test(core): stop output-format global leaking across tests

[GregHolmes]

Problem

BaseCommand.output_result reads the output format from the process-global deepctl_core.output._output_config (via get_output_format()), not from config.get(). But five TestBaseCommand.test_output_result_* JSON tests only stubbed config.get.return_value = "json" — a value the code never reads.

As a result those tests depended on an earlier CLI-invoking test having left the global on "json" (the CLI entrypoint src/deepctl/main.py calls setup_output(), which mutates the global and nothing resets it). They pass in the full suite by collection-order luck and fail in isolation:

# on main
$ pytest packages/deepctl-core/tests/unit/test_base.py
5 failed, 83 passed

In isolation the format is its default "default", so output_result returns early and _output_json / print_json is never called → the assertions fail.

Fix (both halves)

1. Tests set what the code reads. The 5 JSON tests now @patch("deepctl_core.output._output_config", {...}) — the same pattern the yaml/table/csv tests already use — instead of the no-op config.get stub. All 11 format patch sites now spread a shared _OUTPUT_CONFIG_DEFAULTS constant ({**_OUTPUT_CONFIG_DEFAULTS, "format": "json"}) that mirrors the real global's full key set, so the stand-in can't drift from _output_config (previously the dicts omitted agentic).
2. Stop the leak. New autouse fixture in packages/deepctl-core/tests/conftest.py restores the pristine _output_config around every core test, so a format set by one test can no longer bleed into the next.

Half 1 makes the tests correct on their own; half 2 keeps the latent ordering bug from silently returning.

Verification

# in isolation (was 5 failed)
$ pytest packages/deepctl-core/tests/unit/test_base.py   → 88 passed
# whole core unit dir (was 5 failed)
$ pytest packages/deepctl-core/tests/unit                → 354 passed
# full suite
$ pytest                                                 → 976 passed

ruff / format clean (only test files touched). No src/ changes.

Note on the diff

Includes ~90 lines of ruff format reflow in TestConfirmPromptGating / TestIsGuided (the with patch(A), patch(B): blocks became parenthesized with ( … ): form, plus one line-wrap). All behavior-neutral formatter output, no logic change — called out here so it doesn't read as unexplained scope.

Scope

Pre-existing core test hygiene — independent of #92. Reset fixture is intentionally core-scoped to limit blast radius; a root-level version (covering cross-package leakage) can follow if wanted.

🤖 Generated with Claude Code

[dg-coreylweathers]

Approved — verified end-to-end, both rounds.

Re-ran the whole story locally rather than trusting the PR body:

• Reproduced the bug on main: test_base.py in isolation → 5 failed / 83 passed, exactly the five output_result JSON tests, failing on the "default" early-return. Confirmed in source that output_result reads get_output_format() → _output_config["format"], so the old config.get stubs were no-ops as described.
• On this branch: isolation 88 passed, core unit dir 354 passed, ruff format --check clean, no new ruff check findings vs main.
• The autouse reset fixture is correct: pristine snapshot taken at collection time before any test runs, in-place clear()/update() composes cleanly with the tests' @patch swaps, no overlap with the root conftest.
• f1469c1's _OUTPUT_CONFIG_DEFAULTS spread resolves the missing-agentic-key shape drift across all 12 format tests, and the PR body now accounts for the ruff-format reflow.

Two non-gating notes for the record:

1. Full suite here is 975 passed + 1 pre-existing unrelated failure: test_file_info_modified_timestamp (deepctl-shared-utils) uses local-time fromtimestamp() and fails on any machine west of UTC — that's the 975-vs-976 gap between our machines. Deserves its own issue (fromtimestamp(ts, tz=timezone.utc)).
2. Core-scoped reset fixture accepted as-is; root-level version can follow if cross-package leakage ever shows up.