Docs: the 2026-10-17 measurements (largest paints, the channel page's blocking time, real-user data) and how chat history renders in pieces (#219)

ddiff --git a/docs/chat-system.md b/docs/chat-system.md
iindex 8226ca67..738c4085 100644
---- a/docs/chat-system.md
++++ b/docs/chat-system.md
@@@ -240,6 +240,9 @@ unless the viewer chose otherwise; the rules come from `GET /api/chat/filters/fr
  ## History and rendering
  
  `buildChatMessageEl()` in `public/js/chat.js` renders every surface (page, popout, broadcast desk,
--floating widget). History renders off-screen and swaps in once; the last 200 rows of a room are
--cached in localStorage per user for an instant paint on reload. The view follows new messages
--unless the reader scrolled back (`_watchChatPin`).
++floating widget). History renders off-screen: the newest 50 rows are swapped in at once, then the older
++scrollback (up to 500 channel + 500 global rows) is put above them 10 rows at a time, as many pieces as fit
++in each idle moment (`_historyPieces`), keeping the reader where they are. All of it in one task was a
++2.7 s freeze on a phone-speed CPU (2026-10-17). The last 200 rows of a room are cached in localStorage per
++user for an instant paint on reload (its newest 50 render). The view follows new messages unless the
++reader scrolled back (`_watchChatPin`).
ddiff --git a/docs/performance-audit.md b/docs/performance-audit.md
iindex 0489c896..05b94928 100644
---- a/docs/performance-audit.md
++++ b/docs/performance-audit.md
@@@ -8,9 +8,10 @@ says so.
  
  | Tool | What it does |
  |---|---|
--| `scripts/perf/measure-page.js <url> [--runs N] [--phone] [--phone-cpu]` | cold load in headless Chrome over the DevTools protocol, fresh profile and HTTP cache disabled per run; reports medians of transfer, requests, DOM nodes, running animations, FCP/LCP/CLS, long tasks, TBT, script/style/layout time |
++| `scripts/perf/measure-page.js <url> [--runs N] [--phone] [--phone-cpu]` | cold load in headless Chrome over the DevTools protocol, fresh profile and HTTP cache disabled per run; reports medians of transfer, requests, DOM nodes, running animations, FCP/LCP/CLS, long tasks, TBT, script/style/layout time, the LCP element (`lcpEl`) and the element behind the largest layout shift (`clsEl`) |
++| real users (`public/js/ov-rum.js` → `server/web/rum.js`) | one page load in four reports its own TTFB/FCP/LCP/CLS/INP/blocking time and script errors; aggregates per page kind × device class at `GET /api/admin/rum` (staff) and `live_rum_p75` / `live_rum_errors` on `/metrics` |
  | `scripts/perf/ab.js <urlA> <urlB> --runs N` | the same, **interleaved** A,B,A,B… so machine load affects both sides equally |
--| `scripts/perf/trace-page.js <url>` | Chrome trace; attributes style/layout work to the JavaScript that forced it |
++| `scripts/perf/trace-page.js <url> [--tasks]` | Chrome trace; attributes style/layout work to the JavaScript that forced it; `--tasks` lists each main-thread task over 100 ms with what ran inside it |
  | `test/browser/smoke.js` | routes × widths, errors, overflow, navigation, duplicate scripts, resource growth over repeated navigation |
  
  Profiles: **desktop** 1366×900 unthrottled; **phone** 412×900, 4× CPU, 1.6 Mbps/150 ms;
@@@ -178,6 +179,25 @@ Same A/B harness (phone profile, worst of two runs): channel 1.31 → 0.012, cli
  0.127 → 0.011; at 1366 px channel 0.09 → 0.015, VOD 0.19 → 0.001, home 0.06 → 0.01. First paint unchanged.
  `test/first-paint-stability.test.js` holds the pieces in place.
  
++## Measurement 2026-10-17: largest paints, the channel page's blocking time, real users
++
++Phone CPU profile (4× CPU, network unthrottled), production, median of 3:
++
++| Page | Before | After | What it was |
++|---|---|---|---|
++| home: LCP | 2.94 s | 1.15 s | the hero's AI-moment background (`div.hero-bg-layer`) was set by app-home.js after `/api/home/hero`; the server now preloads it and paints it on the first layer (#206) |
++| `/@goosely` (offline): TBT | 3.7 s | 1.2 s | chat.js re-rendered the whole scrollback (500 channel + 500 global rows) in one idle callback, 2.7 s; now the newest 50 rows, then 10-row pieces per idle deadline (#207, #209, #210) |
++| `/@goosely`: LCP | 3.3 s | 2.45 s | the offline screen's "#1 VOD" picture was lazy and waited for `/popular` after the channel request; now eager, high priority, fetched with the channel (#216) |
++| openvibe.video watch page: LCP | 3.8 s | 0.88 s | the `<video>` had no poster, so the largest paint waited for the first frame (MediaHub#94) |
++
++Server side (host, warm): `/api/streams/channel/:username` 120–150 → 95–120 ms (clipper names and AI overviews in one
++query each instead of one per row, #213); `/api/vods/:id` and `/api/clips/:id` now run their six lookups in parallel
++(23–40 ms, #214). Other OpenVibe sites' homes and inner pages: FCP/LCP 0.5–0.9 s, TBT under 210 ms, no overflow at 320 px.
++
++The throttled-network `--phone` profile still shows the home's first paint at 3.3 s (six render-blocking stylesheets
++share the 1.6 Mbps link with ~30 deferred scripts). Real-user data (#217) is the arbiter now: compare
++`live_rum_p75{page="home",device="phone"}` before restructuring the stylesheets.
++
  ## Budgets
  
  `npm run perf:budget` (scripts/perf/check-budgets.js) fails if the home page's raw HTML, eagerly