The controller's target is only a promise: on calm content the encoder
emits a fraction of it, every window looks clean while proving nothing,
and the climb drifts the target into rates the decoder has never seen.
The first motion spike is then the first real test — it fails, and the
decoder is overloaded for the two-window backoff latency (the reported
settle-calm-then-spike stutter).
Three changes, all client-side (no wire/ABI impact, old hosts unaffected):
- Climb gate: a clean window authorizes a climb only when its ACTUAL
delivered throughput reached 3/4 of the current target — the target was
genuinely tested. Calm windows still bank clean credit; the first
loaded window after a clean run climbs immediately.
- Proven-throughput cap: climbs step at most x1.5 past the session's
high-water mark of delivered-and-digested (decode-latency-flat)
throughput, so slow start becomes a bounded experiment instead of a
blind doubling. High-water never decays: calm periods keep a validated
target, so returning motion gets the full rate instantly; shrinking
capacity (thermals) stays the reactive decode signal's job.
- Severe decode excursion: a >45 ms-over-baseline decode spike backs off
after ONE window instead of two — the overload is already on screen.
The pump feeds the window's wire-byte throughput; the byte baseline is
rebased when the startup capacity probe completes so FLAG_PROBE filler
can't poison the proven mark with the link rate.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
enricobuehler
6824c1cc0c