feat(web-shell): show the trajectory over real time, idle included - #12607

What this PR does

The trajectory overview strip gets a second way of laying out time. By default it still cuts out the idle time between turns, so the strip answers where the running time went. A new clock button on the axis, before the zoom buttons, switches it to real time: idle stretches stay on the axis, the two ends of the axis read as local clock times, and the right-hand label says how long the window ran and how much of that was activity, for example "1m elapsed, 1.8s active". The button is a pressed-state toggle. Zoomed windows are labelled with clock times at the precision the window needs, down to milliseconds. The switch is announced through the strip's existing status line.

On a real-time axis a dragged selection can fall entirely inside idle time. The table then says "No request or tool ran in the selected time." with a clear button beside it, and the header reads "Showing 0 of N rows". This is the state the first selection PR left out, because on the compressed axis every stretch touches some record.

Switching mode drops the current selection and zoom. Both are held in the coordinates of the axis they were made on, and the other mode lays the axis out differently, so the same numbers would point at a different stretch of the run. The selection is now tied to the timeline it was drawn on rather than to the loaded window. It lapses in the same render the axis changes, whether that comes from a refresh, another session or a mode switch.

Why it's needed

Part of #12293. Active time answers "was it the model or the tools", but not "how long did this run actually take, and where did it sit waiting". Real time answers that: long pauses before a retry, or between a tool result and the next request, become visible as gaps. This follows the actualTime switch in deepseek-harness's trajectory overview.

Reviewer Test Plan

How to verify

1. Open the Trajectory tab on a session with a few turns. Press the clock button on the overview's axis. The spans spread out with gaps between turns, the axis ends read as clock times, and the right-hand label reads "… elapsed, … active".
2. Zoom into a gap with the wheel and drag inside it. The table is replaced by the "No request or tool ran" message; its clear button restores every row.
3. Make a selection, then switch mode. The selection and zoom are gone and the table is whole.
4. Press the clock button again. The axis goes back to active time and starts at 0.
5. The browser specs under "real time" in web-shell.trajectory.spec.ts drive all of this with real mouse input.

Evidence (Before & After)

Captured by CI's Capture web-shell visuals job on head efc4903: the same session as the existing trajectory captures, switched to real time. Turn 1, with its failed request, retry, read and delegation, sits at the left. Turn 2 sits at the far right after the idle gap. The axis starts at a clock time, the right-hand label reads "1m 18s elapsed, 18.8s active", and the clock button is lit.

Dark	Light

web-shell E2E Smoke on the same head: 192 passed, and the log shows all four "real time" cases below executing. The export renderer bundle measured 1,944,135 bytes in that run, under the 1,970,000 warning line.

Browser specs added, all @smoke:

• spreads the run over real time when switched: first span at the track's left end, last at its right, one minute between turns equal to that share of the track within 2 px, axis start reads as a clock time, table and grid position unchanged
• a selection in idle time empties the table and can be cleared: zooms in on a gap, drags inside it, checks the message and "Showing 0 of 160 rows", then clears it
• keeps the turns a real-time selection covers: a drag from the gap before turn 12 to the gap after turn 15 leaves exactly those four turns
• switching back cuts idle out again and drops the selection

Unit tests: 204 pass across the trajectory hook, loader, projection, range helper, panel and overview. Fifteen mutations were each killed by at least one test. They cover:

• the projection: real time still compressing idle, active time reported as the whole span, the axis origin taken from the first listed row instead of the earliest start, the default mode flipped
• the selection: kept across a mode switch, and the empty-selection message removed
• the toggle and axis: toggle without a pressed state, toggle always asking for the same mode, axis start still printed as 0, clock times ignoring the origin, fractional digits dropped, a 24-hour cycle reading midnight as 24
• the zoom and wording: zoom kept across a switch, switch not announced, selected range named in durations on a real-time axis

Tested on

Environment (optional)

Unit tests locally on Linux. The browser specs run in CI's web-shell E2E Smoke; Chromium cannot run on the machine these changes were written on.

Risk & Scope

• Main risk or tradeoff: "elapsed" is measured from the first drawn record to the last in the loaded window, not from session start. The label says elapsed rather than session length for that reason. On a long real-time axis a one-second request is drawn at the 3 px minimum; zoom shows it properly.
• Not validated / out of scope: the mode is not remembered across sessions or reloads. Selections are not converted between modes; they are dropped. Keyboard control of the strip is the next step, T3c in the ladder.
• Breaking changes / migration notes: none. The default is unchanged. Only the Trajectory tab changes, and hosts opt into it. No SDK, daemon or protocol change.

Linked Issues

Part of #12293