diff --git a/docs/telemetry-contract.md b/docs/telemetry-contract.md
index 3c1b21d..ff285d8 100644
--- a/docs/telemetry-contract.md
+++ b/docs/telemetry-contract.md
@@ -22,6 +22,7 @@ older than 14 days nightly.
 | `player`  | `exo_error`         | `{host, anime}` — native player error            |
 | `player`  | `av_delay_applied`  | `{audioMs, host}` — user dialed in an audio offset (the "a track falls behind" signal) |
 | `session` | `session_summary`   | one row per playback — see below. **Written only by `SessionFeedbackService._write`.** |
+| `error`   | `session_hold_failed` | `{op, sessionId}` — an open session could not be held on disk; see [below](#a-session-the-app-never-saw-end-ended-unseen) |
 | `qa`      | `qa_session`        | `{sid, lines[], truncated?}` — the QA decision log for one playback (issue #92), off by default |
 | `source`  | `source_select`     | one row per source sheet — timings and counts, see below |
 | `source`  | `source_pick`       | **what was chosen and what it beat** — see below |
@@ -64,8 +65,10 @@ The pass is arithmetic over a few dozen candidates, once per pick.
 
 `{sessionId, title, probeKey, host, durationS, watchedS, positionS, stalls,
 droppedFrames, renderedFrames, maxConsecutiveDropped, problemAtS[], promptOutcome,
-rating?}` plus the player's provenance fields (`forcedAudio`, `sideloadedSubs`,
-`resumed`, `seeks`, `seekStormMax`, codec/height/bitrate).
+rating?, ended, lastSeenAt?, playedOn?}` plus the player's provenance fields
+(`forcedAudio`, `sideloadedSubs`, `resumed`, `seeks`, `seekStormMax`,
+codec/height/bitrate). `ended` and the two after it are explained in
+[A session the app never saw end](#a-session-the-app-never-saw-end-ended-unseen).
 
 - **`watchedS` is elapsed watch-clock time. `positionS` is where playback got to.**
   They differ whenever someone seeks, rewatches, or stares at a spinner — a session
@@ -81,6 +84,46 @@ this collection computed before that date counts Android sessions twice**, and
 rows older than the fix still carry the duplicate. `sessionId:isset = true` is the
 filter that excludes them.
 
+### A session the app never saw end (`ended: unseen`)
+
+Added 2026-09-13 (amber-app, first in the build after 1.3.0+38). Until then the
+summary and the amber-api health row were written only when the player closed, and
+a box whose television is switched off with the film on screen never closes it.
+Measured on the kids' box `Xiaomi MIBOX4 · ff3bcf`: **13 of 13 films in 14 days
+left no summary and no health row.** At the end of every one the 5 s position
+save ran at full rate up to the last 45 s `api_usage` tick and then stopped, and
+no row from the hide/pause flush arrived, then or at any later launch (0 of 3572
+`api_usage` rows were delivered late). That box gives the app no callback, so a
+write on `AppLifecycleState.paused` would have recovered none of them.
+
+Now both players hold the open session on disk every 10 s
+(`<appSupport>/open_sessions/<sessionId>.json`). A normal close writes as before
+and deletes the copy once the summary is in the telemetry spool. The next launch
+writes whatever is still held, through the same single writer, with:
+
+| field | meaning |
+|---|---|
+| `ended` | `closed`: the player closed and wrote it. `unseen`: a later launch wrote it from the held copy. Absent on older builds. |
+| `lastSeenAt` | `unseen` only. When the session was last held, device clock, UTC. The row's own `ts` and `created` are the later launch, often the next day. |
+| `playedOn` | `unseen` only, and only when it differs: the `appVersion gitSha` that played it, because a row is stamped with the build that sends it. |
+
+Reading an `unseen` row:
+
+- The counters stop at `lastSeenAt`, at most 10 s before the end, and
+  `promptOutcome` is `notShown`.
+- It does not say how the session ended (TV switched off, power cut, crash, killed
+  in the background), only that the app did not see it.
+- It arrives when the box is next started. A complaint made before that start has
+  no row yet.
+- Its amber-api health row is sent at the same moment, once. A session that was
+  held and then closed normally sends one health row, at the close.
+- A session killed within its first 10 s was never held and still leaves nothing.
+
+`kind: error`, `event: session_hold_failed`, `meta: {op, sessionId}` records a
+hold that could not be written (`put`), let go of (`release`), read back
+(`unreadable`, the copy is deleted) or listed (`recover`). One row per session per
+op, so a full disk is one row, not one per tick.
+
 ## How Claude queries it
 Superuser token (same as releases publishing), then filter/sort the collection: