diff --git a/persistent-memory.d/2026-09-23-the-cache-that-aged-the-thing-it-cached.md b/persistent-memory.d/2026-09-23-the-cache-that-aged-the-thing-it-cached.md
new file mode 100644
index 0000000..7f1c713
--- /dev/null
+++ b/persistent-memory.d/2026-09-23-the-cache-that-aged-the-thing-it-cached.md
@@ -0,0 +1,56 @@
+# The cache that aged the thing it cached
+
+_2026-09-23 · booth_
+
+**Thumbnails: 77.5 MB → 0.78 MB on the biggest gallery, ~100 MB → 1.12 MB on
+the Desk.** The operator found the defect in about a minute of using the live
+redesign. Two lessons, and the second nearly shipped.
+
+## ⚠ LESSON 1 — we parked it on a count and the cost was in the bytes
+
+ROADMAP parked progressive loading on *"the largest gallery is 66 images; at
+that size a lazy grid is almost certainly fine"*, and the parking-lot row said
+*"270 `<img loading=lazy>` may be fine."* **Both count IMAGES. Neither weighs
+BYTES.** 66 is a fine count sitting on 77.5 MB of 1024×1024 PNGs rendered at
+250px — roughly 16× the pixels that reach the screen.
+
+The rule "measure the real booth before optimising it" was followed and still
+produced the wrong answer, because **we measured the dimension that was easy to
+measure rather than the one that determines the experience.** Before parking
+anything on a measurement again, ask what the user would actually feel.
+
+## ⚠ LESSON 2 — a cache INSIDE the thing it describes can age it, and excluding the cache is not enough
+
+`.thumbs/` lives inside the booth on purpose, so it is swept with the booth and
+can never outlive what it describes. But `_newest_mtime` — which feeds
+`is_expired`, which feeds `rmtree` — walks the booth, so **the server writing a
+cache on a mere view counted as the operator touching the booth.**
+
+**The first fix passed its own test and was still wrong.** Excluding every path
+under `THUMB_DIR` does not help, because **creating the directory touches the
+BOOTH DIRECTORY's own mtime**, and `_newest_mtime` SEEDS from exactly that. The
+cache's contents were excluded; its *existence* was the leak.
+
+The fix restores the booth's stamp across the `mkdir`. That cannot hide real
+activity: any file an agent adds is counted by its own mtime in the same walk,
+so the directory stamp is only a seed.
+
+**The blast radius if it had reached the Desk.** The Desk pulls a preview
+thumbnail per booth, so ONE INDEX LOAD would have pushed EVERY booth's expiry
+out — the TTL would never fire again and nothing would ever sweep. Caught by
+design-dev from the outside, hours before the strip landed; verified after the
+fix on the live set: **29 booths, two Desk loads, 22 caches generated, 0 clocks
+moved.**
+
+`.viewed` counting as activity is different and deliberate — that is a record of
+a *person* looking, which U4 says is activity. **A server-written derived cache
+is machinery, like the `.lock` sidecars already excluded.** Any future cache
+inherits this: exclude it from the clock AND preserve the parent's stamp.
+
+## A third, smaller one: the dot-namespace was only skin deep
+
+`booth_items` and `zip_booth` both tested `p.name.startswith(".")` — the FILE's
+name — so `.thumbs/a.png` (name `a.png`) would have rendered as a gallery item
+and shipped inside every zip download. CLAUDE.md invariant 2 promises a dotfile
+costs nothing in item counts, galleries or zips; **that was true only at the top
+level.** Both now skip every dot-prefixed path COMPONENT.
diff --git a/persistent-memory.md b/persistent-memory.md
index ea6a9ca..4e0b4a2 100644
--- a/persistent-memory.md
+++ b/persistent-memory.md
@@ -199,6 +199,7 @@ _As of 2026-09-22:_
 
 ## Recent decisions
 
+- `[2026-09-23]` ⚠ **The cache that aged the thing it cached** — thumbnails 77.5MB→0.78MB; READ BEFORE PARKING ANYTHING ON A MEASUREMENT (we counted images and the cost was in bytes), and BEFORE PUTTING A SERVER-WRITTEN CACHE INSIDE A BOOTH (excluding its contents does not stop it aging the booth) → `persistent-memory.d/2026-09-23-the-cache-that-aged-the-thing-it-cached.md`
 - `[2026-09-23]` ⚠ **The probe that nearly dismissed a live injection vector** — the link board rendered `javascript:` hrefs; READ BEFORE TRUSTING A NEGATIVE RESULT FROM AN OBVIOUS PROBE, and before assuming an existing scheme check is the guard you are looking for → `persistent-memory.d/2026-09-23-the-probe-that-nearly-dismissed-a-live-vector.md`
 - `[2026-09-23]` ✅ **The bug-hunt panel found six defects and five vacuous falsifiers** — READ BEFORE BUILDING ANY FRAGMENT ANCHOR (browsers match raw before decoded, so both sides must be encoded), and before trusting a well-commented diff's guards → `persistent-memory.d/2026-09-23-the-bug-hunt-panel-and-five-vacuous-falsifiers.md`
 - `[2026-09-22]` ✅ **v1.0.0b1 — the v1 target staged as a beta, and a version that was two copies** — READ BEFORE DERIVING A VERSION FROM `importlib.metadata` HERE; it reports a different artifact, and `booth/__init__.py` turns out to be stdlib-only → `persistent-memory.d/2026-09-22-v1-staged-as-a-beta-and-a-second-copy-of-the-version.md`