Synthetic issue-linked-threads RPC performance evidence

Environment: Linux x86_64, AMD EPYC 4585PX (8 visible vCPUs), Node 26.13.1.
Each commit runs its own real T3 dev server and fresh, unseeded T3 home.
No tracker requests or paid provider turns are involved.

Install each checkout with vp i. Start each server from its checkout:
T3CODE_DEV_SEED=0 node scripts/dev-runner.ts dev:server --home-dir <fresh-home> --no-seed --port <port>

Copy driver.ts, stable-driver.ts and small-driver.ts into apps/server, then run
from apps/server (names can be prefixed evidence- as in the actual invocation):
node driver.ts <home> <origin> <initial-output.json>
MEASURE_COUNT=11860 node stable-driver.ts <home> <origin> <stable-output.json>

For the independent small fixture use a second fresh home and server:
node small-driver.ts <small-home> <origin> <small-output.json>

The original driver seeds 1186 then 11860 threads and records initial batches.
The stable driver measures the existing 11860 fixture after 200 alternating
hit/miss warm-up pairs per batch. The small driver seeds only 1186 threads and
uses the same 200-pair warm-up. Each of five batches has 50 alternating hit and
50 miss RPCs, with the same serial workload for both commits.

Threads have deterministic IDs and titles. Every 20th has an issue link;
every 400th links key 4242, all others link unrelated keys. Miss key: 999999.
All data is created through project HTTP API and orchestration.dispatchCommand.
Measurement includes the real issues.linkedThreads WebSocket round trip.
Every response is checked for expected count and stable ordered JSON; the
result hashes are compared across commits by render.py.

Initial short-warmup batches are retained separately and are not used for the
final verdict. PNG reports are rendered from final raw JSON, not browser views.

Final serial pass: run-final.py <base-checkout> <head-checkout> <output-directory>.
It uses final-driver.ts (copy into apps/server as evidence-final-driver.ts),
MEASURE_COUNT to reuse fixtures, and SAMPLES=250. Each fixture has five batches
with 200 warm-up pairs followed by 250 hit and 250 miss calls.
The final four JSON files are the source of the captures; earlier runs are
retained as preliminary observations, not selected for the final summary.
