(orchestration): Optional feasibility probe before agent dispatch

## Component

API or orchestration

## Describe the feature

Add an optional **pre-dispatch feasibility phase**: a read-only agent (or deterministic code-intelligence step) that explores the repo/ticket, estimates scope, and returns a structured feasibility summary **before** allocating an isolated compute session for the full coding workflow.

## Use case

Operators burn compute and turns on tasks that were never feasible as specified. A **deep exploration pass** (code search, dependency tracing, ticket analysis) before dispatch lets humans refine the spec or cancel before an unattended run starts.

## Proposed solution

1. Blueprint flag `feasibility_probe: boolean` (default `false`).
2. New workflow step or lightweight `coding/feasibility-v1` workflow: read-only tools, bounded turns/budget, output schema `{ feasible, summary, suggested_scope, risks[] }`.
3. On `feasible: false`, task transitions to `NEEDS_HUMAN` (see related draft) without opening a PR.
4. CLI: `bgagent submit --feasibility-probe` or separate `bgagent probe`.
5. Slack/thread hydration can feed the same probe.

## Other information

- Distinct from shipped **pre-flight checks** (GitHub reachability, token validity)—this is **task-level** feasibility.
- Complements **PR risk classification** and **toil templates**.

## Acknowledgements

- [ ] I may be able to implement this feature
- [ ] This might be a breaking change

## Acceptance criteria

- [ ] Opt-in probe runs before `RUNNING` when configured; results stored in `TaskEvents`.
- [ ] Probe respects read-only tool policy (no Write/Edit).
- [ ] Operator can promote probe output into task description via API/CLI.
- [ ] Tests cover feasible, infeasible, and skipped paths.

No implementation files are named. Start by tracing the blueprint and task lifecycle, TaskEvents persistence, the existing pre-flight checks, and the CLI submit path. Define the opt-in probe flow, read-only policy, feasible and infeasible transitions, promotion command, and tests for feasible, infeasible, and skipped cases.

Written by the indexing model from the issue text.