Adaptive auto-throttle (AIMD) — steer rate by target latency

[NitinKumar004]

What this adds

Adaptive auto-throttle. Turn it on for a task and Marathon watches the target database's latency and adjusts the rate automatically — backing off when the DB is under load and easing back up when it recovers, using the same additive-increase / multiplicative-decrease law TCP uses for congestion control. Your rate becomes the ceiling.

This is the "polite to production" capability: instead of guessing a safe fixed rate, Marathon finds it and re-finds it as conditions change.

How it works

Before each batch (rate-limited to one probe per 250ms), the worker times a trivial round-trip to the target (SELECT 1) as a health signal. That latency feeds the AIMD controller, which sets the token-bucket rate within [min, ceiling]. When the DB is busy even the probe slows, so the controller reduces the rate; when it's healthy, the rate ramps back toward the ceiling.

• Works in solo and fleet mode (same controller in both the sequential runner and the lease worker).
• The AIMD controller was already implemented and unit-tested; this PR wires it into the workers behind a per-task adaptive flag (migration 011).
• The dashboard speed slider becomes the ceiling for adaptive runs, and still applies live.

Evidence

• TestIntegration_AdaptiveRunCompletes — an adaptive run applies every row exactly once (the feedback loop never breaks correctness).
• TestIntegration_AdaptiveBacksOffUnderLatency — with ~80ms probe latency injected, the controller drives the rate well below its ceiling, and the run still completes correctly.
• Existing AIMD unit tests cover the control law (back-off on spike, gradual recovery, ceiling/floor clamping).

Screenshot

Captured with headless Chrome. The run console shows the AUTO throttle indicator and "ceiling" wording:

Notes

• The probe is a cheap SELECT 1; on a busy DB it slows along with everything else, which is exactly the signal we want. Fleet mode could later use replication lag as an additional signal.
• Opt-in per task via a checkbox in the new-operation wizard ("Adaptive auto-throttle"). Existing tasks default to off — no behavior change.