diff --git a/docs/perf/NEXT-STEPS.md b/docs/perf/NEXT-STEPS.md
new file mode 100644
index 000000000..e0e509b5c
--- /dev/null
+++ b/docs/perf/NEXT-STEPS.md
@@ -0,0 +1,92 @@
+# Where this left off, and what to do next
+
+Written as a handoff. The branch is green and everything below is measured except where explicitly
+marked.
+
+## 1. Verify the one unmeasured change
+
+`c526dd2b7` applies the `content-visibility` treatment to `mediaRequests-requestList`. It is
+typechecked but **not measured** — the session ended first. Every other performance claim on this
+branch was A/B'd before commit; this one was not.
+
+```bash
+# restore two containers on identical data, one per image, then:
+STRESS_BASE_URL=http://127.0.0.1:PORT STRESS_BOARD=default \
+  node --experimental-strip-types scripts/benchmarks/measure-board-paint.mts
+```
+
+Expect **image requests to fall** (the measured effect on `mediaReleases` was 135 → 55). Do **not**
+expect JS heap to move: decoded bitmaps live outside `JSHeapUsedSize`. Three repeats per side — a
+single run wrongly suggested a −16% heap win on the first widget.
+
+If it does not reduce image requests, revert it. Unmeasured complexity is exactly what this branch
+has rejected everywhere else.
+
+## 2. The one lever with real headroom left
+
+**The tRPC router imports every integration eagerly.** A single **16.10 MiB subtree of 460 files** is
+reached by every API route, and because Next preloads route bundles it is resident before the first
+request. This pays twice:
+
+- ~41 MiB of retained bundle source (28% of the live server heap) — V8 keeps a script's source alive
+  while any function in it may still be lazily compiled
+- ~38 MiB of V8 compiler zone memory, which is **proportional to how much JavaScript V8 compiles**
+  and is not reachable by any V8 flag (`--max-opt` is a null result; only `--jitless` removes it, and
+  that breaks rendering)
+
+Nothing else has this profile. Every allocator and V8-level avenue is closed — see
+[REJECTED.md](./REJECTED.md).
+
+Start from `analyze-require-graph.mjs`, which showed the shared subtree, and
+`[root-of-the-server]__0g06wx7._.js` as the chunk that reaches 460 files.
+
+## 3. Known-good but deliberately deferred
+
+| item | prize | why it was not done |
+| --- | --- | --- |
+| ASCII-escape server bundles | **9.45 MiB** | Proven mechanism (2.00 vs 1.00 bytes/char; one char above U+00FF doubles a whole source string). Needs a build-time rewrite of minified **vendor** output — deserves its own PR with its own safety argument |
+| Lazy `mysql2` / `pg` | 21.4 MiB | Three implementations, all broken; see REJECTED.md. Blocked structurally because `createDb` must stay synchronous |
+| `beszelSystemStats` at 1,156 nodes | unknown | Second-heaviest widget, but its cost is chart SVG (`<g>×317`), so reducing it changes what is drawn |
+| `documents` reads 18–32 | unknown | Never investigated; each document carries overhead |
+
+## 4. For the clean PR
+
+The instruction was to close this experimental PR and open one with only the verified fixes, judged
+on lines-of-code to performance ratio. On the evidence, that set is:
+
+| change | size | measured effect |
+| --- | --- | --- |
+| `--max-semi-space-size=4` | **1 line** | −19% idle, −21% peak |
+| `preloadEntriesOnStart: false` | **1 line** | −52% anon at idle; 310 fewer files even at full load |
+| Umami: stop fetching 100k raw events per refresh | 1 file | −26% peak, −93% transient buffers, largest allocation 46.4 → 2.4 MiB |
+| `serverExternalPackages` (mysql2/pg + 8 more) | 1 block | −21 MiB idle |
+| `content-visibility` on media-release cards | 4 lines | −59% image requests, −41% layout objects |
+| ssh2 WASM heap (docker-modem patch) | 1 patch | −15 to −18 MiB external at every stage |
+
+Together these produce the headline: **−26% peak anon, −52% at idle**, no overlapping ranges at any
+stage, for **+18% CPU**.
+
+Two things to leave out or flag:
+
+- **The gRPC patch** removes 112 files and 6 MiB from an isolated `require`, but is **not separately
+  visible in a full run**. It is bundle hygiene, not a measured memory win.
+- **Median board load reads +30%** but is **not established** — individual loads were
+  [506, 383, 1233] and [848, 431, 574] before against [780, 808, 515] and [621, 579, 1004] after,
+  heavily overlapping with no cold-then-warm pattern. If the latency trade matters, measure it
+  properly with many more samples; the suspicion is that it is real and concentrated in
+  `preloadEntriesOnStart`.
+
+## 5. Read this before measuring anything
+
+[METHODOLOGY.md](./METHODOLOGY.md). Every wrong number in this investigation came from an instrument,
+not the application, and several were published before being caught. The short version:
+
+- a heap snapshot **cannot** explain a peak — V8 GCs first and reports survivors only
+- V8 truncates snapshot string contents to ~1 KiB
+- snapshot objects are **not** live native handles
+- measuring without a forced GC measures V8's growth, not a module's cost
+- taking a CDP heap snapshot added **134 MiB to the process being measured**
+- React per-component attribution silently returns **nothing** on a production build
+- a fixed port can measure a **leftover container**
+- **when a server-side change moves a client-side metric, the workload changed** — this is how
+  `--jitless` was caught faking a 24% win by under-rendering the page
diff --git a/docs/perf/README.md b/docs/perf/README.md
index 3e28d0bc8..b19272c41 100644
--- a/docs/perf/README.md
+++ b/docs/perf/README.md
@@ -10,6 +10,7 @@ widgets, 11 integrations, driven through an actual login — not from a syntheti
 | [CLIENT-PERFORMANCE.md](./CLIENT-PERFORMANCE.md) | Dashboard rendering cost in Chrome: DOM, fibers, images, paint |
 | [REJECTED.md](./REJECTED.md) | Ideas that were implemented and measured, then abandoned — with the numbers |
 | [METHODOLOGY.md](./METHODOLOGY.md) | The tools, and the measurement traps that produced wrong answers |
+| [NEXT-STEPS.md](./NEXT-STEPS.md) | Handoff: what is unverified, what to do next, and the set for a clean PR |
 
 ## Headline result