Own human intent through program delivery and add live Proof operations

[kai-linux]

Outcome

Adds durable original intent, project/program ownership and dependent work packages, evidence-gated outcomes, shared attempt/capacity/reservation budgets, human handoffs, revision/pause/cancel controls and transactional status delivery. Default planning manages accepted commitments instead of speculative growth. Adds a private read-only Proof dashboard beyond GitHub Projects.

Root-cause fixes

• Missing result/no-diff cannot override independently verified delivery ([task] In the Readme, change the title from Agent OS to Agent-OS #355 regression).
• Decomposition cannot close parents or silently discard work packages.
• Legacy retries and board cleanup cannot override managed ownership.
• Deployment checks survive merge-triggered issue closure.
• PR handoffs and notifications retry without restarting workers.
• External effects require configured adapters, delegated targets and durable receipts.

Verification

794 Agent-OS tests pass in an isolated environment with pinned Proof 0.2.0. 43 Proof tests pass; kai-linux/proof PRs #1 and #2 are merged. Changed-path F checks, shell syntax, diff checks and staged secret scan pass. HTTP and Playwright checks cover desktop/mobile, live data, filters, details, escaping and outage/stale snapshots. Browser fixtures were not inserted into live records.

Boundaries

See docs/delivery.md. Historical claims and unknown costs are not invented as success/zero cost. CLI workers remain trusted host processes, not an OS-enforced sandbox. Real media publication, provider billing receipts and business-impact evaluations remain unproven. No new cron or high-access execution jobs are installed by this patch.

[kai-linux]

Self-review completed. Checked the queue/dispatcher compatibility boundary, transaction and outbox behavior, parent/child/dependency gates, shared limits, revision and process-group fencing, action receipt replay protection, private dashboard/auth/export boundaries and actual versus claimed completion. Review found and fixed mixed hierarchy cycles, parent-resume handling, silently dropped work packages, merge closure before deployment, historical KPI contamination and stale result files.

794 tests pass locally using installed Proof 0.2.0 (not an editable checkout); 43 Proof tests pass. Secret scan, shell syntax, F checks and whitespace checks pass. Remaining scope boundaries are explicit in docs/delivery.md; this does not claim an OS sandbox, actual billing coverage, or an executed media publication pilot.