phase: 04_llm_chart_readability

**Phase 04 final verification pass — all tasks already in `complete/`; verified + fixed 3 defects found in verification.**

- Verified `/llm` live (both engines, gufo-only instance on :8099, `llm_monitor` on :8095) in a real browser; dev server left running/healthy.
- Completed task-04's truncated layout fix: `#engines` now carries `.items-start` (the CSS rule existed but nothing used it) — gufo 544 px vs vLLM 715 px, no blank card; pinned by `test_llm_engine_panels_are_not_stretched_to_their_tallest_sibling`.
- Fixed: event rows said the subject twice ("gufo gufo is answering again"), and `/ups` mangled mid-message names ("sent to apc: OK" → "sent to  OK") → shared `Dash.eventText()` (strips `UpstreamError: ` + a *leading* name only, mirroring `alerts._clean_error`) used by `/system`, `/llm`, `/ups`; new test.
- Fixed: count sparklines autoscaled rollup averages (queue max 0.1379) into full-height spikes → `sparkline({minSpan})`, `minSpan: 1` on sessions/queued (spike 30 px → 4 px); assertions added.
- Tests: `bash .agents/validate.sh` → **EXIT=0** (`uv run pytest --cov --cov-fail-under=94`: **485 passed, 7 skipped**, coverage **96.87%**; `ruff check src tests agent`: All checks passed; `pyright`: 0 errors; `node --check` all static JS clean; repo hygiene clean; `uv pip list --outdated` empty). `uv run pytest -m live --live` → **7 passed**.
- Completion criteria: heroes largest per engine ✔ (38.4 px vs 16.8 px row values); sparkline per cell + selected window ✔ (5/5 gufo, 7/7 vLLM; 6 h chip moved KV max 53 %→58 %); mean+p95 without hover ✔ (243 / 48.8 / 285 ms feet); provenance tag ✔ (`Δ10.0s`/`Δ8.0s` + `title`=exact source); single-engine ✔ (1 panel, totals hidden, zero red); screenshots ✔ `data/logs/llm-verify-{both-engines,charts-and-slots,single-engine,monitor-reference}.png` + `NOTES-chart-set.md`; pytest/cov/ruff/pyright ✔; earlier phases unchanged-behaviour ✔ (all 01–03 tests pass).
- Deviation (documented, not code): throughput heroes stay **neutral ink** — `severity.py` has no throughput row and 0 tok/s idle is healthy; colour comes from rated cells (KV) + reachability. Kept per `NOTES-chart-set`.
- Deviation: could not append the task-04 written verdict to `NOTES-chart-set.md` (it lives in `todo/`, which I must not edit). Verdict: ours reads better than `:8095` — bigger throughput numbers, per-cell shape over the selected window, percentiles inline; `llm_monitor` only sparklines decode+KV and has no window control.
- Docs: AGENTS.md updated (485 test count, `sparkline` `minSpan` rule, `Dash.eventText` event-row rule, `/llm` panel `items-start`); README reviewed in full — commands, config table (all `config.py` vars present), architecture bullets, page descriptions verified accurate; no other inconsistencies found.
- Note for the harness: `.agents/phases/todo/03…/NOTES-verify.md` and `todo/04…/NOTES-chart-set.md` are untracked notes left behind by their phases (task files moved, notes did not).
- Next pending phase: `05_ups_hero_battery_glyph`.

AGENTS.md

@@ -167,9 +167,10 @@ Responsibilities, one line each:
   `Dash.hero(el, {label, value, unit, digits, severity, sub, bar})` renders/patches a
   `.hero` big number (plus an optional `.bar`) from that class — it never compares a number
   to a threshold. `Dash.severityFor(value, warnAt, failAt)` is the documented fallback for
-  values the API does not rate; converting its remaining call sites is phases 04–05's job
-  (`/` was converted in phase 02, `/system` in phase 03 — those pages colour only from the
-  `severity` verdicts the API sends, and no `severityFor(` call survives in their scripts).
+  values the API does not rate; converting its remaining call sites is phase 05's job
+  (`/` was converted in phase 02, `/system` in phase 03 and `/llm` in phase 04 — those pages
+  colour only from the `severity` verdicts the API sends, and no `severityFor(` call survives
+  in their scripts).
 
 Metric naming: dotted (`cpu.usage_pct`, `mem.used_pct`, `disk.used_pct@/srv`,
 `gpu.busy_pct@0`, `llm.decode_tokens_per_second`, `ups.charge_pct`), `@stream` for per-device
@@ -215,7 +216,7 @@ cp .env.dev.example .env.dev              # dev-only: ssh tunnels + live targets
 ./scripts/dev-server.sh status            # per-source ok/DOWN
 ./scripts/dev-server.sh logs
 
-uv run pytest                             # hermetic: 472 passed + 7 live-skipped, no network
+uv run pytest                             # hermetic: 485 passed + 7 live-skipped, no network
 uv run pytest --cov                       # coverage gate (96.9%; floor 94%)
 uv run pytest -m live --live              # opt-in: real fleet / engines / upsd
 uv run ruff check src tests agent && uv run pyright
@@ -302,6 +303,30 @@ muted unit, mono `.sub` footnote (`Δ10.0s`, "at current load"), coloured by the
 family and no second palette, and `tests/test_api.py` fails if either appears or if a CSS
 rule reaches for a token that `:root` does not declare.
 
+A cell-scale trend is `DashCharts.sparkline()` — the `llm_monitor` `drawSpark` idiom, so a
+percent metric is plotted against `fixedMax` (**never** autoscaled: an 8 % KV window that once
+held a 92 % spike must not read flat, and the axis gets top/bottom scale marks so "a third of
+the way up" says a third of what). `align: 'right'` pins the newest sample right and grows the
+line leftwards inside a real `span` in seconds — x stays time, not sample index. Stroke colour
+comes from the payload `severity` via the `--ok`/`--warn`/`--fail` tokens only; an unrated
+series draws a faint `--text-faint` stroke with no fill, because a shape is not a verdict.
+Below `minPoints` (2) it prints `no history` rather than drawing a confident line through one
+point, `labels: {current, max}` renders the `now / max` foot through the caller's `format`, and
+there is no animation at all (`prefers-reduced-motion` holds by construction). `minSpan` is the
+other half of the same honesty for count metrics: history arrives as rollups, so eight seconds
+of one queued request averages to 0.14 and an axis fitted to that draws a full-height spike for
+a queue nobody was in — `/llm` passes `minSpan: 1` on sessions and queued so its axis is never
+finer than one request.
+
+An event row names its subject **once**. Every feed (`/system`, `/llm`, `/ups`) renders the
+message through `Dash.eventText(event)`, which drops the `UpstreamError: ` wrapper and a
+*leading* duplicated subject the way `alerts._clean_error()` does for the overview — `poller.py`
+messages start with their own label, and each row already prints that label as a chip, so
+"gufo gufo is answering again" reads as two problems. A name in the middle of a message stays:
+the UPS rows used to `replace()` it anywhere and mangled "Quick battery test sent to apc: OK".
+This is presentation only — the stored event and `/api/events` keep every word, and the raw
+error string still surfaces in `/api/health`.
+
 The `/system` host card is one row of six heroes — CPU, MEM, TEMP, DISK, GPU, NET (the
 `.heroes` grid; `tests/test_api.py` pins the slots and their order on the served script).
 Colour comes from the payload's `severity` dicts and nowhere else; a missing number or
@@ -316,7 +341,29 @@ simpler surface, not a less complete page.
 A `Δ<window>` label is a measurement, never a config value: `sources/host.py` sends the window
 the estimator actually used as the payload's `window_seconds`, and `/system`'s footnote prints
 those (a range when the hosts disagree) instead of the poll interval, which only coincides with
-them at the defaults.
+them at the defaults. `/llm`'s hero `.sub` prints the rate cell's own `window` through the same
+`provenance()` tag, and the hero's `title` carries the exact `source` string — "78 tok/s" must
+always say whether it was scraped, measured `Δcounter/Δt`, or estimated (`llm_monitor`'s
+`setThroughput` idiom; dropping the tag to look cleaner is a regression, not a simplification).
+
+**The `/llm` page reads per engine** (phase 04, `NOTES-chart-set.md`): a row of two
+`Dash.hero()` throughput numbers (decode, prompt — deliberately **uncoloured**, because 0 tok/s
+on an idle server is healthy and `severity.py` has no throughput row), then a compact `.cells`
+list — label left, tabular value right, a `sparkline()` fixed at 160px so rows line up, and a
+mono `p95 …` foot under every latency mean (no hover; p95 needs no stored series). Rows are
+engine-scoped by the `only` flag: gufo never grows a TTFT mystery, vLLM never a spec-accept
+hole. `preemptions` / `device_lost` are **alarm chips that exist only while non-zero** (`.chip.fail`,
+container emptied and `hidden` at zero — the rule is the presence of an incident, never a graded
+threshold); lifetime token counters are a quiet footnote line. **One `/api/llm/history` request
+per refresh** carries every chart *and* spark series (`HISTORY_SERIES`; the endpoint answers
+unknown names as empty series on purpose), and the two series that addition required
+(`llm.tpot_seconds` vLLM-only, `llm.spec_accept_rate` gufo-only) are stored additively in
+`sources/llm.py` HISTORY_METRICS. The engine-totals card kept only engines-live/active/queued
+and **hides below two engines** — a cross-engine sum of one engine duplicates that engine's own
+panel, and the summed-KV bar went because `fleet_totals` carries no `severity` to render. The
+four trend charts stay but at height 110 *below* the panels, each with an `ariaLabel`. Panels
+sit at their own height (`#engines` carries `.items-start`): a 5-row gufo panel next to a 7-row
+vLLM one otherwise grows a blank bottom, and a blank panel reads as "no data".
 
 **Charts belong to the host they chart.** `/system` has two sections — the host panels and the
 transition events — and nothing else: the fleet-wide chart stack, the nine-figure fleet-totals