diff --git a/HANDOFF.md b/HANDOFF.md
index 9664884..b66a13b 100644
--- a/HANDOFF.md
+++ b/HANDOFF.md
@@ -1,40 +1,40 @@
-# HANDOFF — 2026-11-12 (bitrate correction: all four units landed)
+# HANDOFF — 2026-11-12 (bitrate + audio-clock fixes; sync number pending)
 
 ## State
 
-- `main`; `b265d42` is the latest pushed commit. Everything after it is local.
-- **Bitrate saga closed.** An @8 stream averaged 5.1 Mbps because the args were capped VBR, not CBR:
-  measured on the bundled build, a static frame at `-b:v 2000k` gave 9 kbps with NVENC's default and
-  1951 kbps with `-rc cbr`. Not the machine (1.8 ms render, 0 stalls) and not YouTube (no step-down
-  logged). `700c7a8` fixes it: `-rc cbr` for NVENC/AMF, nothing for QSV (it infers CBR from
-  bitrate == maxrate) and nothing for OpenH264 (no measurable difference).
-- `1e3a4f7` pre-release switch (`OutputProfile.EntitlementEnforced = false`) lets a Release build select
-  60 fps, and suppresses the licence row while nothing is limited. Flip at launch.
-- `1708667` ladder retargeted to 1080p60@12 / 1080p30@10 / 720p@6; `93719a2` makes the step-down floor
-  the tier's YouTube minimum (6/5/3) instead of a flat 3.
-- `f2cab5c` the creator sets the rate (per-tier Mbps box, persisted, clamped only at the YouTube minimum);
-  `b0f8263` files each session locally and proposes a rate between sessions on consent alone.
-- Verification: **672 tests green, clean build with 0 warnings** (`./scripts/verify.sh`).
+- `main`; `b265d42` is the latest pushed commit. Everything after is local.
+- **Bitrate saga closed.** An @8 stream averaged 5.1 Mbps because the args were capped VBR, not CBR
+  (measured: 9 kbps vs 1951 kbps at a 2000k target on a static frame). `700c7a8` adds `-rc cbr` for
+  NVENC/AMF, nothing for QSV (infers CBR from bitrate == maxrate) or OpenH264 (no measurable gain).
+- Ladder `1708667` (12/10/6); floor `93719a2` (tier YouTube minimum); rate box `f2cab5c`; rate history +
+  consent proposal `b0f8263`; pre-release 60 fps switch `1e3a4f7`.
+- **Audio drift closed** (`47f409f`, `102cb96`, `751d9b1`): the mix is driven by the monotonic clock,
+  counts only audio whose write reached the pipe, and starts its clock at session open. Measured
+  across three recordings: +2.1% long → 226 ms short → 137 ms short → **lengths now match (34 ms in
+  13 s)**.
+- **What is left is one number.** `5ae5280` gives the mic delay a 0..500 slider (off/175/350) and logs
+  the applied value. Measured on the last clip: `-185 ms`, i.e. sound still 185 ms EARLY, so the value
+  to try is **350–360 ms**. `Audio.MicSyncOffsetMs` is stored per profile.
+- **675 tests green, 0 warnings.** `scripts/avsync.py` measures any recording in seconds.
 
 ## Next action
 
-- **Creator check pending in a real Release stream**: 60 fps selectable, live bitrate sitting near the
-  rated number instead of drifting, the Mbps box and the rate prompt behaving.
-- Resume the queue: **40·A/C/D → Move → 9·5 → 14·16/17/18 → 66 → 67 → 68 → 69 → 70 → 36**.
-- Ask creator before TASK 68 (first-run ON-AIR default); settle the Release single-instance mutex
-  before the gold pass. Do not push absent a checkpoint request.
+- Creator: set the mic delay to 350, record ~15 s, run `./scripts/avsync.py "<clip>" --crop
+  764x764+1032+159`. Target: `offset` inside ±30 ms and `lengths match`.
+- Then resume the queue: **40·A/C/D → Move → 9·5 → 14·16/17/18 → 66 → 67 → 68 → 69 → 70 → 36**.
+- Do not push app commits absent a checkpoint request.
 
 ## Landmines
 
-- `Assets/starting-background.jpg` is modified in the working tree, **awaiting the creator's visual check** —
-  do not commit it with another unit.
-- **Two load-sensitive flakes, not regressions**: `ChromaKeyRealClipTests.RealClip_PlaysAtItsNativeRate`
-  and `AudioPipelineTests.Mix_WithFiltersDuckAndGain_Lands_On_AudioPipe` failed once each in a full run
-  and passed alone; verify.sh only tolerates its known pair, so rerun the single test first.
-- `RatedFor` must not cache before `_layoutStore` exists (the ctor resolves the default tier first), or
-  the shipped default shadows the creator's stored rate for the whole run.
-- Every rate change is a full encoder relaunch (`ChangeEncoderAsync`), so mid-show rate changes stay out.
-- A tier's rate is read through `MainViewModel.EffectiveQuality`; a call site using the raw tier
-  streams the shipped default.
-- `MainWindow` tests need a temporary `LayoutPathOverride`, cleared in `finally`; build with the Windows
-  `dotnet.exe` host from WSL and read render numbers in Release.
+- `Assets/starting-background.jpg` is modified in the working tree, **awaiting the creator's visual
+  check** — do not commit it with another unit.
+- **Never read a field the constructor assigns later** (`RefreshRateProposal()` at ctor 354 reads
+  `_framePump`, built at 468 — 8 failing tests, `e654a16`).
+- **Never pipe a long job through `tail`** — it buffers, so a live gate looks dead, and verify's phase-3
+  build writes to `/dev/null`. Use `vstest --logger:"console;verbosity=normal"`; it names the failure.
+- A killed test run leaves `testhost.exe` holding `LlamaCasty.dll`; every later build then fails on the
+  copy. `taskkill.exe /F /IM testhost.exe` before rebuilding.
+- `ChromaKeyRealClipTests`, `AudioPipelineTests.Mix_*`, `AppLogRotationTests` and
+  `NotificationAreaIntegrationTests` flake under load; each passes alone.
+- Every rate change is a full encoder relaunch; rates go through `MainViewModel.EffectiveQuality`.
+- `MainWindow` tests need a temporary `LayoutPathOverride` cleared in `finally`; read render numbers in Release.