turn-intent/Wake, recovery, session, and event snapshot was byte-identical; updated_at and work_revision did not change.
• Watchdog observation covered 20 real ticks at 59,982-60,004 ms intervals. The query visited only Running rows with LIMIT 100; the shared scan used 9-25 ms of its 250 ms budget.
• The exact temporary App PID was terminated after the test; both isolated listeners were gone.

Performance verdict: pass.

Visual screenshots are not included because this PR's meaningful acceptance evidence is lifecycle persistence, database immutability, API cadence, hidden-window behavior, and background-resource ownership rather than a visual layout change.