Real-app TAR preview proof

Run each checkout's dev server with T3CODE_DEV_SEED=0. Use vp i first.
Generate fixtures: python3 seed.py FIXTURE_DIRECTORY
Initialize that directory as a git repository. Add it through each app's project CLI:
node apps/server/src/bin.ts project add FIXTURE_DIRECTORY --title "Archive preview proof" --base-dir .t3
Seed a replay thread through each app's CLI:
node apps/server/src/bin.ts thread start FIXTURE_DIRECTORY go --title "Archive proof" --provider devReplay --model multi-turn --wait --json --base-dir .t3
Pass the returned thread ID and actual web URL to the driver:
node measure.mjs CHECKOUT WEB_URL THREAD_ID base mixed.tar OUTPUT_ROOT 5
Repeat for head mixed.tar, base control.tar, and head control.tar (five runs each).

Each run uses a fresh Chromium browser; the same synthetic workspace is served by each isolated database.
Worker startup is paused through CDP only to install identical timing listeners; timing starts on open-message dispatch and ends at archive reply. Three forced GCs precede Runtime.getHeapUsage. GC is outside the timed interval.

copy-probe.mjs adds a wrapper around Uint8Array.prototype.slice, counts returned payloads exceeding 10 MiB, and exercises the same app once per side. It hides the file explorer after measurement to capture readable stills. These separate probe timings are not included in the five-run timing statistics.

Original five-run batch videos are cropped to the Files panel and trimmed to begin at the visible Files listing; playback speed is unchanged. Stills are from the separate copy probe, after two matching screenshots 500 ms apart.

comparison.json contains every measurement from the four final five-run batches. No completed run in those batches was dropped. Earlier navigation probes and partial batches were followed by the final full batches after fixing a pairing-reload race.

Worker parse medians: base 15.6 ms, head 0.9 ms; full spreads 10.9 ms and 0.2 ms. End-to-end click-to-listing medians are 909.6 ms and 865.0 ms, but their ranges overlap and the difference is smaller than the spread, so no user-visible speedup is established.