refactor(output): one column renderer, one kind word, one status text

[ryanlewis]

Closes #243

What changed

printTasks, printProjects, printAreas and printTags each declared a local row struct, measured every column with lipgloss.Width and padded with padCol before joining with a gap. Four copies of the same twenty lines, and the copy inside printTasks also carried the terminal-width column dropping. There is now one table in internal/output/table.go.

Rows of cells go in, measured and padded lines come out. The column dropping is a dropOrder field naming the columns that may be given up, in the order they go, rather than a pair of booleans and a hand-written width sum.

The rule the four copies shared is written down once: every column is padded to its widest cell except the last one declared, which goes out as it is. Dropping a column does not move that rule, which is why a narrow listing still pads its title column. That is the behaviour the copies had; it is now stated rather than implied.

printTasks keeps its group-header logic, which no table can own, because it has to write a header between two rows. table.lines() hands back one line per row for exactly that, so the header fold from #205 is untouched.

One place each for the kind word and the status text

t.Type == model.TypeProject was branched on six times across output.go and agent.go. It is now isProject(t), and the word that reaches the output is kindWord(t), which takes "task" and "project" from model's type codec, the same source the JSON type field renders from.

kindWord is deliberately not TaskType.String(). A heading, or a code Things has yet to write, reads as a task on these surfaces, which is what the CLI has always shown. The doc comment says so.

statusText no longer keeps its own switch over the three statuses. It capitalises model.Status.String(), so the detail block, the JSON and the agent brief cannot drift apart, and the unrecognised code still renders "Unknown" because the codec already falls back to "unknown". checklistLine takes its "(cancelled)" from the same source.

statusIcon stays as it is. [ ], [~] and [x] are glyphs rather than the status words, and Markdown's two-state checkbox in a brief is a third thing again. Folding those together would have meant inventing a rendering, which this change is not for.

Proof the output did not change

The golden test went in first, in its own commit, before anything was refactored.

TestRenderGolden renders a fixed fixture through every public entry point: Print for tasks, projects, areas and tags and for a *model.Task, PrintTaskList with and without a view label, PrintTaskWithChecklist for a to-do, a project and a repeating item, PrintHint, and PrintAgentBrief for a task, an open project, an empty project, a closed project, a repeating item and a note carrying its own code fence. Empty listings are cases too. Each case renders at both ends of the colour profile, so the ANSI is pinned as well as the text, and the task listings render at three terminal widths so both drop steps are covered.

The refactor commit does not touch testdata/render_golden.txt. That is the claim, and the file's absence from that commit's diff is the evidence.

The golden values were captured from unmodified rendering code. The one production change in that first commit is the seam the test needs: termWidth becomes a var, like nowFn beside it, so a test can pin a width instead of depending on how the test binary's stdout happens to be attached. The body is unchanged. origin/main moved to 61542c3 while this was in review, but #260 does not touch internal/output, so the captured bytes are still the bytes of the base this merges onto.

Against the live database, 226 invocations of the built binary are byte-identical to origin/main, 1.8 MB of output in all: nine views plain, with --include-completed, as JSON, with --color always and with --no-hints; projects, areas and tags in the same three forms; and 43 real items through show, show --agent, show --json and show --color always. Exit codes are compared too. The database was read only through the binary.

Every line in the golden file ends in |, because a padded column emits trailing spaces an editor or a hook would otherwise eat, and ANSI escapes are written \e so the file stays readable and greppable.

Review

/code-review --fix at high effort found one defect, and it was mine, in the golden test rather than in the refactor.

TestRenderGolden returned early when the document matched and otherwise reported from inside its per-case loop. A difference the case splitter does not see, such as trailing whitespace or text before the first header, both of which it drops, left the loop with nothing to report and the test passed with got != want. Appending two newlines to the golden file reproduced it. For a test whose whole job is pinning bytes that is the one outcome it must not have, so it now fails loudly and names the regeneration command.

The review separately re-derived the extraction rather than trusting the tests: that table.fits reproduces the old arithmetic including the two-step drop and the absent re-check after the last drop, that the last declared column is the one left unpadded whether or not it survived, that kindWord and statusText are equivalent to the switches they replace down to the unknown-code arm, and that the golden file is the same blob in both commits. It also confirmed the golden output is timezone-stable, checked under TZ=Pacific/Kiritimati and TZ=UTC, and that making termWidth a var introduces no race.

Noted, not fixed

TestListQueryGoldenSQL in internal/db has the same gap the review found here: it reports only from inside its per-case loop, so a difference outside a ### view= block would pass. It is a separate package and a separate change, so it is left alone.

Not touched

No command surface moves, so internal/skill/SKILL.md and the pages under docs/content/ have nothing to say about this. The (project) marker, the group header fold, the project icons and the agent brief's wording are all documented contracts and all unchanged.