Issue 37877 performance reproduction

Platform: Linux x86_64 VM, 8 vCPUs (AMD EPYC 9455), Node v26.13.1.
Base: d405bd94ac45ae94b602a9bdacea08b513771c01 (merge-base of head with origin/main)
Head: 87ece8c7e20f5479470ee97b7e8a2be88f704b9a

Each side uses its own checkout, `vp i`, and a fresh database created by
`T3CODE_DEV_SEED=0 vp run dev`. Run one server at a time. Leave dev tracing
at its defaults. No customer database is used. No production provider is used.
Add that checkout as a project with `node apps/server/src/bin.ts project add
<checkout> --title evidence --base-dir <checkout>/.t3`. Start two threads with
`node apps/server/src/bin.ts thread start . 'turn 1' --title big --provider
devReplay --model command-output --wait --json --base-dir <checkout>/.t3`,
then repeat with title small.

Issue the MCP bearer with `node issue-token.mjs <checkout> <local-token-file>`.
This calls the same issueMcpClientSession method as the thread CLI. Do not upload
the token file. Read the origin and PID from .t3/userdata/server-runtime.json.

Set MCP_URL=<origin>/mcp and MCP_TOKEN_FILE=<local-token-file> for both drivers.
Run `OUT_DIR=<new-output-directory> node grow.mjs <big-thread-id>
<small-thread-id> <server-pid> <checkout>/.t3/userdata/statev2.sqlite 1013`.
Growth sends one real Replay turn and awaits t3_thread_wait before the next.
It records process CPU ticks / persisted-event sequence delta between checkpoints.
The 1.5-second settle delay only separates growth from checkpoint measurements.

After growth ends, run `OUT_DIR=<another-new-output-directory> node measure.mjs
<big-thread-id> <small-thread-id> <server-pid> <before-or-after> <commit-sha> <checkout>/.t3/userdata/statev2.sqlite`.
It initializes one long-lived plain Node streamable-HTTP MCP client, warms each
thread with 20 reads, and alternates thread order across seven rounds of 100 reads.
Each read uses {limit:20, runLimit:1, maxCharsPerItem:2000}, with the default page.
Latency measures the complete HTTP MCP call. CPU measures server process utime
plus stime via /proc/PID/stat, with CLK_TCK=100, divided by 100 calls per round.
Every response is checked against the initial semantic response for its thread.
The semantic response excludes environment-specific IDs, links, and timestamps;
its SHA-256 must match across base and head for each fixture. Latency samples,
round medians, CPU per call, and semantic responses are kept in measurement.json.

Repeat exactly on the other commit with a new database. Compare the median of
seven round medians and their min-to-max spread. Do not use unit tests as proof.

The scenario runner was not used: docs/operations/scenarios.md explicitly
excludes timing/performance regressions. This is a real server/CLI proof.
The named head changes bounded thread reads, not SQLite cache pragmas.
The large 15.5 GB database condition is not reproduced by this synthetic fixture.
