fix(harbor): restore task.toml [metadata] as Task columns on load

[rohitsudhakar1]

load() documents its rows as carrying task.toml's [metadata] as columns, and _HarborTask still parses and stores the config for it, but the mapping was dropped in 5b7110a ("Simplify the v6 contract surfaces") and nothing has read config since. Every row loaded from a Harbor dataset comes back with columns=None.

Repro

Any Harbor dataset whose tasks carry a [metadata] table:

from integrations.harbor import load

# task.toml:  [metadata] \n category = "networking" \n difficulty = "easy"
[row.columns for row in load("path/to/dataset")]
# -> [None, None]

Impact

The failure is quiet. A converted terminal-bench-style dataset loses the category / difficulty / tags each task ships with, so results can't be filtered or grouped by them. Because the field is merely absent rather than wrong, nothing surfaces the loss — the docstring on load() still promises the behaviour, and _HarborTask.config is still populated for a consumer that no longer exists.

Fix

Restore the mapping behind a small _columns helper. Metadata is left unset rather than emptied when a task.toml has no [metadata] table, so tasks without it don't gain an empty column set.

Tests

Three tests, using the existing conftest fixtures (which already carry [metadata]):

• test_load_surfaces_task_toml_metadata_as_columns — metadata reaches Task.columns
• test_load_leaves_columns_unset_without_metadata — no [metadata] table stays None
• test_load_dataset_carries_per_task_metadata — metadata stays per-task across a dataset whose rows are grouped by build context

The first and third fail on current main:

FAILED integrations/tests/test_harbor.py::test_load_surfaces_task_toml_metadata_as_columns
FAILED integrations/tests/test_harbor.py::test_load_dataset_carries_per_task_metadata
2 failed, 1 passed

and pass with the fix:

uv run pytest integrations/tests/test_harbor.py -q   ->  19 passed
uv run ruff format . --check                         ->  258 files already formatted
uv run ruff check .                                  ->  All checks passed!
uv run pyright                                       ->  0 errors, 21 warnings  (parity with baseline)

The full suite is 725 passed with 6 pre-existing failures in test_workspace.py / test_capability_backing.py around setpriv and shell-uid sandboxing. I verified those fail identically on an unmodified checkout on this machine, so they're unrelated to this change.

---

Note

Low Risk
Localized Harbor loader fix with no auth, runtime, or export path changes; behavior matches existing Task.columns usage on sync/upload.

Overview
Restores Harbor load() mapping so each row’s Task.columns comes from task.toml’s [metadata] table (e.g. category, difficulty, tags), matching the documented contract that was dropped in a prior simplification.

A new _columns helper copies [metadata] when present and returns None when it’s missing or empty, so tasks without metadata don’t get an empty column dict. Metadata stays per task when rows are grouped by shared environment/ build context.

Adds three tests covering metadata surfacing, unset columns without [metadata], and per-task metadata across grouped datasets.

^{Reviewed by Cursor Bugbot for commit 3b97097. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jdchawla29]

addressed in #516