Evidence procedure

Build both commits independently in separate worktrees with `vp i` then `vp run build:desktop`.
Base: b387e3a179d7495321019363c5ba11e847834cd0
Head: 8c676078c53763f74c94687e5e313b3820076c07

Usage:
  MODE=navigation headless-desktop node driver.cjs <checkout> <base|head> <new-output-directory>
  MODE=mutation headless-desktop node driver.cjs <checkout> <base|head> <new-output-directory>

The driver launches the actual compiled desktop main process and bundled renderer in Electron 44.10.1,
with --no-sandbox --disable-gpu. Playwright's Electron API supplies renderer key events. Each launch starts
its own backend. T3CODE_HOME points at checkout-local isolated state, T3CODE_DEV_SEED=0 prevents live
state copying. The two homes were separately created by their own commits; no database copying.

The first launch seeds a project and nine user messages through the real project/thread CLI, running
Replay (dev), multi-turn, after each message. Every user message has 3,000 lines of ten 'the', 30,000
matches per message, 270,000 total. Subsequent launches reuse only their own stopped synthetic thread.

Window: 1440x900; light theme; Linux x86_64; 8 vCPU; 16 GiB RAM; Xvfb.
After opening the thread by its application URL, the driver scrolls the timeline to top, presses Ctrl+F,
types 'the', and waits for the full count. Workload starts three seconds after the counter appears.
Navigation: 300 Enter presses, awaiting each Playwright press then pausing 40 ms.
Mutation: a fixed external div appended to document.body; textContent changed every 100 ms for 20 s.
Each workload is a fresh Electron launch. The order is base-navigation, head-navigation, base-mutation,
head-mutation, repeated three times. Workloads run sequentially on the same machine.

Renderer RSS is read from /proc/<renderer-pid>/status every 50 ms outside the renderer. Duration is
wall time of the scripted workload, including deliberate pauses and automation round trips; it is not
CPU time or a display-frame latency measurement. Registered CSS Highlight ranges, total match count,
active range count and visibility are captured before and after. The head does not change the count or
navigation: all 12 controlled runs start at 90001/270000; navigation ends at 90301/270000, mutation at
90001/270000. There is one active, visible range throughout those endpoints.

Videos record the workload and settling period. The private environment display label in the details
card is covered in published videos and screenshots; the chat, counter, highlights and action are
untouched. Published stills originate from two identical screenshots 500 ms apart after the workload.
measurements.json contains all twelve runs and raw RSS samples, not pilot runs.

Two extra control launches (one per commit) repeat the same 20-second mutation with the search query
cleared and find closed. Their raw samples appear separately under closedFindControls. These are
controls only, not included in the three-run medians. One interrupted mutation capture was rerun
because its final stills did not match; completed measurements and capture pairs only are reported.

measurement-driver.cjs produced the primary three-run series. driver.cjs adds a recorder-ready wait
before measurement. Published recordings and their stills use separate runs of driver.cjs so they
include the entire action. These extra capture runs appear separately under recordingRuns, excluded
from the three-run medians. The search workload itself is identical.
