migration: mark executed 2026-08-01, record deviations from the written procedure

migration/README.md

@@ -1,5 +1,35 @@
 # Plex Migration — Windows → Linux, nothing lost
 
+> ## ✅ EXECUTED 2026-08-01 — M0 through M4 complete in one day
+>
+> Plex is live on Linux with its original identity (`eb69ec8a…`), watch state
+> intact, direct play verified with a real session. In soak (M5). This file is
+> now the **record of the procedure**; deviations that a future re-run must
+> know about:
+>
+> 1. **VACUUM is mandatory after the remap.** The no-backslash gate greps the
+>    raw DB file and hits deleted rows lingering on SQLite **free pages**.
+>    Plain sqlite3/python cannot open the DB (custom `icu_root` collation —
+>    do NOT fake it); run VACUUM with **Plex's own SQLite** via
+>    `docker run --entrypoint`.
+> 2. **Expect phantom rows from dead eras.** 1,715 backslash rows survived
+>    the remap legitimately: Sync+ cache from old D:/E:/G: drive letters
+>    (+ orphan `media_items`), 6 rows for a dead `\\tesla` NAS, and 115 UNC
+>    blobs in `download_queue_items.decision_result`. All deleted; 3 real
+>    theme-music paths re-pathed. The remap itself was correct — these
+>    predate the current library roots.
+> 3. **Verify staged copies with `rsync -an --delete --stats`** (0 transferred
+>    / 0 deleted), never `du -sb` equality — directory apparent sizes differ
+>    across filesystems.
+> 4. **First boot needs `systemctl start --no-block`** in cloud-init runcmd —
+>    `enable` alone misses the boot it runs in (fixed in `user-data`).
+> 5. Post-go-live, the Plex image pin was **reversed** (Mike, 2026-08-02):
+>    `:latest` + watchtower daily 05:00 with ntfy. Watchtower needs
+>    `DOCKER_API_VERSION=1.44` against Engine 29.
+>
+> Cleanup 2026-08-02: staging dir + `.pre-remap`/`.pre-syncfix` sidecars
+> deleted. Rollback net = NAS `media_pc` backup + Plex's scheduled DB backups.
+
 Supersedes the earlier "forward-only, fresh Plex" decision. That call was made
 before the library was measured. At **27 TB, 25,148 video items and 337 GB of
 already-generated preview cache**, rebuilding fresh costs one to three weeks of
@@ -200,6 +230,11 @@ paths remain, that the counts moved 1:1, that the "Dr. Korval" tag and the
 four Liaden book summaries are **untouched**, and finally stats 200 random
 remapped paths against the real filesystem.
 
+**As executed:** the no-backslash gate additionally required deleting phantom
+Sync+ cache rows from dead drive-letter eras, 6 dead `\\tesla` rows and 115
+`download_queue_items` UNC blobs, then **VACUUM with Plex's own SQLite**
+(see the banner at the top) before the raw-file grep came back clean.
+
 No claim token. The server already has its identity. The compose bind-mounts
 `/srv/data/plex-media` over `.../Plex Media Server/Media`, so the migrated
 thumbnails are in place from the first start.
@@ -236,5 +271,6 @@ decision, not a cleanup step.
 Before M3 step 3 (the format), `D:` is intact and everything is reversible on
 the data side — Windows itself is gone after M1, but a reinstall from the USB
 takes ~20 minutes and every copy still exists. After the format, the NAS
-`media_pc` backup is the recovery source: it outlives every step here, and
-`migrate-db.sh` leaves a `.pre-remap` copy beside every database it touches.
+`media_pc` backup is the recovery source: it outlives every step here.
+(The on-box `.pre-remap`/`.pre-syncfix` sidecars were deleted 2026-08-02
+after the soak-start all-clear.)