Shell pending-items performance proof

Requirements: Linux x86_64, Node 26.13.1 (node:sqlite StatementSync.stat/resetStats),
git, Python 3 with Pillow, vp, and two checkouts with dependencies installed via vp i.
Base: 8edefc8818aed1222428af5cf4d92df3bf566689
Head: 6538014ca1a68623fef81735760f9d2d61cc9bfb

Keep these scripts together in a scratch directory outside the checkouts.
Run:
  python3 instrument.py BASE_CHECKOUT HEAD_CHECKOUT
  python3 driver.py BASE_CHECKOUT HEAD_CHECKOUT
  python3 report.py RUN_DIRECTORY_PRINTED_BY_DRIVER

The instrumentation.patch shows the same measurement-only edits applied to both
commits. It times the actual pending-items statement in the real SQLite reader
worker, from iterate() through exhaustion; output IO occurs after timing.
VM steps come directly from StatementSync.stat('vmStep'), reset for each read.
No query, binding, planner setting or returned result is changed by the instrument.

The driver starts each real server with a fresh private home and no automatic
seeding. It adds a synthetic git project through the CLI, starts a local ACP
provider, and sends BULK1000, BULKHANG, Stop, BULK14000 and BULKHANG. The provider
emits real command tool calls through the adapter and orchestrator. No SQL rows
are inserted by the scripts. There is also one completed workspace-preparation
command, giving 1,001 and 15,001 completed command items.

Each measurement drives authenticated GET /api/orchestration/shell, the shell
snapshot endpoint used on reconnect, without resumeFrom so it reads rows anew.
Each history size receives 5 warmups and 5 groups of 21 measured reads. Every
read must return exactly one running command. The report compares its functional
fields (fresh UUIDs and timestamps differ between commits).

before.json and after.json contain all 105 statement timings and VM step counts
per scale, HTTP request times, query plans from the same node:sqlite version,
the executed SQL and binds, and the final pending row. PNGs are readable captures
of the report output. Timing claims refer to the live SQL statement, not browser
reload or whole-app latency. The driver stops every server it starts.
