fix(stall): T2 amplification kill — resume-edge pacing + ABR starved-window guard #53

Merged

enricobuehler merged 3 commits from worktree-stall-ride-through into main 2026-08-05 06:35:32 +00:00

Conversation 0 Commits 3 Files Changed 3

+234 -45

enricobuehler commented 2026-08-05 05:47:34 +00:00

Owner

Copy Link

Copy Source

Stall program T2 (planning: design/vdisplay-disturbance-immunity.md §9). A 300–750 ms capture stall must degrade the stream by its own duration, not by seconds of crater.

The mapping (what turns a hole into a crater): a pure capture stall creates no frame-index gap and no loss — the recovery machinery correctly stays silent. The damage is all at the resume edge: the first frame after the hole carries seconds of scene delta, and the native pace budget min(0.9 × time-to-deadline, overflow @ 3× rate) let the deadline term clamp a multi-interval overflow into the remainder of one interval — a line-rate blast that overruns the socket tx-buffer (the field WSAENOBUFS(10055) + loss_ppm edge-spike fingerprint across all three RDNA4 standby-sink cases). The self-inflicted loss then costs recovery-IDR round trips (each another oversized blast), and the ABR reads dropped/flushed/kf-storms as SEVERE — ×0.7 per event — and since #36 could even latch a phantom decode cap from a pair of stall-shaped windows at the same rate.

Commit 1 — fix(host/send): send_pacing::native_budget() (pure, unit-tested). With the rate cap active, the overflow's wire time at the proven ~3× stream rate is the budget, bounded by a 100 ms ceiling; the deadline never under-cuts it. Steady-state frames produce byte-identical schedules (the rate term already decided there); PUNKTFUNK_PACE_FACTOR=0 keeps the legacy deadline-only spread; the GameStream plane's Moonlight-pinned schedule is untouched.

Commit 2 — fix(core/abr): a deciding window that delivered under ¼ of target (capture stall, outage) still backs off, but takes the same "not a knee sample either way" arm as a draining backoff — it neither latches a decode cap nor erases the reference a genuine choke set. Two new tests model the 5 s stall cycle (no-latch) and reference-preservation around a starved interruption.

Verified no-change-needed: keyframe-storm capping already ships (IDR cooldown 750 ms/2 s + RFI-echo swallowing); in-place reconfigure_bitrate (no IDR) already exists for NVENC/AMF, so the SetBitrate-rebuild cost is confined to fallback backends.

Gates: core 346 lib tests green (--features quic; 44 abr, 2 new) · host clippy --all-targets -D warnings + 9 send_pacing tests green (linux/amd64 container) · full CI board green on the PR head (the earlier windows reds were main's pad-audio audio_wasapi breakage, fixed in #57).

Post-merge follow-up: field A/B with the case-#3 reporter (LG UltraGear / RX 9070 XT) — expected: freeze duration ≈ raw stall duration, no edge loss, no bitrate crater.

Stall program T2 (planning: `design/vdisplay-disturbance-immunity.md` §9). A 300–750 ms capture stall must degrade the stream by its own duration, not by seconds of crater. **The mapping** (what turns a hole into a crater): a pure capture stall creates no frame-index gap and no loss — the recovery machinery correctly stays silent. The damage is all at the **resume edge**: the first frame after the hole carries seconds of scene delta, and the native pace budget `min(0.9 × time-to-deadline, overflow @ 3× rate)` let the *deadline* term clamp a multi-interval overflow into the remainder of one interval — a line-rate blast that overruns the socket tx-buffer (the field WSAENOBUFS(10055) + `loss_ppm` edge-spike fingerprint across all three RDNA4 standby-sink cases). The self-inflicted loss then costs recovery-IDR round trips (each another oversized blast), and the ABR reads `dropped`/`flushed`/kf-storms as SEVERE — ×0.7 per event — and since #36 could even **latch a phantom decode cap** from a pair of stall-shaped windows at the same rate. **Commit 1 — `fix(host/send)`**: `send_pacing::native_budget()` (pure, unit-tested). With the rate cap active, the overflow's wire time at the proven ~3× stream rate *is* the budget, bounded by a 100 ms ceiling; the deadline never under-cuts it. Steady-state frames produce byte-identical schedules (the rate term already decided there); `PUNKTFUNK_PACE_FACTOR=0` keeps the legacy deadline-only spread; the GameStream plane's Moonlight-pinned schedule is untouched. **Commit 2 — `fix(core/abr)`**: a deciding window that delivered under ¼ of target (capture stall, outage) still backs off, but takes the same "not a knee sample either way" arm as a draining backoff — it neither latches a decode cap nor erases the reference a genuine choke set. Two new tests model the 5 s stall cycle (no-latch) and reference-preservation around a starved interruption. **Verified no-change-needed**: keyframe-storm capping already ships (IDR cooldown 750 ms/2 s + RFI-echo swallowing); in-place `reconfigure_bitrate` (no IDR) already exists for NVENC/AMF, so the SetBitrate-rebuild cost is confined to fallback backends. **Gates**: core 346 lib tests green (`--features quic`; 44 abr, 2 new) · host `clippy --all-targets -D warnings` + 9 send_pacing tests green (linux/amd64 container) · full CI board green on the PR head (the earlier windows reds were main's pad-audio `audio_wasapi` breakage, fixed in #57). **Post-merge follow-up**: field A/B with the case-#3 reporter (LG UltraGear / RX 9070 XT) — expected: freeze duration ≈ raw stall duration, no edge loss, no bitrate crater.

enricobuehler added 2 commits 2026-08-05 05:47:35 +00:00

fix(host/send): a stall-resume frame paces at the proven rate instead of blasting ... e5416646f9

Stall program T2 (amplification kill), the resume-burst half. The native
pace budget was min(0.9 × time-to-deadline, overflow at ~3× stream rate) —
for steady-state frames the rate term is smaller and decides, but for an
OVERSIZED frame (a capture-stall resume carrying seconds of scene delta, a
cold IDR) the deadline term clamped a multi-interval overflow into the
remainder of ONE: an instantaneous many-×-stream-rate blast that overruns
the socket tx-buffer and loses the very frame that would have ended the
freeze. Field fingerprint across three RDNA4 standby-sink cases:
WSAENOBUFS(10055) + loss_ppm spikes at stall edges, then a recovery-IDR
round trip per retry while the client shows 'current bitrate 0.1'.

The budget is now the overflow's wire time at the pace rate itself
(send_pacing::native_budget, pure + unit-tested), bounded by an absolute
100 ms ceiling so a pathological frame can't park the send thread; the
deadline stays a target, never a license to blast. Steady-state frames
produce byte-identical schedules (the rate term already decided);
PUNKTFUNK_PACE_FACTOR=0 keeps the legacy deadline-only spread; the
GameStream plane's Moonlight-pinned schedule is untouched.

Gates: host clippy --all-targets -D warnings + 9 send_pacing tests green
(linux/amd64 container), fmt clean.

fix(core/abr): a starved window is never a decode-knee sample ... e5ca213339

Stall program T2 (amplification kill), the phantom-latch half. A deciding
window that delivered under a quarter of the target rate (a host-side
capture stall, an outage, a mid-window pause) carries starvation-shaped
distress — a jump-to-live flush, a keyframe-ask burst — that the decode-cap
latch read as decoder evidence: under a periodic capture stall (the RDNA4
standby-sink field cases, one stall every ~5 s) every edge offers another
'backoff' at the SAME rate, and one pair latches a phantom decoder knee at
whatever rate the display driver happened to interrupt. The session then
fights the cap's re-probe ladder (+12.5% per 16-128 clean windows) for
minutes on a decoder that was never the problem.

Starved windows still back off (real damage deserves the safe response) but
take the same 'not a knee sample either way' arm as a draining backoff:
they neither latch a decode cap nor erase the reference a genuine choke
set, so a real knee's pair still finds itself around the interruption. The
¼ bar sits deliberately far under the ×¾ utilization bar climbs require.

Gates: 44 abr tests green (2 new: the stall-cycle no-latch scenario and the
reference-preservation scenario), full core lib suite 346 green
(--features quic), fmt + clippy clean.

enricobuehler referenced this pull request 2026-08-05 06:15:21 +00:00

fix(client-core): pad-audio references the WASAPI module by its mounted name #57

enricobuehler added 1 commit 2026-08-05 06:22:48 +00:00

Merge branch 'main' into worktree-stall-ride-through 1d6f4760f3

enricobuehler marked the pull request as ready for review 2026-08-05 06:35:22 +00:00

enricobuehler merged commit 110ac9b663 into main 2026-08-05 06:35:32 +00:00

Sign in to join this conversation.

Reviewers

No Reviewers

Labels

No items

No labels

Milestone

No items

No Milestone

Projects

Clear projects

No projects

Assignees

Clear assignees

FernandoMX enricobuehler (Enrico Bühler) kszlim nannku neoney sassycrown

No Assignees

1 Participants

Notifications

Subscribe

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: unom/punktfunk#53