Each finished broadcast is filed locally — tier, probed encoder, rated, settled, steps
spent — and Models/RateHistory turns that into at most a suggestion. The app never
applies one: it asks before the next stream, "Use it" sets the rate, "Keep mine"
remembers that exact number so the same question does not return until the evidence
moves. Nothing leaves the machine.
Two rules keep it from nagging. The high-water mark is the highest rate that
*finished* a session with zero steps, not the highest seen, so one easy session (a
menu screen, a quiet game) cannot set a target the machine cannot hold. And it is
asymmetric: raising needs 10% above the current rate across five clean sessions of the
same tier and encoder, lowering needs three step-downs in five — fixing damage should
not wait for a percentage, and risking it should.
The encoder key is real, not decorative: StreamHealth.VideoEncoder is now filled from
the probe (FfmpegEncoder), so history recorded against NVENC is not applied to a
machine that later runs QSV. At startup, before anything has streamed, the last
session on that tier is the honest guess.
672 pass, 0 warnings.
The tier table cannot know a creator's connection, so it should not pretend to. The
Mbps box under the picker is theirs: the tier chooses geometry and frame rate, the
box chooses bits. Stored per tier (`Video.Rate.<tier>`), so tuning 1080p60 does not
touch 720p30, and none means "the shipped default" rather than zero.
Clamped at the tier's YouTube minimum (`RateLadder`), because a target under it
cannot hold the resolution; no ceiling otherwise, because a wrong guess is exactly
what live step-down exists to catch, and a creator who types 25 on a fibre line
should be allowed to find out. The tooltip carries YouTube's number for the tier.
`MainViewModel.EffectiveQuality` is the single substitution point, and the encoder
options, the step-down policy and the clip buffer all read it — so a rated tier can
never stream at the shipped default because one call site was missed.
One ordering trap found by the test: the constructor resolves the default tier
before the layout store is open, so a read at that point must not cache, or the
shipped default shadows the creator's stored rate for the whole run.
670 pass, 0 warnings.
Step-down used a flat 3 Mbps (4 for vertical) as its floor — a number that had
nothing to do with the tier being streamed. With CBR holding the target, that
floor is a real picture-quality floor, so it has to follow YouTube's published
H.264 table (support.google.com/youtube/answer/2853702): 6 at 1080p60, 5 at
1080p30, 3 at 720p.
Models/RateLadder is that table as pure policy. Vertical is not a separate row:
1080x1920 and 1920x1080 are the same pixel count, so the frame rate alone picks
the band. StepDownPolicy takes the floor as an optional constructor argument and
refuses one above the starting rate, so an unusual start cannot begin under its
own floor. Omitted, the historical 3/4 floor still applies, which keeps the
existing policy tests meaningful and lets a caller opt in.
One consequence worth knowing: the three-step budget can stop above the floor
(12 → 9.6 → 7.7 → 6.2 from a 20% loss), so the floor is a limit, not a target.
669 pass, 0 warnings.
Every 1080 tier carried 8 Mbps, inherited from TASK 4's "mainstream ceiling,
upload headroom stays comfortable". It was never derived from anything: the same
number served 30 fps and 60 fps, landscape and vertical, which no per-tier
derivation would produce. With CBR now actually holding the rate (700c7a8) that
number becomes the picture quality, so it has to be chosen rather than inherited.
New defaults sit between YouTube's H.264 recommendation (17 at 1080p60, 14 at
1080p30, 8 at 720p — support.google.com/youtube/answer/2853702) and the rate
practitioner OBS guides converge on for YouTube (9–12 Mbps at 1080p60). Vertical
mirrors its landscape pair: 1080x1920 is the same pixel count as 1920x1080.
The free-profile fallback in OutputProfile.Resolve moves with it (8.0 -> 10.0) so
a build with no shipped tiers still emits the 1080p30 rate rather than a stale
one.
668 tests (1 audio-pipe flake under full-suite load, passes alone), 0 warnings.
The stream was configured as "@8 Mbps" and averaged 5.1. The cause is the argument
set, not the machine and not YouTube: `-b:v K -maxrate K -bufsize 2K` is capped
VBR. Measured on the bundled BtbN build with a static 640x480 frame at -b:v 2000k:
default rate control 9,858 bytes / 8 s = 9 kbps
-rc cbr 2,001,989 bytes / 8 s = 1,951 kbps
-rc vbr 9,858 bytes / 8 s = 9 kbps
On high-motion testsrc2 the same target produced 2006 kbps (CBR) against 1741
(VBR): VBR undershoots by however easy the picture is, and a composited scene
with a still background is very easy.
Spelling is per encoder, verified rather than assumed:
- h264_nvenc, h264_amf: `-rc cbr`.
- h264_qsv: nothing. It has no `-rc`, and libavcodec/qsvenc.c (select_rc_mode,
FFmpeg d263fce2b2) selects MFX_RATECONTROL_CBR when bitrate == maxrate — which
the triad already sets. Passing -rc here would be an unknown option.
- libopenh264: nothing. `-rc_mode bitrate` changed no measurable output on this
build (3,832 vs 3,830 bytes), so it stays out rather than adding a line whose
effect nobody can demonstrate.
Reference: YouTube's encoder spec requires CBR
(support.google.com/youtube/answer/2853702).
668 tests pass, build 0 warnings.
Records the 5.1-vs-8 finding as an args defect (no -rc, so NVENC ran VBR) rather
than a machine or ingest limit, the four units that follow, and two new landmines.
Assets/starting-background.jpg stays out of this commit: it is the creator's
replacement art, still awaiting the visual check recorded in the landmines.
The frame-rate boundary is correct as a product rule, but it only bites in a
Release build, where DevBuild.IsDevelopment is false and DeveloperMode cannot
unlock anything — exactly the build whose performance numbers are worth reading
(see docs/ai/environment.md: Debug pixel paths are several times slower). 60 fps
could not be tested at all.
OutputProfile.EntitlementEnforced is false: every shipped tier is selectable by
any build, and HasProOutput — the one property the tier setter, the dropdown's
strikethrough and the locked-item style all read — answers yes regardless of
ownership. static readonly rather than const, so the consumed `||` does not fold
into unreachable operand code (same reason as DevBuild.IsDevelopment).
The licence indicator row is suppressed while the switch is off: it exists to
explain what the boundary costs, and claiming "Steam ID Limit: 60fps" to an
unowned build would be a false statement. Docs corrected in place — the enforced
rule is unchanged, only its state is recorded.
Tests: one integration case covers both switch states (unowned build reaches
1080p60 with the switch off, 1080p30 with it on), so flipping at launch leaves no
red test behind. The two suites that anchor on the pinned licence row early-return
while it is absent; EntitlementTests covers that state meanwhile.
667 pass, build 0 warnings.
Add a persisted global App Settings switch. Off keeps the existing per-severity
lifetimes (errors remain sticky); on makes every alert non-expiring and closes
it when its body is clicked. The NotificationService supplies the current
setting to each toast request.
StickyNotificationIntegrationTests covers persistence and both request modes.
618 tests green, clean build with 0 warnings.
Reference: Notification.Wpf exposes a CloseOnClick request option:
https://github.com/G-Stas/Notification.Wpf
App Settings now chooses one of six ToastArea positions (left/right ×
top/center/bottom), defaulting to the existing bottom-right. The setting is
persisted per profile and the toast area moves immediately; edge gutters keep
bottom notifications above the footer.
ToastPositionIntegrationTests covers live repositioning and restart persistence.
617 tests green, clean build with 0 warnings.
Reference scan: OBS Notifier supports user-configured notification placement:
https://github.com/DmitriySalnikov/OBSNotifier
App Settings now chooses one of six ToastArea positions (left/right ×
top/center/bottom), defaulting to the existing bottom-right. The setting is
persisted per profile and the toast area moves immediately; edge gutters keep
bottom notifications above the footer.
ToastPositionIntegrationTests covers live repositioning and restart persistence.
617 tests green, clean build with 0 warnings.
Reference scan: OBS Notifier supports user-configured notification placement:
https://github.com/DmitriySalnikov/OBSNotifier
The FFmpeg ProcessFailed event arrives off-thread. Start recovery on the UI
dispatcher, publish retry status there, and route StopStream back to the
dispatcher before changing bound state or alert collections. This prevents a
recovery exception from becoming an unhandled WPF CollectionView crash.
AttentionAlertStackTests covers worker-thread recovery reporting and stop.
616 tests green, clean build with 0 warnings.
A quality step-down now closes the named audio pipe before stopping/flushing the
pump, then opens a fresh server before restarting the pump. This prevents ffmpeg
from waiting on the old audio input and avoids reconnecting to a broken pipe.
EncoderRestartTests pins the audio-stop → video-stop → audio-start → video-start
order. 616 tests green, clean build with 0 warnings.
App Settings can now restore only the staged canonical scene. The VM saves
pending edits first, resets that scene, reloads immediately, and keeps the same
scene staged. Other scene edits survive; the action is disabled while live or
recording. 615 tests green, clean build with 0 warnings.
Reference: OBS Scene Collections keep scene/source sets distinct:
https://obsproject.com/kb/scene-collections
App Settings now offers a clearly described all-scenes reset. It is disabled
while live or recording and reloads the restored layout immediately. The
integration test drives the real VM command and checks the Settings binding.
614 tests green, clean build with 0 warnings.
Reference: OBS Scene Collections are a distinct saved set of scenes and sources:
https://obsproject.com/kb/scene-collections
Build-mismatch restoration now keeps the currently selected webcam identity,
restores default scene placements against that camera, and leaves cameras absent
when none is selected. The integration regression test pins the creator's camera
choice and restored default geometry.
612 tests green, clean build with 0 warnings.
The per-scene reset restores that scene's settings, layers, webcam placement,
and background-path settings while leaving other scenes' current edits alone.
A focused integration test covers the restore/isolation behavior.
611 tests green, clean build with 0 warnings.
Reference: OBS scene collections store scenes and sources as a saved set:
https://obsproject.com/kb/scene-collections
A mismatched or missing current-build stamp restores the immutable default
snapshot atomically before the VM loads scenes. The new integration case edits
current on build A, opens on build B, and verifies the factory layout returns.
610 tests green, clean build with 0 warnings.
Reference: OBS scene collections keep scenes/sources as a separate saved set:
https://obsproject.com/kb/scene-collections
The matching-build persistence slice is in place: existing layouts are captured
once as defaults, first-run factory scenes are saved and captured before the
creator can edit, and same-build startup continues to load current rows. Default
asset references survive current-layout saves.
One integration test covers the current-layout roundtrip and immutable default.
609 tests green, clean build with 0 warnings.
Reference scan: OBS stores scene collections separately from output profiles;
this split keeps layout state independent in the same SQLite store:
https://obsproject.com/kb/scene-collections
608 green, 0 warnings. Completes the layer half of TASK 65.
- **(+) → Replay Clip** adds an ordinary media layer with an app-owned path, one per scene, silent by
default (`MediaMuted`) and looping only outside Live — the compositor needs no special case.
- **F10 saves → the Live layer plays it once, then hides** (`SessionCompleted` → invisible, guard
cleared). Every other scene with a replay layer keeps looping the kept set, advancing on each clip's
completion, wrapping at the end.
- **The guard the creator asked for:** while that Live replay is playing, **F10 is inert** with a plain
explanation. That is what replaces the impossible "exclude the layer from its own clip" — with one
encoder and two muxes the clip *is* the program, so the fix is to not clip a clip.
`ReplayLayerTests` drives the real VM: one layer per scene, muted by default, once-through in Live,
looping in BRB, and a stray F10 with clips off writes nothing.
Slice 5a, scaffolding only. 607 green.
- `Source.IsReplayLayer` + schema column + roundtrip: a media layer that shows saved clips instead of a
chosen file. The compositor needs to know nothing — it is a media layer whose path the VM owns.
- **Two spec corrections from the creator (2026-11-10), recorded in the task:**
1. The immediate replay of a fresh clip happens **only in the Live scene**. While it plays, **F10 is
inert** ("replay in progress").
2. That guard replaces the "layer excluded from its own clip" idea, which is **impossible** with one
encoder and two muxes: the clip is a second mux of the same frames, so it necessarily contains what
viewers saw. No second encode, no fiction about a clean feed.
Remaining for 65: the add-layer menu entry, `ClipSaved` → play (Live: once then hide; elsewhere: loop the
kept set), `SessionCompleted` → advance or hide, and silent-by-default audio.
Slice 4. 607 green, 0 warnings. Clips remain **off by default** and, while off, the ring is never
constructed — so no reader thread and no encoder output exist until someone asks for the feature.
- **F10** (`HotkeyId.SaveClip`, default F10, remappable in the hotkey dialog) snapshots the ring, writes
the file, prunes to the retention count and toasts the name. Empty ring → "nothing to clip yet"
instead of a zero-byte file; clips off → says so and points at Settings.
- **Settings**: Enable, length (10 or 15, default 10), keep (3/4/5, default 3). Persisted in the layout
store and **coerced on load** — a bad stored value falls back to a legal one, so the ring can never be
sized for a length the UI cannot show.
- The ring is sized from length × tier bitrate × 1.25 and rebuilt if either changes; it is created at
session start and dropped at stop, so memory is only spent while a show is running.
- The drawer states the boundary in the creator's terms: "A capture, not an editor: for trimming,
captions or a library, use a video editor."
Still to come: the replay layer itself (play once then hide in Live, loop the kept clips elsewhere).
Slice 3 of TASK 65. 606 green, 0 warnings.
`ClipLibrary` is the whole of clip "management": a file lands in `clips\` beside the recordings, the
oldest is deleted past the retention count, and `Recent()` returns newest-first because that is also the
replay layer's playlist order. No gallery, no prompt, no naming.
Two guards the tests pin: an **empty snapshot never becomes a zero-byte file** (a save that fails loudly
is better than a file that won't open), and **pruning only touches our own prefix** — another file in the
folder survives.
`RingBytesFor(seconds, bitrate)` is the memory cap the ring is sized with: exact wire size × 1.25
headroom, because bitrate moves and the ring trims whole fragments.
Still to come: the replay layer, the F10 hotkey, and the Settings switch.
Slice 2 of TASK 65. 602 green, 0 warnings. **Clips are still off**: no `ClipBuffer` is passed, so the
ffmpeg command line is byte-for-byte what it was before this commit, and a test asserts exactly that
(the original args are a strict prefix of the new ones).
- `FfmpegArgs` appends a fragmented-MP4 output to `pipe:1` only when `ClipBuffer` is set. One-second
fragments (`-frag_duration 1000000`) rather than the 4 s GOP, so a clip can start inside the window
instead of at the last keyframe. Stats already own stderr, so stdout was free.
- `Fmp4FragmentSplitter` turns that byte stream into whole fragments: `ftyp`+`moov` is the init
segment, each `moof`+`mdat` is a fragment. It parses box headers itself, survives a box being cut
anywhere across reads, and **resyncs instead of throwing** — a clip fault must never take the stream
down.
- `FfmpegEncoder` reads it on its own task, never blocking the frame path. Any exception logs and
disables clips for the session; the broadcast is untouched.
- `IEncoderProcess` gained `StandardOutputBytes` (the seam already had the text reader for probing).
Next slices: the replay layer + retention, then the Settings switch with disabled-but-remembered controls.
Slice 1 of TASK 65, pure memory, no encoder change yet.
A 1080p BGRA frame is 8.3 MB, so fifteen seconds of raw composite is ~7.4 GB — the 15 MB figure in the
spec is the *compressed* stream. So the ring holds fragmented-MP4 bytes that ffmpeg will write to a
pipe, and a save is init-segment + surviving fragments in order.
Two rules the tests pin: the cap is by **bytes** (bitrate moves; seconds do not), and the newest
fragments always win, because a clip that ends at the keypress is the entire point. The init segment does
not rotate and cannot be reused across encoder starts, so `Clear` drops it too — an MP4 without its moov
is noise.
Next slice: ffmpeg writes the buffer output to stdout while `-stats` stays on stderr, and the pump feeds
this ring. Not done here on purpose — an encoder change is the riskiest edit in the app and deserves its
own commit and test.
595 green, 0 warnings. Completes TASK 55 (slice 1, log rotation, shipped 2026-10-22).
- **`AlertClipCleanup`**: the default alert clip is materialised into `%TEMP%` for ffmpeg to read, and
nothing ever deleted it — one orphan per profile that ever played an alert. Now: delete this session's
file in `Shutdown`, and sweep `ytLive-alert-*.mp4` at startup for anything a crash left behind. Only
the prefix we own is touched; a directory is never deleted.
- **`DiskSpaceGuard`**: a recording that dies of a full disk is a trust breach, so the numbers are a
pure policy — refuse to start under **1 GB** (toast; the stream still starts, recording does not),
warn under **5 GB** at start ("about eighty minutes" at 8 Mbps ≈ 3.6 GB/h), and one mid-show warning
under **2 GB** — the last useful moment before the encoder fails for a reason it won't explain.
Unreadable drives (a USB stick that just left) return null and never block a running recording.
Tests are policy-level on purpose: filling a real drive is not a test. `DiskHygieneTests` proves the
sweep only removes our own files and that the thresholds behave at both ends.
591 green, 0 warnings.
- `Scene.SceneMutesMic` / `SceneMutesDesktop` are **nullable booleans**: null = leave the mix alone, so
a scene is passive until the creator says otherwise. Schema columns added, NULL for every existing
layout except one deliberate default.
- **Applied once when the scene takes the PROGRAM output** (`TransitionToScene`, only when live or
recording). Staging a scene offline is auditioning, not a stream edit — the test asserts that.
- **The creator is boss:** a manual mute survives any transition into a scene with no opinion; a scene
that does have one wins on its next entry. `SetMicMuted`/`SetGameMuted` are idempotent so a repeated
"mute" cannot clobber the level that unmute would restore.
- **BRB ships muted** (creator spec) for new layouts *and* via a one-time migration for existing rows
where the creator has never expressed an opinion.
- UI: two three-state checkboxes under the layers list ("When this scene goes live"), styled with
`PropertyPillCheck` so the theme test holds.
`SceneAudioTests` covers the whole rule in one integration test: default state, offline staging does
nothing, live entry mutes, manual unmute survives a passive scene, an opinionated scene wins, desktop
untouched, and all three states roundtrip through save/reload.
Needs the creator's eye: the checkboxes are new chrome in the left panel (indeterminate = passive).
Creator-supplied capture of the main window: five scenes across the top, the Starting scene with its
three callouts in the preview, the layer panel, and the audio dock.
Re-encoded with the bundled ffmpeg instead of the WPF resize tool (scale 1400 + q4) — one command, the
same tool the rest of the assets use — and that shortcut is recorded in MyMistakes next to the original
recipe so it is not re-derived a third time.
Creator ruling 2026-11-09: finish the queue before the new features, gold pass always last.
**Retired (platform's surface, not ours):**
- 34 scheduled streams — YouTube owns scheduling *and* the reminder that makes it worth anything.
- 50 publish recording — Studio owns upload (resumable, quota-aware, explains rejections), and an
in-app uploader would widen the OAuth consent to whole-channel video management for one convenience.
The app adds "Show in folder" instead.
**New tasks:** 65 replay clips (the creator-side buffer YouTube does not give you), 66 timed ticker
messages, 67 per-event alert videos, 68 edit-while-live + first-run On-Air.
**TASK 65 spec** is the interesting one: 15 s is a clip (10/15/20, default 15); the ring holds encrypted
*encoded* fragments in RAM — raw 1080p frames would be ~7.4 GB for 15 s — so there is no steady disk
write and no crash debris; F10 saves; the replay layer plays once then hides in Live and loops the last
1–3 elsewhere; silent by default; three kept; the layer is excluded from its own clip (hall of mirrors).
**TASK 14 gains Branch 3:** mic input gain ahead of the gate (table boom mics fall under it), webcam
blank hotkey, opacity/geometry parity with a collapsed Geometry disclosure, Show-in-folder.
**Decisions recorded:** the replay-buffer reversal (it was retired in 2022-era reasoning as "different
product"; the case for it won), the platform-surface rule, and voice triggering deferred — always-
listening in a streaming app hears the game, so the content is its own false-positive source.
Queue refactored to 56 lines by archiving completed rows and the superseded surge plan; HANDOFF at 40.
Last slice. 590 green, 0 warnings.
- **Shed before you charge the viewer.** On machine evidence the policy's first act is
`ShedCaptureCadence`: `ScreenCaptureFrameSource` widens its conversion interval 10 ms → 33 ms (≈100
→ 30 conversions/s) through a new `ScreenCaptureManager.ShedLoad`. The output tier, the declared fps
and the viewer's picture are untouched — we stop paying for readbacks we were throwing away. It
costs no step budget; if the starvation persists, the fps ladder runs on the next clean window.
- The shed is reset on every stop path, so a cheap session cannot leave a later one capped.
- **The session line is a Logs-tab line, not a session-report line.** The session report is the money
summary (Super Chats, per currency); quality chatter there would bury the numbers it exists for. The
cause-tagged line already lands in the log at the moment of the step.
TASK 33 is complete: two ladders, budgets, floors, ambiguity, one-way with recovery behind a flag, the
relaunch, the copy, and the cost shed. 18 tests pin the ladders, the copy and the restart.
Slice 3. 589 green, 0 warnings.
- **`FramePump.ChangeEncoderAsync`** is a full pump stop/start, not an encoder swap. A swap leaves an
in-flight submit pointing at a disposed process, and that faults the loop — which ends the
broadcast, the one thing a quality step-down must never do. `StopAsync` already awaits the loop and
flushes, so this reuses the tested path and costs one short freeze, which the ladder accepts.
- **`MainViewModel.StepDown`**: one planner per broadcast, subscribed to `StatsReported`, unsubscribed on
every stop path. The pump thread only decides; the restart is posted to the dispatcher, because the
act stops the loop that raised the event.
- **The creator's choice is not rewritten**: `_encoderTierOverride` changes what the encoder emits while
the dropdown keeps showing what they picked, and the badge adds `· lowered` so the two never disagree
silently.
- Copy comes from `StepDownPlanner.Describe` (machine cause owns it, upload cause blames the upload);
the log gets the cause-tagged line for support.
Test: `EncoderRestartTests` runs a real pump with a fake encoder, changes the tier, and asserts a NEW
encoder started at the new fps/bitrate and produced a frame — the regression that would otherwise ship
as "it lowered the badge and the stream died".
Remaining for 33: shed our own capture cost before the fps step, and the session-report line.
Creator's Release take `ty-20261109-1142-036-0005.mp4` (1080×1920 @ 30, 300 s) measured by cropping the
webcam tile and diffing frames: 29.8 distinct images/s overall, median 30 s window 30.0/s, worst 24/s,
one 4-frame repeat run in 9026 frames. The app's own lines agree (`webcam 28.4–30.0/s` against camera
`arrivals 29.9–30.1/s`).
- The 25.6/s failure recorded on the 2026-10-05 16:9 take does **not** reproduce in the vertical tier,
so the vertical take no longer blocks the cadence units.
- Recorded the measurement trap in `frame-pipeline.md`: a threshold of 1.0 on a downscaled tile
reports long "0.0/s" stretches during low motion — duplicates sit below 0.02, a still face at
0.3–0.7. Read the histogram before believing a count.
- Vertical-tier camera check reported 1–6 pass; both creator-hands items in TASKS.md are closed.
(The mic being off for the first 2:42 does not affect video cadence.)
Slice 2. 588 green, 0 warnings.
- `FramePumpStats` and `FramePump.StatsReported`: the 5 s window the log line already describes, as
data. `Dropped` is the ENCODER's counter (frames submitted, not accepted); `Stalls`/`AvgRenderMs`
are the pump's. `FramesRendered` is deliberately not used for classification — a static scene
reuses the paste cache and renders few frames while completely healthy, so it would read as
starvation.
- `StepDownPlanner`: one per broadcast, owns the stateful policy, converts stats → sample, exposes the
landed tier for the restart. It never acts — the VM owns the effects, which is what keeps the ladder
testable without an encoder.
- `Describe` / `ReportLine`: the sentences a creator sees. Machine cause says LlamaCasty couldn't keep
up and the connection was not the problem; upload cause blames the upload and promises no further
thrash. A test asserts the machine copy never leaks "dropped" — mechanism strings stay in the log.
Remaining for the task: VM subscription, the encoder restart at the landed tier, the toast, and
shedding our own capture cost before stepping fps.
Pure policy, no wiring yet. 12 unit tests, 584 green overall, 0 warnings.
- **Machine starvation** buys less work: 60 → 30 fps first, then 1080 → 720. The vertical tier has no
resolution step — its floor *is* its only geometry.
- **Upload congestion** buys fewer bits: `current − max(0.5, current × lossFraction)`, floored at
3.0 Mbps (16:9) / 4.0 (vertical). Resolution is deliberately not a bandwidth lever: 1080p30 and
720p60 cost the same, so dropping pixels at an unchanged bitrate relieves nothing.
- **Budgets** as ruled: three steps for a 1920-wide start, one for 720 and vertical. 15 s of evidence,
30 s cooldown, ambiguous evidence does nothing at all.
- **Recovery** is `allowClimb`, default off (the one-way ruling): the machine carries its ceiling from
construction, climbs one step after 180 s of clean windows, and a climb followed by a relapse inside
120 s freezes the session at the low value — one flap pair is proof enough.
Inputs are the poster's own counters (`PumpStalls`, `PumpDrops`, `AvgRenderMs`, `EncoderDrops`), not a
throughput reading: `StreamHealth.CurrentBitrate` is ffmpeg's *encoded output* rate and sits at target
even while the socket is gagging, so it cannot size a step.
`scripts/session-stats.py` turns every `startup*.log` into one row per FramePump session: tier,
scene, measured fps, drops, stalls, average/worst render, capture rate, and the build — with the
sink marked, because that is what makes the numbers quotable:
- **RECORD** — machine only, no socket. The cleanest minimum-config evidence.
- **TEST** — machine + socket, no audience; recognised by `Test session dock open` /
`Test chat simulation`, which the creator pointed out are recorded too.
- **LIVE** — network and audience included.
- **UNKNOWN** — no marker; not evidence.
Verdicts use the average render against the frame budget, not the worst: the worst is dominated by
the 100 ms first-frame bake (a session that "holds" can still show worst 35 ms). `holds` = no drops,
no stalls, average under half the budget.
First run over the existing logs already separates the two halves of TASK 33: static Starting scenes
render at ~1.5 ms, the same pump with display capture running at 35–46 ms — the capture path, not the
overlays, is where a budget streamer loses the frame budget.
Design settled with the creator. Two findings drove it:
- **Resolution is not a bandwidth lever.** 1080p30 and 720p60 are both 6 Mbps in the shipped table,
so "step down to 720p" relieves nothing against congestion; and 60→30 at the same bitrate relieves
nothing either. Congestion therefore steps **bitrate only** (8.0 → 6.4 → 5.1 → 4.1, ~20%), which is
what OBS Dynamic Bitrate does — cited, adopted, with its climb-back discarded for v1.
- **The other half has no OBS equivalent**: machine starvation (the game and the app on one PC) steps
fps, then resolution, and sheds LlamaCasty's own capture cost *first*. That is the app taking
responsibility for a stream it is helping to starve.
Budgets as ruled: 3 steps for a 1080 start, 1 for 720 and vertical, hard floors at 720p30 /
Vertical 1080p30, hold-and-warn at the floor, never disconnect. Trigger ≥5% drops over 15 s twice,
30 s cooldown, ambiguous evidence does nothing.
New rule in `ui-surface.md`: a toast names the cause, the culprit and the remedy. Mechanism strings
("All pipe instances are busy", HTTP codes, exit codes) are log findings, never user-facing copy.
Audited the shipped feature surface against the list. Seven subjects were missing, three of them
distinguishers:
- **Green/blue screen removal** — auto-detect + opt-in suggestion, and it keys alert clips too.
- **Video file layers** — clip decodes, audio joins the mixer, per-layer mute/level.
- **Edit the broadcast on air** — title/description/tags/visibility, same fields prefill Go Live.
- Global hotkeys (F1–F5 / F6 / F7 / F8 / F9, OS-level), layout autosave, record with no sign-in.
- Resilience as one subject: ingest pre-flight, reconnect, Back on air, crash-safe fragmented MP4.
Corrections: the transition row claimed Move, which is still queued — now says so. The after-surge
list gained Move and scene-linked audio (35) and dropped entries that have shipped (media file
picking) or that belong to ready rows (resilience). 32 ready subjects, 8 later.
Creator went live on the Debug build while the Release build was recording and got an empty
title/description, with the Stream Info tab filled in on screen. Both processes write to the same
profile: `BroadcastForm` is loaded ONCE at startup and each keystroke persists straight to the store,
so the instance that opened the dialog was holding the form as it was when it launched — before the
other instance saved those fields. The dialog copied from that stale copy.
`CreateGoLiveDialog` (new, extracted so it can be tested without showing a window) calls
`BroadcastForm.Load()` before composing the dialog. That is what the form's own doc comment always
claimed ("startup / Go Live refresh") but the refresh was never wired. It cannot lose an edit: every
keystroke is already persisted.
The dialog also logs `Go Live dialog: title='…' description=N chars`, so the next report of "the
dialog was blank" says immediately whether the values reached the dialog or never left the store.
Test: `BroadcastPullOutTests` writes metadata through a second `LayoutStore` handle (the other
instance) after the VM has loaded and asserts the dialog carries it. 572 green, 0 warnings.
gramps