[lauridskern]

The comparison and TraceSift cards now use the recording card's layout: the charts or the open finding on the left, and a sidebar with an overview, a list to move through, and the actions.

The comparison card drops its range table and legend. Each chart's heading shows a figure per run (peak CPU, memory change, average FPS). A summary compares every measure across runs, for the whole runs or the selection, and says where a shorter run covers only part of a selection. The sidebar lists each run's sharp changes in words, using fixed per-second thresholds: CPU 10 points, memory 50 MiB, FPS 15. Choosing a change selects it in every run. Memory can switch between absolute values and change from each run's first reading. The runs list hides and shows runs.

TraceSift shows one finding at a time as a bar or treemap. A summary line and legend of fixed height keep the card from shifting when you move across the chart or open a part, and a Back button steps out of a part. A profile timeline and a ranked list move between findings, and returning from Explore lands on the finding that was explored. The project folder picker is gone: Investigate hands the finding to the chat as written, because the chat already works in the reader's project. Series colours match the recording card, and severity uses the host's status colours where it sends them.

In hosts with composer pills (Codex), Ask about a comparison and TraceSift's Investigate now attach their reference as a pill and send a one-line message, as recording actions already did. Before, they pasted the whole prompt or handoff into the chat. Hosts without pills still get the full text. One helper in host-bridge.ts now attaches the pill for recordings, comparisons and TraceSift.

Telemetry keeps its existing names. New are ui.comparison.change_selected, ui.comparison.memory_scale_changed, ui.tracesift.finding_selected, and the ui.comparison.context_attach and ui.tracesift.context_attach timings. The card no longer emits ui.tracesift.project_*. docs/telemetry.md records each change. No titles, IDs, ranges or prompt text are sent.

Tests cover sharp-change detection, the comparison summary, coverage notes, change list, memory scale, the TraceSift finding, legend, Back and list, returning from Explore, and the pill paths, including fallback without pills, attachment failure, duplicate clicks and timeouts. The UI, MCP app and shared suites passed 446 tests, and the core telemetry suite passed 32. The only visual snapshot nearby is the TraceSift tool's upload pane, which this change doesn't touch; I didn't refresh snapshots locally because Docker wasn't running.

Rebased onto #69, with this change's comments condensed to match. The server tools mobile_get_tracesift_project and mobile_pick_tracesift_project now have no caller in the card. Removing them can be a separate change.

[lauridskern]

Before and after screenshots, rendered from main and from this branch with the same recordings and profiles (light theme, 728 px wide).

Performance comparison

The range table becomes a per-measure summary, and each run's sharp changes are listed in the sidebar.

Before	After

TraceSift, CPU profile

Findings show one at a time, with a timeline and a list to move between them. The project picker is gone.

Before	After

TraceSift, React profile

React commits use the same layout, with the severity beside the duration.

Before	After