docs: rewrite the README around one worked example

[abrichr]

The README opened with six bolded-lead feature bullets and then spent a third of its length on things this package does not do. I replaced it with one worked example that produces an HTML file, and everything in the new file was run against an editable install of this branch.

What was structurally wrong

The old opening was a category description followed by a bullet list ("Component-based", "Composable", "Standalone HTML"). A reader could not picture anything until line 40, and the first runnable thing was a snippet that built an HTML fragment nobody could see. Now the second command in the file writes a viewer and the next line is a picture of it.

The rest of the length came from nobody deciding what mattered. Two tables of 22 components, a source tree, four screenshot captions, and a Documentation list of eight files. The signatures moved to docs/COMPONENTS.md, read out of the installed package with inspect.signature rather than transcribed. The source tree went; git ls-files answers that better.

What was inaccurate

None of this survives into the new file:

• The Audio Transcript Feature section, about 40 lines describing timestamped transcription synced to playback. The string transcript does not appear anywhere in src/. It is a feature of openadapt-capture, whose create_html reads transcript.json out of a capture directory.
• The four screenshots and their captions. scripts/generate_readme_screenshots.py calls openadapt_capture.visualize.html.create_html, so those images are openadapt-capture's viewer, not this one, and the captions point at an "audio transcript (right sidebar bottom)" that is not visible in any of them. The PNGs stay in docs/images/; they are just no longer presented as this package's output.
• The Synthetic Demo Viewer section. It tells you to open synthetic_demo_viewer.html, which does not exist, and links SYNTHETIC_DEMOS_EXPLAINED.md, which also does not exist. It carried an unsourced "33% -> 100% accuracy" claim I could not check.
• SEARCH_FUNCTIONALITY.md was linked from the repository root and lives at docs/SEARCH_FUNCTIONALITY.md.
• "Standalone HTML: Generated files work offline, no server required." The page fetches Alpine from cdn.jsdelivr.net. I aborted that request in a headless browser: the summary cards and the filter dropdowns paint, the task list renders empty, and nothing responds to a click.
• "5 production viewers available". There are generators for benchmark, capture and segmentation. The other two entries were the examples/ modules for openadapt-ml and openadapt-retrieval.

What I verified

I first ran everything against openadapt-viewer==0.2.0 from PyPI in an empty venv, which was a mistake worth recording: the wheel still carries the hardcoded DEFAULT_CAPTURE_PATH and the capture.db reader, both of which #15, #17 and #19 have already fixed on main. Writing that up as current behaviour also tripped tests/test_no_hardcoded_paths.py, which bans an absolute home path in any tracked file including docs. The second commit re-runs everything against an editable install of this branch.

Every command and snippet in the README was executed there and the output blocks are that run's stdout: openadapt-viewer demo, the PageBuilder example, generate_capture_html, benchmark --data against a directory holding episodes.json plus recording.db, the $OPENADAPT_CAPTURE_RECORDING form, and catalog stats. Signatures come from inspect.signature against the install rather than from the source tree. Every external URL returns 200 and every relative link resolves to a file that exists. pytest tests/ is 154 passed, 79 skipped locally, and ruff check . is clean.

The screenshot is a real openadapt-viewer demo run. scripts/generate_demo_screenshot.py regenerates it, and because the demo data is an unseeded random.random() a rerun will not match task for task. The README says so.

Four things the "What it doesn't do" section documents

Written down rather than fixed, because this is a docs change.

1. The page is not offline-safe, as above.
2. generate_capture_html writes <link href="src/openadapt_viewer/styles/episode_timeline.css"> and a matching <script src=...episode_timeline.js> into the page. Both files ship inside the installed package. Resolved relative to the output HTML they are 404s, so the episode timeline is unstyled and inert unless the output lands at a source checkout root.
3. benchmark --data on a directory holding a pre-2026-07-17 capture.db prints Generated: and writes a viewer with zero tasks. LegacyCaptureError subclasses FileNotFoundError, so the except (FileNotFoundError, ValueError, KeyError) in generate_benchmark_html swallows it and falls through to load_benchmark_data. The migration command that recording_db took the trouble to name never reaches the user. Reproduced against a real legacy directory.
4. openadapt_viewer.__version__ is "0.1.0" while pyproject.toml says 0.2.0.

One more, visible in the committed screenshot but not called out in the README: the task detail panel renders below the task list rather than in the second column of the 1fr 2fr grid, which stays empty.

Could not verify

The --open flag on demo, benchmark, capture and segmentation calls webbrowser.open; I did not exercise it headless. uv sync --all-extras in the Development section comes from .github/workflows/test.yml rather than a local run, so CI on this PR is the check on it.