diff --git a/CLAUDE.md b/CLAUDE.md
index 2aed479..aad376b 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -383,6 +383,71 @@ Non-goals for this phase (same discipline as every phase so far):
   someone with database superuser access; that's an operational control,
   out of scope here and named explicitly, not silently assumed away.
 
+## What "done" looks like for Phase 5
+
+**Status: shipped.** A ground-up frontend redesign — visual direction,
+a real design system, navigation/IA, charting, dashboard panels,
+query/search, and alerting UI — plus an accessibility pass, all verified
+against a live docker-compose stack with real seeded data, not just
+`npm run check`/`npm run build` passing. See `/docs/design-system.md`
+for the token system and component library, and
+`/docs/phase-5-runbook.md` for the full verification log, including five
+real bugs this phase's live-verification discipline caught that a
+type-checked, successfully-building frontend would not have surfaced on
+its own.
+
+The visual direction ("Signal": near-neutral grayscale UI, color rationed
+to the four-tier severity system plus a single interactive accent, real
+dark-mode-as-default) was picked from three proposed directions before
+any token or component work started, per an explicit stop point in this
+phase's brief. The charting library (ECharts, over Observable Plot and
+raw D3 — see the design-system doc for the reasoning and the verified
+bundle-size/perf numbers) was likewise confirmed before being wired into
+every panel type, the second explicit stop point.
+
+Two of the five bugs this phase's verification caught were backend bugs
+with no connection to the frontend redesign itself, only surfaced
+because getting real dashboard/alert data to verify the new UI against
+required actually exercising write paths nothing had exercised since
+Phase 4's `tenant_id` migrations landed:
+
+- `alerting`'s `rulestore.Create`/`ApplyTransition` never populated the
+  `tenant_id` column Phase 4 added to `alert_state`/`delivery_log` (with
+  a `NOT NULL` constraint) — every alert rule created against a
+  Phase-4-or-later database silently failed. Existing rows all had a
+  value from Phase 4's backfill migration, which is exactly why this
+  went uncaught: Phase 4's own verification never created a *new* rule
+  post-migration, and its runbook already discloses that Docker access
+  was lost partway through that phase.
+- `dashboard_panels`'s `viz_type` CHECK constraint was never updated
+  alongside `heatmap`'s addition to the Go/TS validators — a three-place
+  change (Go validator, TS union, DB constraint), not two.
+
+Both are fixed (`alerting/internal/rulestore/store.go`,
+`metadata/migrations/0035_add_heatmap_viz_type.sql`) and confirmed
+against a live stack: rule creation → evaluation → firing → a real
+(failed, to an intentionally fake webhook) delivery attempt, and a
+heatmap panel created, persisted, and rendered end to end. See the
+runbook for the other three findings (one more real product bug — a
+`findIndex`/nullish-coalescing bug in the chart-pivoting logic that made
+every `single_stat` panel render `0` — and two real accessibility
+findings caught by axe-core against live-rendered pages with real data,
+not fixture data or empty states).
+
+Non-goals for this phase (same discipline as every phase so far):
+- No query-language or data-model changes beyond the one narrowly
+  justified exception: `heatmap` as a `VizType`, needed to feed a new
+  visualization, not a new query capability.
+- No changes to tenant isolation, RBAC, SSO, or audit logging — Phase 4's
+  surface area is untouched; this phase is presentation-layer only.
+- No mobile-phone-width layout — responsive verification stops at
+  tablet-landscape width, per the brief's explicit scope ("laptop/
+  tablet-landscape," not phone-width).
+- No fuzzy search in the command palette, no data-grid virtualization for
+  very large result sets, no chart types beyond the five built
+  (time-series, bar, single-stat, heatmap, top-N) — real, disclosed
+  future work, not oversights.
+
 ## When in doubt
 Ask before: changing the pinned stack, adding a new external dependency
 that pulls in a large transitive tree, or making an architectural decision
diff --git a/docs/design-system.md b/docs/design-system.md
new file mode 100644
index 0000000..a5358cc
--- /dev/null
+++ b/docs/design-system.md
@@ -0,0 +1,371 @@
+# Sentry Design System (Phase 5)
+
+Direction: **Signal** — cockpit/ICU-monitor instrumentation logic. Color
+is a rationed resource: the UI is near-neutral grayscale everywhere, so
+the four severity tiers land as genuinely the most saturated colors
+anywhere on screen, not one saturated color competing with a dozen
+decorative ones. See the Phase 5 design-direction review (three options
+were presented; Signal was the one picked) for the two rejected
+directions and the full rationale.
+
+This doc is the source of truth for the token system and component
+library `/web/src/lib/components/ui` implements — read it before adding
+a new color, spacing value, or component, rather than reaching for a
+literal hex or one-off markup the way Phase 0-3's pages did.
+
+## Status
+
+**Status: Phase 5 shipped.** All ten tasks are built: the token system,
+the `ui/` component library, the persistent sidebar nav + command
+palette, a real ECharts-based charting layer (5 chart types, drill-down,
+zoom), the dashboard panel rebuild (GridStack drag-and-drop + live-preview
+panel editor), the query/search redesign (CodeMirror syntax highlighting
++ autocomplete, sortable/resizable/expandable results), the alerting UI
+redesign (severity-state pill + delivery timeline), and an accessibility
+pass driven by real axe-core runs against the live app (not static
+analysis) — see the Accessibility section below for what that actually
+caught. See `/docs/phase-5-runbook.md` for the full verification log.
+
+Verified in a real browser against a live docker-compose stack with real
+seeded data (not just fixtures): dark/light/system theme and
+comfortable/compact density switching persisting across navigation and
+reload; command palette open/filter/keyboard-navigate/go-to; every nav
+route including dashboard/alert detail pages with real panels and real
+alert history; all 6 panel viz types rendering real query results; a
+firing alert rule's full state-history timeline; keyboard-only operation
+of the query editor, results table (sort/expand), and dashboard grid.
+
+## Fonts
+
+Self-hosted (not a Google Fonts CDN link — no runtime dependency on a
+third party for the app to render correctly): `web/static/fonts/`,
+licensed under the SIL Open Font License (see that directory's
+`LICENSE.txt`).
+
+| Role | Typeface | Why |
+|---|---|---|
+| UI (nav, labels, headings, body) | Overpass | Originally drawn for U.S. highway signage — engineered to be read correctly, fast, under bad conditions. A more honest reason to pick a typeface for an incident-response tool than "it looks modern." |
+| Data (log tables, code, query bar, numbers) | Overpass Mono | Same family as the UI face — one typeface end to end removes even the small cognitive cost of a font pairing, matching the direction's overall restraint. |
+
+Both are variable fonts (one file covers the full weight range), loaded
+via `@font-face` in `web/src/lib/styles/fonts.css`.
+
+## Color tokens
+
+Defined in `web/src/lib/styles/tokens.css`. Dark is the literal default
+— `:root` defines the dark palette directly, light is the override
+(both `@media (prefers-color-scheme: light)` for an unset preference and
+`[data-theme="light"]` for an explicit one) — not a retrofit where light
+is `:root` and dark is bolted on. Component CSS must only ever read a
+token, never a literal hex; that's what makes the theme/density toggles
+a token swap instead of a per-component rewrite.
+
+| Token | Dark | Light | Use |
+|---|---|---|---|
+| `--color-bg` | `#0a0a0b` | `#f7f7f8` | Page background |
+| `--color-surface` | `#17181a` | `#ffffff` | Cards, tables, inputs |
+| `--color-surface-raised` | `#1e2023` | `#ffffff` | Hover states, popovers |
+| `--color-border` / `--color-border-strong` | `#2a2c2f` / `#3a3d41` | `#dfe0e2` / `#c7c9cc` | Dividers, input borders |
+| `--color-text` / `--color-text-muted` / `--color-text-faint` | `#f0f0f1` / `#85888d` / `#8a8d92` | `#101113` / `#6b6e73` / `#75787d` | Body text hierarchy |
+| `--color-accent` / `--color-accent-strong` | `#3fb6ff` | `#0b84d6` | Interactive elements only — links, primary buttons, focus rings, active nav. Never reused for severity (see below); a semantic color competing with the brand accent defeats the point of "color means something." |
+
+### Severity tiers
+
+The schema carries seven OTel severities
+(`TRACE`/`DEBUG`/`INFO`/`WARN`/`ERROR`/`FATAL`/`UNSPECIFIED`, see
+`/storage/README.md`). Seven colors would be seven things to memorize at
+a glance; `web/src/lib/severity.ts`'s `severityTier()` collapses them to
+four accent tiers plus one "quiet" state:
+
+| OTel severity | Tier | Dark | Light |
+|---|---|---|---|
+| `TRACE`, `DEBUG`, `UNSPECIFIED` | quiet | `#85888d` | `#6b6e73` |
+| `INFO` | info | `#4c8dff` | `#1a63d6` |
+| `WARN` | warn | `#f5c242` | `#8c6800` |
+| `ERROR` | error | `#ff6a39` | `#c94b1e` |
+| `FATAL` | critical | `#ff2d78` | `#c21362` |
+
+Chosen as a blue → amber → orange → magenta progression — hue *and*
+lightness both shift at every step, so no two adjacent tiers rely on
+red-vs-green to be told apart, and the sequence should survive grayscale
+and protanopia/deuteranopia simulation. That reasoning hasn't been
+verified with an actual simulator yet — do that before treating it as
+confirmed accessible, not just plausible.
+
+Each tier also has a translucent `-bg` token (e.g. `--color-sev-warn-bg`)
+for chip/pill backgrounds. These are tuned independently of the solid
+foreground colors above, not derived from them by a fixed formula: a
+translucent color tints *toward* its own hue as alpha increases, which
+for these saturated, low-luminance hues (critical's magenta especially)
+*lowers* contrast against the foreground text the higher the alpha goes
+— the opposite of the intuitive "more opaque background is safer"
+assumption. `--color-sev-warn` itself was darkened in light mode
+(`#9c7300` → `#8c6800`) for the same underlying reason: axe-core caught
+the *plain* light-mode warn text failing AA (4.32:1) against white
+before any background was even involved. See the Accessibility section.
+
+Use `<SeverityBadge severity={row.severity} />`
+(`ui/SeverityBadge.svelte`) to render one of these — it owns the OTel
+string → tier mapping so call sites can't invent a sixth color by hand.
+For non-log-severity status (success/danger/neutral/accent — e.g. a
+"saved" confirmation), use `<Badge tone="...">` instead; the two are
+deliberately separate components so a form-validation color can never
+accidentally collide with a log-severity one.
+
+## Type scale
+
+14px base (dense-first — this is a tool for reading log tables, not
+marketing copy), 1.2 modular ratio:
+
+`--text-xs` 11px · `--text-sm` 13px · `--text-base` 14px · `--text-md`
+16px · `--text-lg` 20px · `--text-xl` 28px · `--text-2xl` 40px
+
+`--font-weight-normal` 400, `--font-weight-medium` 600,
+`--font-weight-bold` 700. `font-variant-numeric: tabular-nums` is set
+globally on `body` so columns of numbers/timestamps align.
+
+## Spacing, radius, shadow
+
+8px-based spacing scale: `--space-1` through `--space-8` (4px, 8px,
+12px, 16px, 24px, 32px, 48px, 64px). Radius: `--radius-sm` 4px,
+`--radius-md` 6px, `--radius-lg` 10px, `--radius-full` (pills). Shadow
+is used sparingly, matching Signal's restraint — only real overlays
+(modals, the command palette, tooltips) get one (`--shadow-sm/md/lg`);
+inline UI never does.
+
+## Density
+
+Two presets, one token swap — `html.density-compact` overrides
+`--row-height`, `--row-padding-y/x`, `--panel-padding`, `--control-height`,
+and drops `--text-base` to `--text-sm`. Comfortable (the default) suits
+dashboards and forms; compact suits log tables and query results. Toggle
+via `$lib/density.svelte.ts`'s `setDensity()`/`toggleDensity()` — a
+global, persisted (`localStorage['sentry.density']`) preference, not a
+per-page setting, so switching it on one page carries to the next.
+`web/src/app.html` has a synchronous inline script that applies the
+stored value before first paint, so there's no flash of the wrong
+density on reload; keep that script's storage key/values in sync with
+`density.svelte.ts` if either changes.
+
+## Theme
+
+`$lib/theme.svelte.ts`, same persisted/synchronous-apply shape as
+density. Three states — `'dark' | 'light' | 'system'` — but unlike most
+apps, the *unset* default is `'dark'`, not `'system'`. That's the actual
+point of "real dark mode as the default, not an afterthought": a
+first-time visitor on a light-OS machine still lands in dark. `'system'`
+is available as a deliberate opt-in for anyone who wants their OS
+setting to win instead.
+
+## Components (`web/src/lib/components/ui/`)
+
+Import from the barrel: `import { Button, Input, ... } from
+'$lib/components/ui';`
+
+| Component | Notes |
+|---|---|
+| `Button` | `variant`: `primary`/`secondary`/`ghost`/`danger`. `size`: `sm`/`md`. Renders `<a>` when `href` is passed, `<button>` otherwise. |
+| `Input` | Thin styled wrapper over `<input>`; `bind:value`, `invalid` sets `aria-invalid` + a danger border. |
+| `Select` | Styled wrapper over native `<select>` (real `<option>` children) — not a custom listbox, so it keeps native keyboard/screen-reader behavior for free. |
+| `Badge` | Generic status pill. `tone`: `neutral`/`success`/`danger`/`accent`. |
+| `SeverityBadge` | `severity` prop takes a raw OTel string; maps to a tier internally. Use this instead of `Badge` for anything log-severity-shaped. |
+| `Table` | A styling wrapper around real `<table>` markup (native semantics matter for screen readers) — pass real `<thead>`/`<tbody>` as children. Density-aware via the row tokens above. Not a data grid; sort/resize (Phase 5 task 6) layers on top later. |
+| `Card` | `title`/`actions` (a snippet) header, `padded` toggle. |
+| `Modal` | Built on native `<dialog>` — real focus trap, Escape-to-close, and top-layer stacking, not hand-rolled. `bind:open`, `title`, `footer` snippet. |
+| `Tooltip` | Pure-CSS hover/focus tooltip, `role="tooltip"` + `aria-describedby`. |
+| `Tabs` | Renders the tab list only (roving tabindex, arrow-key nav); the caller renders each panel's content on `bind:active` and owns `id="panel-{id}"`/`aria-labelledby="tab-{id}"`. |
+
+`CommandPalette.svelte` and `NavSidebar.svelte`
+(`web/src/lib/components/`, not under `ui/`) are app-shell components,
+not general-purpose library pieces — mounted once in `+layout.svelte`.
+
+## Navigation & tenant switching
+
+Persistent sidebar (`NavSidebar.svelte`): Search / Dashboards / Alerts /
+Data Sources / Settings, active-route highlighting, a command-palette
+hint, and the theme/density quick toggles. The tenant indicator calls
+`getCurrentSession()` (`$lib/api.ts`), which wraps `POST
+/internal/authorize` — an endpoint that already existed (`api/authz.
+HTTPAuthorizer` and `alerting` already call it) and was already reachable
+from the browser (`enterprise-auth`'s whole mux is behind
+`WithCredentialedCORS`, not just the tenant-picker routes), so this
+needed zero backend changes.
+
+**Known limitation, not yet solved at the API level**: there is no
+"list my other tenant memberships while already logged in" endpoint —
+`GET /auth/memberships` only works during the short-lived pending-login
+window (see `enterprise/internal/loginhandler`'s doc comment), not for
+an established session. So "Switch tenant" in the sidebar re-triggers
+login (`/auth/oidc/login`) rather than offering an inline dropdown —
+logging in again is the only tenant-switching mechanism this product
+actually has today, and if there's more than one membership it
+naturally lands back on `/select-tenant`. A real inline switcher would
+need a new `GET /auth/my-memberships`-shaped endpoint (or equivalent);
+that's real, disclosed future work, not a bug in what's built.
+
+## Command palette
+
+`⌘K`/`Ctrl+K` anywhere in the app. Indexes the five static nav
+destinations plus live-fetched dashboards and alert rules
+(`listDashboards()`/`listRules()`, via `Promise.allSettled` so one
+failing endpoint doesn't blank the other's results or the static items).
+Arrow keys + Enter to navigate, Escape to close (native `<dialog>`,
+same reasoning as `Modal`). Substring filter, not fuzzy — no new
+dependency pulled in for it, matching this repo's "boring,
+well-understood dependencies" convention.
+
+## Charting (`web/src/lib/charts/`)
+
+Built on **ECharts**, via modular imports (`echarts/core` plus only the
+chart/component modules actually used — `LineChart`, `BarChart`,
+`HeatmapChart`, `TooltipComponent`, `GridComponent`, `LegendComponent`,
+`DataZoomComponent`, `VisualMapComponent`, `MarkLineComponent`,
+`CanvasRenderer`), not the full bundle. Picked over Observable Plot (SVG
+rendering hits a real performance ceiling at volume, and no built-in
+zoom/pan or legend-toggle) and raw D3 (too much hand-rolled engineering
+for chart types this standard). Verified, not assumed: the lazy-loaded
+chart chunk is 211,975 bytes gzipped, and a synthetic 30,006-row/6-series
+stress fixture (`/dev/charts`, an unlisted dev-only route) renders its
+first two frames in ~50ms on a production build — `npm run dev`'s
+~3.1s figure for the same fixture is pure dev-mode/unminified-JS
+overhead, not a real perf number, and was confirmed as such before being
+discarded.
+
+Five chart components, all consuming the same `{columns, rows}`
+`QueryResult` shape Phase 2's query endpoint has always returned:
+`TimeSeriesChart` (multi-series line, legend toggle), `BarChart`
+(including stacked), `SingleStat` (big number + sparkline + trend),
+`Heatmap`, `TopN` (ranked horizontal bars). `EChart.svelte` is the shared
+base wrapper (init/resize/dispose lifecycle).
+
+**No query-language change was needed for multi-series or drill-down** —
+both are pure frontend reshaping of the existing tabular output:
+
+- `pivot.ts`'s `pivot(columns, rows, config)` turns a "long" result (one
+  row per series+x pair, e.g. `stats count by service, timestamp`) into
+  one series per distinct value of the grouping column. Its value-column
+  auto-detection has one sharp edge worth knowing if you touch it: when
+  a result has only one non-x column (a bare `stats count`, `single_stat`'s
+  most common shape), `Array.prototype.findIndex` returning `-1` for "not
+  found" must be checked explicitly — `-1 ?? fallback` never falls back,
+  because `-1` isn't `null`/`undefined`. This was a real, shipped bug
+  (every `single_stat` panel silently rendered `0`) caught by seeding a
+  live dashboard with real query results rather than only fixture data,
+  not by any static check.
+- `drilldown.ts` strips a panel's query down to its pre-`stats` filter
+  portion, appends a clicked series/x-value as a new filter term, and
+  computes a tight time window around a clicked timestamp — then
+  navigates to the Search page via URL params (`?q=&earliest=&latest=`).
+
+`theme.ts`'s `readChartTokens()` reads real computed CSS custom-property
+values (`getComputedStyle`) so charts render in the actual active
+theme's colors rather than a hardcoded palette — guarded with an
+`SSR_FALLBACK` object for adapter-static's prerender pass, where
+`document` doesn't exist.
+
+**The one backend change in this phase**: `heatmap` as a `VizType`
+(`api/dashboards/types.go`'s `validVizType()`, `web/src/lib/api.ts`'s
+`VizType` union, and the `dashboard_panels` table's `viz_type` CHECK
+constraint — three places, not two; the DB constraint mirrors the Go
+validator and was originally missed, which meant a heatmap panel passed
+API validation and then failed on insert. See
+`metadata/migrations/0035_add_heatmap_viz_type.sql`). Justified under
+the brief's "no query-language or data-model changes except what's
+strictly needed to feed richer visualizations" exception — a heatmap is
+a viz type, not a new query capability.
+
+## Dashboard panels
+
+Drag-and-drop grid is **GridStack**, already a Phase 3 dependency — no
+new library needed, satisfying the brief's "a maintained library is
+fine, don't hand-roll grid physics." `PanelEditor.svelte` (a `Modal`)
+replaces the old inline add-panel form: a debounced live preview reuses
+`PanelViz` directly, so what you see while editing is pixel-identical to
+what renders on save, not a separate preview renderer that can drift
+from the real thing. Empty/loading/error states use `EmptyState` and
+`Skeleton` rather than a blank panel or a raw error string.
+
+## Query & search (`web/src/lib/query-editor/`)
+
+**CodeMirror 6**, not a hand-rolled textarea-plus-overlay highlighter —
+picked specifically because autocomplete needs real cursor-aware popup
+positioning, which a plain textarea can't give you. `language.ts` is a
+`StreamLanguage` tokenizer for the pipe grammar; its `token()` function
+must return real `@lezer/highlight` tag names looked up by string
+(`'controlKeyword'`, `'operatorKeyword'`, `'name.function'` for
+tag+modifier pairs) — a custom `Tag.define()` object's `.toString()`
+looks plausible but silently fails to highlight anything, a real bug hit
+and fixed while building this. `completions.ts` provides context-aware
+suggestions: stage keywords after `|`, stats functions after `stats`,
+field names elsewhere.
+
+`ResultsTable.svelte`: sortable columns (click header — a real `<button>`
+inside the `<th>`, not a clickable `<th>` itself, so it's keyboard-operable
+for free), resizable columns (pointer-drag handles, mouse-only by
+design — the `role="separator"` handle is intentionally not in the tab
+order, matching how most apps treat column resize as a mouse affordance
+rather than a keyboard one), and expandable rows. The row-expand
+interaction was originally a bare `<tr onclick=...>` with no keyboard
+equivalent at all — a real gap the accessibility pass caught (not via
+axe, which doesn't flag missing keyboard handlers on custom widgets;
+caught by manually tabbing through the page), fixed by adding
+`tabindex="0"`, `role="button"`, `aria-expanded`, and an `Enter`/`Space`
+`onkeydown` handler alongside the existing click handler.
+
+"Add as panel to dashboard" (`AddToDashboardModal.svelte`) lets a query
+built on the Search page become a saved panel without hand-copying the
+query string into the dashboard editor.
+
+## Alerting UI
+
+`AlertStatePill.svelte` reuses the log-severity color tiers rather than
+inventing a second color vocabulary: `ok → quiet`, `pending → warn`,
+`firing → critical`. `DeliveryTimeline.svelte` reframes the existing
+`delivery_log` data (no new backend fields) as a vertical timeline —
+"why didn't I get paged" is a chronological question, and a flat table
+answered it less directly than a timeline does.
+
+## Accessibility
+
+Automated checks ran with **axe-core**, injected into the live app in a
+real browser and driven against a real seeded docker-compose stack
+(dashboards with all 6 panel viz types populated with real query
+results, a firing alert rule with real delivery history) — not just
+static markup or empty/loading states, since several of the real
+findings below only appear once real content renders. Real violations
+found and fixed:
+
+- `color-contrast`: `AlertStatePill`'s critical tier at 4.04:1 (below
+  the 4.5:1 AA threshold) when rendered on an opaque `--color-surface`
+  card, not the plain page background most severity chips happen to sit
+  on — see the Severity tiers section above for why translucent
+  backgrounds made this worse, not better, and why quiet/info (dark) and
+  quiet/warn/error (light) were fixed alongside it as the same latent
+  bug, not yet triggered elsewhere only because no page had rendered
+  those specific tiers on an opaque surface yet.
+- `landmark-main-is-top-level` / `landmark-no-duplicate-main` /
+  `landmark-unique`: the layout shell's content wrapper was a second
+  `<main>` nested around each page's own top-level `<main>`.
+- `landmark-one-main` / `region`: `/data-sources` was the one page using
+  a plain `<div>` instead of `<main>` as its top-level element.
+- `heading-order`: `Card` and `EmptyState` both jumped from `<h1>` to
+  `<h3>`, skipping `<h2>`.
+- `aria-input-field-name`: the CodeMirror query editor's `.cm-content`
+  had no accessible name — fixed via `EditorView.contentAttributes.of({
+  'aria-label': ... })`.
+- `empty-table-header`: the alerts list table's trailing actions column
+  had a bare `<th></th>` — fixed with visually-hidden text (`.sr-only`,
+  now a shared utility in `app.css`) rather than a visible "Actions"
+  label that would've added a fifth, less-necessary heading.
+
+Keyboard-only operation was verified by tabbing through, not just
+inferred from markup: sidebar nav → command-palette hint → theme/density
+controls → into main content, with a visible focus ring
+(`--focus-ring`, a two-layer box-shadow so it's visible on both light
+and dark surfaces) at every stop; the query editor is reachable and
+labeled; sort buttons are real `<button>`s; the results table's row
+expansion works via `Enter`/`Space` after the fix above. Responsive:
+the sidebar collapses to an off-canvas drawer under 860px
+(`NavSidebar.svelte`'s `mobileOpen`/`onCloseMobile`, a transform-based
+slide-in with a backdrop button), verified down to tablet-landscape
+width, not just described in CSS.
diff --git a/docs/phase-5-runbook.md b/docs/phase-5-runbook.md
new file mode 100644
index 0000000..675a2d6
--- /dev/null
+++ b/docs/phase-5-runbook.md
@@ -0,0 +1,155 @@
+# Phase 5 runbook
+
+Extends `/docs/phase-0-runbook.md` through `/docs/phase-4-runbook.md`.
+Read those first. Phase 5 touched only `/web` and the two narrow,
+justified backend changes called out below — no new services, no new
+pinned-stack components.
+
+## What's actually been verified
+
+Every claim below was checked against a live `docker compose up`
+stack with real seeded data (not fixture files, not mocked stores),
+in a real browser (Claude in Chrome), not just `npm run check` and
+`npm run build` passing. Five real bugs were found this way — two of
+them backend bugs unrelated to the frontend redesign itself, only
+surfaced because getting real chart/panel/alert data required actually
+exercising the write paths those bugs were in. That's the same
+"a passing check and a working feature are not the same claim"
+discipline every prior phase's runbook has held to.
+
+## 1. Bring up the stack
+
+```sh
+docker compose up -d --build
+cd web && npm run dev   # localhost:5183, talks to localhost:8080/8081 by default
+```
+
+No new services this phase. `docker compose ps` should show the same
+set as Phase 4.
+
+## 2. Seed real data
+
+`hack/benchmark-fixture` (Phase 2) still works unchanged:
+
+```sh
+cd hack/benchmark-fixture
+go run . -count=5000 -batch-size=500 -concurrency=4
+```
+
+Create a dashboard covering all 6 panel viz types and an alert rule
+with a real notification target, either through the web UI or via the
+API directly. One thing worth knowing if you script this the way this
+runbook's own verification did: the query language requires a leading
+base term before any pipe stage — `stats count by host` alone is a
+syntax error (`unexpected COMMA after query` or similar), because
+`Parse()` always parses a base filter/free-text term before it'll
+accept a `|`. `earliest=-1h | stats count by host` is the idiomatic
+"match everything in the current window" form (see
+`/docs/query-language-reference.md`'s `earliest=-24h severity=ERROR |
+stats count by service` example) — this isn't new to Phase 5, but is
+easy to trip over when scripting panel creation instead of using the
+query bar, which always has a real base term from its own default
+query.
+
+## 3. Dashboard panels — all 6 viz types
+
+Time-series line, bar, single-stat, top-N, table were all real Phase 3
+capabilities re-rendered on the new chart layer; heatmap is the one new
+type this phase added. Verify each renders against real query results,
+not just the `/dev/charts` synthetic fixture route (unlisted, dev-only —
+useful for perf testing, not a substitute for exercising the real panel
+CRUD + query path).
+
+**Real bug caught here**: `single_stat` panels rendered `0` for every
+query, including ones the API confirmed returned real data (`{"columns":
+["count"],"rows":[[5000]]}` from `POST /query` showed `5000`, the panel
+showed `0`). Root cause in `web/src/lib/charts/pivot.ts`'s value-column
+auto-detection: `Array.prototype.findIndex` returns `-1` for "not
+found," and the fallback used `??`, which only substitutes on
+`null`/`undefined` — `-1 ?? fallback` evaluates to `-1`, not the
+fallback. For a single-column result (`stats count`, single-stat's most
+common shape), the x-column fallback (index 0) excludes the only column
+from the search, `findIndex` always returns `-1`, and the panel silently
+read `row[-1]` (`undefined`) as its value. Fixed by checking for `-1`
+explicitly instead of relying on `??`. This would not have been caught
+by `/dev/charts`'s synthetic multi-column fixture data — it only
+reproduces with a genuinely single-column result, which only a real
+`stats count`-shaped query against the live API produces.
+
+**Real bug caught here, backend**: creating a heatmap panel returned
+`{"error":"viz_type must be one of table, line, bar, single_stat,
+top_n, heatmap, got \"heatmap\""}`-shaped failures at two different
+layers in sequence. First, the running `api` container was stale
+relative to the Phase 5 source change adding `heatmap` to
+`validVizType()` — rebuilding (`docker compose build api`) fixed that.
+Second, after rebuilding, panel creation still failed with a Postgres
+`23514` check-constraint violation: `dashboard_panels`'s `viz_type`
+CHECK constraint (`metadata/migrations/0002_create_dashboard_panels.sql`)
+was never updated alongside the Go validator, so `heatmap` passed API
+validation and then failed on insert. Fixed with a new migration,
+`metadata/migrations/0035_add_heatmap_viz_type.sql` (drop and recreate
+the constraint — Postgres has no `ALTER CHECK`). Worth remembering for
+any future `VizType` addition: it's a three-place change (Go validator,
+TS union, DB constraint), not two.
+
+## 4. Alert rule creation
+
+**Real bug caught here, backend, unrelated to the Phase 5 redesign
+itself**: creating any alert rule against a freshly migrated database
+failed with `inserting initial alert_state: ERROR: null value in
+column "tenant_id" of relation "alert_state" violates not-null
+constraint`. Phase 4 added `tenant_id` to `alert_state` and
+`delivery_log` (migrations `0022`/`0023`, backfilled via a join through
+`alert_rules.id`, then set `NOT NULL`) but
+`alerting/internal/rulestore/store.go`'s `Create` and `ApplyTransition`
+were never updated to populate it on new inserts — every existing row
+had a value from the backfill, so this was invisible until the first
+rule created *after* that migration ran, which nothing in Phase 4's own
+verification happened to do (Phase 4's Docker access was lost partway
+through, per its own runbook's disclosed gap). Fixed: `Create`'s
+`alert_state` insert now passes `r.TenantID` explicitly;
+`ApplyTransition`'s `delivery_log` insert resolves it via `(SELECT
+tenant_id FROM alert_rules WHERE id = $1)` since that function only
+receives a rule ID, not a full `Rule`. Confirmed fixed against the live
+stack: rule creation, evaluation, firing, and a real delivery attempt
+(to a fake Slack webhook URL — a real HTTP 404 back, logged in
+`delivery_log`, which is itself the intended behavior for an
+unreachable target) all completed end to end.
+
+This means **every core-mode alert rule created against a Phase-4- or
+Phase-5-migrated database before this fix was silently broken** — worth
+knowing if debugging an environment provisioned between those two
+points.
+
+## 5. Accessibility sweep
+
+Automated: axe-core, injected into the live app via a temporary
+`web/static/axe-test-temp.js` script tag (removed before commit — do
+not check it in; recreate from `node_modules/axe-core/axe.min.js` if you
+need to re-run this). Checked every route, including the ones that only
+render meaningfully with real backend data: `/`, `/dashboards`,
+`/dashboards/[id]` (with all 6 panel types populated), `/alerts`,
+`/alerts/[id]` (with a real firing rule and delivery history),
+`/alerts/new`, `/settings`, `/data-sources`, `/select-tenant`,
+`/dev/charts`. Real findings and fixes are cataloged in
+`/docs/design-system.md`'s Accessibility section — don't duplicate that
+list here; the summary is zero outstanding violations across all of the
+above, in both themes, confirmed by re-running axe after each fix
+rather than assuming a fix worked.
+
+Manual: keyboard-only tab-order walk through the sidebar, command
+palette, query editor, and results table (see the design-system doc's
+Accessibility section for the one real gap this caught that axe
+structurally can't — a custom interactive element with no keyboard
+handler at all doesn't trip any axe rule, since axe checks markup/ARIA
+correctness, not "does this thing respond to Enter").
+
+## 6. What's not verified
+
+Same category of gap as every prior phase's disclosed limitations, not
+a new one: real external IdP login, a real multi-container Kubernetes
+deployment, and load-testing the chart layer against ClickHouse-scale
+(millions-of-rows) result sets rather than the 30K-row synthetic
+`/dev/charts` stress fixture — that fixture validates client-side
+render performance, not query performance at that scale, which is
+already covered by Phase 2's own benchmark.