fix(harbor): recognise multi-step tasks in the converter

[Ulusha]

_is_harbor_task requires a root instruction.md, so valid multi-step tasks aren't recognised — their instructions live under steps/<name>/, declared by a [[steps]] array in task.toml. load() then raises no Harbor tasks found or drops every task at parse.

Reproducible on Harbor's own examples/tasks/hello-multi-step-simple:

_is_harbor_task(): False

though Harbor's own tooling builds and runs it fine.

Fix: recognise a task by a parseable task.toml plus an instruction in either form — root instruction.md (single-step) or a [[steps]] array (multi-step). _is_harbor_task and _parse_task share the check.

Behaviour note: _parse_task now rejects an unparseable task.toml (previously tolerated with an empty config), since detection must read it.

Tested: 3 single-step + 2 multi-step tasks accepted; non-Harbor folder, task.toml without instruction/steps, and broken task.toml rejected.

---

Note

Low Risk
Scoped to Harbor task discovery/parsing in the integration layer, with new tests; behavior change is limited to which task directory layouts are accepted.

Overview
Harbor load/detect no longer treats a root instruction.md as mandatory. Tasks are valid when task.toml parses and either that file exists (single-step) or the config defines a [[steps]] array (multi-step, instructions under steps/<name>/).

_read_task_config, _task_has_instruction, and updated _is_harbor_task / _parse_task share that rule. task.toml alone (no root instruction and no steps) is excluded. Broken task.toml still loads only if a root instruction.md is present (empty config); multi-step dirs without parseable config are dropped because steps cannot be detected otherwise.

Tests add make_multi_step_task and cover detect/load for multi-step tasks, mixed datasets, and bare task.toml rejection.

^{Reviewed by Cursor Bugbot for commit f6ef295. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 0198416. Configure here.}