[lukecotter]

What

The Call Tree's three views (Time Order, Aggregated and Bottom-Up) now use lv-grid, the grid from #1166. The Call Tree's Tabulator tables are deleted. There is no setting to go back.

The Bottom-Up Tabulator table stays, because the Analysis view uses it. The Analysis, Database and inspector grids stay on Tabulator. Each gets its own change later.

User-visible changes

• The Call Tree is much faster on large logs. The numbers are below.
• Cmd/Ctrl+C copies the table in every view. The copy holds the shown columns, tab-separated, and every row that passes the filters, closed rows too. Before, only Bottom-Up copied, and only its top-level rows.
• Bottom-Up ungroup no longer overflows the stack.
• The category tint setting changes an open Call Tree at once.
• A row picked in the inspector scrolls into view only when it is off screen, as before.

Numbers

Measured on a 95 MB log of 580,091 rows: the median of 5 runs, and first render the median of 3 loads. Each cell gives the time in ms, then the longest task in ms.

Action	Tabulator	lv-grid
Time Order, sort by self time	473 · 525	166 · 0
Time Order, expand all	254 · 655	74 · 0
Time Order, find	641 · 17,639	500 · 0
Time Order, drag a column wider	60 of 60 frames dropped	0 dropped
Time Order, heap after all runs	477 MB	31 MB
Bottom-Up, group by type	2,500 · 2,636	296 · 0
Bottom-Up, sort by total time	373 · 392	18 · 0
Aggregated (965 rows), every action	17–47	17–34

All the tables, with the method, are on the results page.

Known:

• First render has one long task in every view, 179–187 ms. It is the garbage collection after the parse, and Tabulator has it too.
• In Aggregated, the synchronous tree build adds to that task: 344 ms, where Tabulator takes 334 ms.
• Earlier runs saw a 52–123 ms task in Bottom-Up group by, but only in some runs. It did not show in these 5 runs, and its cause is not known.

Screenshots

Dark	Light

The three views, Time Order, Aggregated and Bottom-Up:

Commits

Each commit is one concern, so a review by commit is easiest.

1. Time Order and Bottom-Up on lv-grid, then a steady scroll bar (estimated row heights from line counts).
2. Aggregated on lv-grid, and its bench with a Tabulator baseline recorded before the delete.
3. Copy in every view.
4. The colorize setting applies to an open grid.
5. Reveal keeps the scroll when the row is in view (goTo takes scrollIfVisible).
6. Remove the Call Tree's Tabulator tables and the helpers only they used.
7. Docs, the screenshot script, and new screenshots in lana/assets/1_24.
8. CHANGELOG.

Checks

• pnpm lint passes, and jest passes, 2904 of 2904.
• Anchoring, in Playwright, in all three views: after a toggle, sort, sort again, clear sort, filter off or filter on, the row in view stays at the same pixel in every frame. Four clicks on a twisty with no mouse move toggle it each time.
• Checked by hand in the dev host: find, the inspector both ways, go to row, the Name link, copy, the column menu, the colorize setting, and a light and a dark theme.

Stacked on #1166.

Known open items

• filterOn long task on the 580k-row bench log. In 1 of 5 runs, Time Order filterOn shows one 55–85 ms task. A Chrome trace shows it is a V8 full garbage collection (MajorGC), with no grid code in the task. The collection falls in whatever step runs when the heap fills.
• Follow-ups from review, not in this PR:
  • Number the find marks from the store's count, not from the DOM text, so a cell no longer has to mark its find text.
  • Give a row filter a cache key in the grid store, so a host does not have to keep its filter objects.
  • Use one cancel mechanism in the Call Tree for the view switch, the visibility wait and the builds.