fix: keep stdout parseable, correct advertised examples and the auto-JSON claim

[GregHolmes]

Fixes the three output-correctness issues filed today. Stacked on #103 — base is chore/release-hardening, not main, because the #104 fix builds on the plugin_manager console change there. Rebases onto main cleanly once #103 merges.

Closes #104, closes #105, closes #106.

1. Diagnostics off stdout on the failure path (#104)

auth, client, timing and base_group_command declared bare Console() instances bound to stdout and printed diagnostics through them, so dg -o json <cmd> on an auth failure wrote 132 bytes of prose to stdout and left stderr empty. get_status_console() warns about exactly this in its own docstring.

Those four now use the shared stderr_console. base_command is deliberately not moved — its console writes the payload (tables, JSON, raw text from _output_*), which belongs on stdout. It moves from a bare Console() to the shared stdout instance so it stops silently missing the agentic no-color/highlight config.

Moving the prose off stdout left stdout empty, which is still unparseable, so the guard's except now emits an ErrorResult through the normal output path. It is a no-op in default mode, so humans see no duplicate of the stderr diagnosis.

dg -o json … (invalid key)	before	after
stdout	132 B prose	78 B, {"status": "error", …}
stderr	0 B	132 B prose
json.load(stdout)	JSONDecodeError	parses

2. Advertised examples that don't parse (#105)

dg usage advertised --days 30 and --start/--end; the real options are --start-date/--end-date and there is no --days. The sweep the issue suggested found the same error twice more:

• dg debug stream — debug has audio, browser, network, probe, toolkit. The stream debugging its agent_help describes is probe.
• dg profiles --show default — profiles has --switch, --current, --list.

Worse than a help-text typo, because the same array is what --agent-friendly emits: an agent asking the CLI how to use itself was handed four commands that fail.

tests/unit/test_command_examples.py now parses 125 examples, one test each, via parse_args — no handler invocation, no network. Examples are shell snippets rather than argv, so it extracts only the dg … segments; pipelines, upstream producers and trailing # comments are not ours to validate, and command substitution is skipped. dg debug toolkit's script subcommands are exempt with a stated reason — they come from a manifest fetched by toolkit refresh, so a clean checkout has only refresh.

3. Auto-JSON attributed to piping (#106)

The landing page promised output auto-switches to JSON when stdout is piped. setup_output only flips the format when is_agentic() is true, which needs 3+ soft signals; a plain pipe scores 1–2 because TERM is set in any normal terminal. Verified directly — with both streams non-tty and TERM=xterm, is_agentic() is False and setup_output("default") leaves the format at default.

Two instances beyond the three the issue names:

• README.md:319 said the switch happens in "a non-TTY environment (pipes, CI, or AI coding tools)". CI and AI tools are right; pipes are not.
• index.astro:433 is a JSON-LD FAQPage answer — structured data Google can surface as a rich result, so the false claim travels further than the page.

The agent-mode copy at :645 and llms.txt:34 already attributed the switch correctly and are untouched.

One deliberate deviation from the issue

#106 offered keeping "Clean stdout channel. No surprises in pipes." if #104 landed first. #104 landed, but the claim is still not true, so the bullet is softened instead. Probing failure paths across ten commands found two more polluters:

dg -o json ffprobe --path /nonexistent        -> 105 B, prose then JSON
dg -o json debug audio -f /nonexistent.wav    -> 988 B, prose then JSON

Both come from bare Console() instances in their own command packages. Unlike the deepctl_core ones, those consoles are mixed — they also carry the human-readable display that legitimately belongs on stdout in default mode — so routing them needs per-call-site judgment across ~30 command packages rather than a module-level swap. Out of scope here; worth a follow-up issue.

Tests

• test_output_channels.py — which shared console each module holds; the JSON failure payload parses while default mode stays silent on stdout; plus an AST sweep asserting no deepctl_core module declares a bare Console(). Confirmed to fail when one is reintroduced.
• test_command_examples.py — 125 parametrised example checks, with a test_examples_were_discovered guard so the suite cannot silently degrade to testing nothing if the entry point groups go stale.

1232 passed, 6 skipped (from 1107). ruff check + format, mypy: clean. npm run build in web/ succeeds; the corrected strings render and the FAQPage schema still parses.

Note for review

base_command.py is CRLF in the repo (7 of 235 Python files are, with no .gitattributes). An earlier pass normalised it to LF, which turned a 23-line change into a 1027-line diff; the history was rewritten to preserve CRLF so the diff stays readable. The stray CRLF files are probably worth a separate .gitattributes cleanup rather than drive-by normalisation.