Sharpen method summaries, CLI errors, and review UI; add dark mode

[Raman369AI]

Summary

A review pass from a first-time user's view (someone who knows some Python), run against the bundled example, Threadline itself, and agent-kanban-pm.

Method summary accuracy

• False "Changes": a callee calling param.method() counted as possibly mutating the caller's argument, even for email: str. Parameters typed as immutable (str, int | None, Optional[str], tuple[int, ...]) are now skipped. Changes that remain only possible show as dashed chips with the source in the tooltip.
• @dataclass constructions were "probably": any decorator downgraded the call. Imported @dataclass / @total_ordering classes now resolve as direct calls. @staticmethod stays "probably", as the curated corpus requires. Two corpus cases were added.
• Side effects were missing: Calls now lists methods called on objects the method receives (db.execute, db.commit, repository.save). Library calls show solid, with the library method in the tooltip; unresolved ones are dashed. Clicking one highlights its line.
• Raises: the summary shows e.g. Raises: HTTPException when a method raises.
• Search ranking: exact method names first, then name prefixes; project code before tests and lambdas. The live server and saved-HTML search are kept in sync.

threadline review CLI

• Errors print as plain text rather than JSON; the agent query commands still output JSON.
• A bad --base names the revision; a missing --output folder is caught before analysis.
• Warns when no Python files are found; request logging is off.
• Falls back to a free port when 4173 is busy, unless --port is given.
• Every option has --help text.

UI

• Commands: test modules with a main guard move to a collapsed "Test modules you can run" group (19 → 2 real commands on this repo).
• Clearer labels: the call-map context names the selected step, module cards show the line instead of repeating the file, <module> is gone, and Coverage is explained in plain language.
• Dark mode: the ~190 hard-coded colors became 45 light-dark() tokens. The page follows the system setting, and a Theme button cycles System / Light / Dark (remembered, and applied before first paint). Below the header, light mode is pixel-identical to before.
• Header: a ☰ button replaces "Hide sidebar"; the Coverage button is removed.
• Instructions: a corner button replaces "Source only · not executed". Its panel has How to read and Coverage sections, and the analysis-issues badge opens Coverage directly.
• Resizable code pane: drag the divider, or use the arrow keys; double-click or Enter resets it; the width is remembered.
• The reading guide now fits narrow windows; before, it overflowed.

Testing

• Unit tests: 162 pass, including new tests for immutable parameters, search rank, plain-text review errors, the empty-repo warning and the busy-port fallback.
• test_semantic_accuracy.py: 25 cases pass, 2 of them new.
• tests/browser_smoke.py in all four modes (live, --changes, --html, --html --changes): all checks pass, including new ones for summary calls and raises, the theme toggle and persistence, divider drag and reset, and Coverage inside Instructions.
• tests/browser_accessibility.py --browser chrome: 84 checks pass, and axe now runs in both themes at each of 24 states with zero violations.
• tests/release_smoke.py on a freshly built wheel passes (theme.js is packaged and served). mypy (CI scope) and node --check pass.
• Not run locally: native Safari (the macOS CI job) and Windows.

Notes for review

• Two existing test expectations changed on purpose: test_usability.py now expects test modules in their own testCommands catalog category, and one browser check looks up each module card's scope id because cards no longer repeat the file name.
• New asset threadline/static/theme.js, added to the server routes, the HTML export, CI node --check, and the release smoke test.

🤖 Generated with Claude Code