# SQLite V1 upgrade proof

Linux x64; Node 26.13.1. The web client was Chromium, 1280x800 CSS pixels,
light theme, device scale 2. Both checkouts were installed with `vp i`.
The runtime proof uses the actual server, not tests or an imported service function.

1. Download the published `t3@0.0.45` Linux x64 release and start `t3 serve`
   with a new, explicit `--base-dir`. Pair the browser. Decline importing provider history.
2. Initialize an empty local git project, add it with `t3 project add`, and send
   these two prompts using Codex GPT-6-Luna (the old release has no dev replay):
   - Write a four-line poem about SQLite pages returning to disk. Do not use tools.
   - Now give one sentence explaining what a freelist is. Do not use tools.
3. Wait for both completed responses. Stop the V1 server and wait for process exit
   before copying its database, so committed WAL content has reached the main file.
   The seed had auto_vacuum=0, page_count=81, freelist_count=0, page_size=4096,
   one thread and four messages (two user, two assistant).
4. Copy this stopped `state.sqlite` into two fresh homes' userdata directories.
   Start each checkout with `T3CODE_DEV_SEED=0`, an explicit isolated home, and
   automatic project bootstrap disabled. Never use live user data.
5. Pair each web client, decline provider-history imports, and open the migrated
   thread directly using its environment and thread identifiers. Both clients
   rendered all four original messages. Take two screenshots 500ms apart;
   only retain a still when they are pixel-identical.
6. Observe normal maintenance with read-only PRAGMA queries every five seconds.
   Do not issue VACUUM, delete rows, trigger maintenance or alter its clock.
   The server logs on both sides show one imported/hydrated thread and two
   imported assistant messages, then compaction deleting 26 events and 25 receipts.
7. `observe.py` captures the web runs. `live-proof.py` repeats the same first-time
   import on fresh copies with the real `serve` command and records its live,
   read-only console viewer using Xvfb/wish/ffmpeg. Its arguments are documented
   at the top. The raw video is a live screen recording at normal speed; the
   short copy is explicitly labelled 10x playback. No frame is reconstructed.
8. Continue observing multiple once-per-minute reclaim passes. Stop both servers
   and inspect again. Compare SHA-256 hashes of both original V1 files against
   the seed; no source file should change.

Base: ed101ef641fe75e1a86c7ff7ef34a09955765307
Head: 8b1156cd1abbd0dce33432bef62f3d18b7c09806

The V1 seed SHA-256 was
50b5ee48f4cdb5ec270fb19c1b3840493db0d8bf95e90c0e58556fabec634605.
Database files are not uploaded because they contain local workspace metadata.
Only synthetic conversation screenshots, safe selected log events, read-only
measurements and portable drivers are published.
