Motir

Produce the six figures the fast lane's latency budget is judged against, from inside the motir-core deployment, and record them on this card. No PR — this is an operator action, marked done on Yue's confirmation. Transcribing the figures into lib/jobs/latencyBudget.ts is MOTIR-3464's job and explicitly not this card's.

Why this is its own card

It was scope on MOTIR-3463, whose cutover finished on 2026-08-26 with every other criterion discharged. The reading needs HOURS of whole-lane tail that the cutover does not — the lane became whole at 15:36:59Z — and leaving it attached kept MOTIR-3463 open, which held MOTIR-3475 and MOTIR-3418 blocked on work they do not need. Splitting it lets the cutover close on what it actually achieved.

⚠️ WAIT for a tail worth reading — this is the card's main judgement

The budget is a p95, and a p95 over a quiet evening is not the same statistic as a p95 over a working day. docs/decisions/job-lane-occupancy.md §6 attributes the Inngest tail to arrival burstiness, so a window containing no burst measures the thing the budget was NOT written about.

Measured on the day of the cutover, work-item/transitioned arrived at 12Z n=1 · 13Z n=12 · 14Z n=39 · 15Z n=14 — a 39× spread across four consecutive hours. Take the reading over a window that spans at least one working period, state the window, and say whether it contained a burst. A --hours 24 window taken late in a working day is the cheapest way to get one.

⚠️ Two traps that produce a WRONG answer rather than an error

Both were hit and proved on 2026-08-26; neither announces itself.

1. The pooled url returns nothing and says so in a sentence that sounds like a measurement. scripts/experiments/engine-fastlane-lag.mjs:161 reads process.env.DATABASE_URL, which inside the machine is the POOLED url — motir_app, rolbypassrls = false — while job_run / job_event are relforcerowsecurity. Every policy matches nothing, so the script prints:

ledger rows matched: 0
samples: 0 — no fast-lane engine runs in this window. Nothing to report.

over a window containing dozens of successful engine runs. Override with DATABASE_URL_UNPOOLED. Filed as MOTIR-3593; if that lands first this trap is gone, so check its status before working around it.

2. The script is not in the image, and it self-guards on its own filename. /app/scripts does not exist — the runtime is a Next standalone output. Upload the file, and the basename must end engine-fastlane-lag.mjs:

if (process.argv[1] && process.argv[1].endsWith('engine-fastlane-lag.mjs')) { await main(); }

Saved as anything else it exits 0 and prints nothing at all. Also: curl is not installed in the image — use node with fetch for any HTTP.

The command

git -C <motir-core> show origin/main:scripts/experiments/engine-fastlane-lag.mjs > /tmp/lag.mjs
B64=$(base64 -w0 /tmp/lag.mjs)
fly ssh console -a motir-core --machine 7817663f103648 -C \
  "sh -lc \"echo $B64 | base64 -d > /app/engine-fastlane-lag.mjs && \
    DATABASE_URL=\\\$DATABASE_URL_UNPOOLED node /app/engine-fastlane-lag.mjs --hours 24 2>/dev/null\""

⚠️ Every fly secrets set replaces the containers and wipes /app, so re-upload after any roll — and prove the probe answers before trusting an empty result. A silent instrument and a silent system are indistinguishable.

Acceptance criteria

• All six figures are recorded on this card verbatim: measuredOn, windowHours, samples, medianMs, p95Ms, maxMs.
• The exact command that produced them is recorded beside them, including which connection string was used.
• The window is stated, and whether it contained an arrival burst — with the per-hour job_event counts for work-item/transitioned over that window as the evidence, not an assertion.
• samples is greater than zero and the card says how many events the window held, so a thin sample cannot be mistaken for a good one.
• The figures are compared against the recorded Inngest baseline (median 1 300 ms, p95 29 400 ms, max 93 300 ms, n=556, 72 h) and against FAST_LANE_LATENCY_BUDGET.p95Ms (5 000 ms), stating plainly whether the budget is met.
• lib/jobs/latencyBudget.ts is NOT edited by this card.

Context refs

• scripts/experiments/engine-fastlane-lag.mjs — the probe, its HOW TO RUN block (currently wrong on both counts above), and the :161 connection-string read
• lib/jobs/latencyBudget.ts — the budget and the Inngest baseline this reading is compared against; NOT edited here
• docs/decisions/job-lane-occupancy.md §6 — the burstiness attribution the window must be chosen against
• MOTIR-3593 — the probe defect; may remove trap 1
• MOTIR-3463 — the cutover that made the lane whole, and where this scope came from