[P1][AGENT-01] Task graph and scoped orchestration

Programme: #1
Epic: #20

## Objective
Implement the core harness as a durable supervisor-owned task graph, not an open-ended LLM loop.

## Scope
- Task lifecycle: Draft → Scoped → Planned → Running → AwaitingApproval → Verifying → Completed plus Suspended/Cancelled/Failed/Blocked/NeedsReconciliation.
- DAG/state-machine representation with dependencies, inputs, outputs and explicit success predicates.
- Specialist subjobs with scoped context bundles, allowed capabilities, deadlines and budgets.
- Execution leases/fencing tokens, checkpoints and resumability.
- Concurrency control: parallel safe reads; serialize conflicts over mutable accounts/carts/repos/budgets/desktops.
- Plan revision proposals that cannot expand authority.
- Cancellation cascading and orphan-job cleanup.

## Harness rules
- The supervisor owns durable state; provider thread/session IDs are secondary references.
- Subagents cannot recursively spawn unbounded trees or inherit ambient authority.
- Completion requires verification, not a model saying “done”.

## Acceptance criteria
- [ ] Subagents cannot expand their grant/tool/resource scope.
- [ ] Conflicting access to one mutable external resource is serialized/fenced.
- [ ] Cancellation revokes leases before later worker output can dispatch new effects.
- [ ] Task resumes from persisted state after worker/app restart.
- [ ] Budget/deadline exhaustion produces explicit bounded failure/degraded state.
- [ ] Success predicates and verification outcomes are durable and inspectable.

## Dependencies
- CORE-02
- CORE-03
- SEC-02

**First phase:** P1
**Maturity target:** P3
**Owner:** harness-pc-providers

## Task issues
- [ ] #439 `AGENT-01.T01` — Define graph nodes, effects and state transitions
- [ ] #440 `AGENT-01.T02` — Implement durable task submission and planning
- [ ] #441 `AGENT-01.T03` — Build the ready-node scheduler
- [ ] #442 `AGENT-01.T04` — Implement scoped specialist delegation
- [ ] #443 `AGENT-01.T05` — Implement conflicting-resource locks and fencing
- [ ] #444 `AGENT-01.T06` — Implement bounded tool execution and context management
- [ ] #445 `AGENT-01.T07` — Add independent verification nodes
- [ ] #446 `AGENT-01.T08` — Implement checkpoints, stop and recovery
- [ ] #447 `AGENT-01.T09` — Expose progress, budgets and revisions
- [ ] #448 `AGENT-01.T10` — Run graph race, budget and recovery tests

The issue names no files, tests, or entry points; begin by reading dependencies CORE-02, CORE-03, and SEC-02, then use task issues #439–#448 as the implementation map. Done means meeting the listed acceptance criteria for scoped authority, fencing, cancellation, recovery, bounded failure, and durable verification.

Written by the indexing model from the issue text.