Frontend Performance Baseline — 2026-08-07 (Phase 3D, slice 1)

Scope: miz-oki-command-center-ui — bundle-analysis wiring, first code-splitting pass on heavy visualization routes, fleet-grid virtualization decision, and the route-size budget this slice establishes.

Truth discipline (TRUTH.md): every number below is a measurement of a local production build (npm run build, Next.js 15.5.21, standalone output, built in the working tree on 2026-08-07). These are not live-infrastructure measurements — nothing here upgrades to a "verified result" serving claim until the same build is deployed and measured on Cloud Run. The budget in §5 is a rule going forward, not an observed guarantee.

1. Method

• Baseline build: npm run build on the pre-change tree (exit 0), route table captured from the build output ("First Load JS" per route + shared chunks).
• Changes applied: analyzer wiring (§2), next/dynamic at four page-level import sites (§4), no edits to the heavy components themselves.
• After build: npm run build again (exit 0), same capture.
• Noise floor: the two builds ran ~40 minutes apart in a shared working tree while two sibling work streams edited disjoint zones. Every route not touched by this slice moved by exactly +1 kB (the shared baseline went 103 kB → 104 kB, "other shared chunks" +0.9 kB — the next/dynamic lazy-loading runtime entering the shared graph, plus any concurrent sibling edits). Deltas of ±1 kB are therefore noise; the changed-route deltas below are 40–480× that floor.
• Route count: 399 app routes in both tables — none added or removed.

2. Bundle analyzer (opt-in)

@next/bundle-analyzer (already in devDependencies) is wired in miz-oki-command-center-ui/next.config.mjs behind ANALYZE=1:

ANALYZE=1 npm run build   # writes treemap reports under .next/analyze/
npm run build             # ANALYZE unset -> wrapper disabled, config untouched

Measured 2026-08-07: with ANALYZE unset the exported config's keys are byte-identical to the pre-wrapper config (no webpack hook injected); with ANALYZE=1 the analyzer's webpack hook is present. The Phase-3F security headers block is preserved verbatim.

3. Changed routes — before / after (measured local build)

First Load JS = what a cold client downloads to render the route.

The heavy libraries did not disappear — they moved into async chunks fetched when the component actually renders (with skeleton fallbacks in the page's own loading idiom). /kg/live's per-route size rose 1.3 kB (the dynamic loader + fallback markup) while its first load dropped 43 kB.

Shared chunks (all routes):

4. Code-split inventory

Applied (next/dynamic at the importing page only; SSR left at its default in all four cases — measured 2026-08-07, none of these components reads window/document at module scope, so server-render semantics are unchanged):

CausalGraphExplorer sits in the same tab group as the two mapped ecosystem targets and carries the same libs, so it was split in the same pass — the three tabs now each load on first activation.

Skipped — already dynamic before this slice (measured):

Not split: PlanVisualization (also imports GraphCanvas) is imported only by an archive-ui/ page outside the app build — no live route pays for it.

5. Budget rule (established by this slice)

• Command-center pages must not exceed their captured baseline First Load JS (table below; family ceiling today: 125 kB). A PR that pushes any /command-center/* route above its baseline needs an explicit, measured justification in the PR body.
• The three changed routes must not regress above their after-numbers in §3 without the same justification.
• These budgets are measured local-build numbers (this doc); re-baseline only with a fresh npm run build route table committed alongside the change.

(Baseline column from the pre-change build; the after build read each +1 kB via the shared-chunk delta — the budget tracks the route's own weight, so ±1 kB shared-baseline movement does not count against a page.)

6. Fleet-grid virtualization — measured decision: NOT WARRANTED, SKIPPED

The work item named app/cells/page.tsx; that path does not exist in the tree. The real 36-cell fleet surface is app/command-center/cells/page.tsx (Phase 2C), which also sits outside this slice's writable zone. Measured render counts of that page (2026-08-07):

• Static registry table: 36 rows (FLEET_REGISTRY, FLEET_CELL_COUNT = 36).
• Live health table: 36 rows + probes outside the transcription (bounded by the Boss probe fan-out; 0 in the static registry case).
• Illustrative certification walkthrough: 7 mock cards (CELLS in lib/command-center/data.ts).

Worst case ≈ 79 rendered rows/cards — far under the ~200-row threshold at which @tanstack/react-virtual pays for its complexity. Virtualization was therefore not applied anywhere in this slice (honesty over checkbox): the library remains installed and unused, ready for a genuinely long surface (for example a future events/audit table that renders hundreds of rows in one scroll container).

7. Remaining candidates (largest routes not yet split)

Baseline First Load JS, measured local build 2026-08-07. Attribution beyond the direct imports noted below needs an ANALYZE=1 treemap — that is what §2 is for.

Note for the recharts pair: recharts is chart-per-page here, so the /analytics pattern (dynamic at the page import site) applies directly.

8. Gates at completion of this slice

• npx tsc --noEmit — exit 0.
• npm test — 83 files / 514 tests, all passed (matches the slice-start baseline).
• npm run build — exit 0 before and after the changes (lint + TS build gates enforced by next.config.mjs remained on).

9. Slice 2 (2026-08-07, session 9) — the recharts pair split

Truth discipline: every number below is a measurement of a local production build (npm run build, Next.js 15.5.21, standalone output, 2026-08-07). Nothing here is a live-infrastructure claim.

9.1 Method

• Baseline build: the session orchestrator's fresh full build of the untouched tree (exit 0; route table captured before any wave-1 edits).
• Changes applied: the §7 recharts pair only — each single-file 'use client' page body moved verbatim into a colocated client component (email-intelligence-client.tsx, unified-revenue-client.tsx; sole edit = the exported function name) and each page.tsx became a thin next/dynamic wrapper per the /analytics precedent. Zero content/copy changes to the rendered page bodies.
• After build: one npm run build (exit 0) in the shared working tree while two sibling work streams edited disjoint zones (lib/api/** + proxy routes; lib/bff/adapters/boss.ts + app/api/bff/kg/* + app/command-center/kg-live/*) concurrently.
• SSR semantics: left at the default in both wrappers — measured before deciding: neither page body reads window/document at module scope (email-intelligence fetches only inside effects/handlers; unified-revenue's module scope reads only process.env, server-safe).
• Fallback idiom (measured first, §4 discipline): email-intelligence has its own loading early-return (RefreshCw w-8 h-8 animate-spin text-blue-500
• "Loading Email Intelligence..." centered at h-screen on the default surface) — the dynamic fallback mirrors it byte-for-byte plus role="status". unified-revenue has no loading early-return, so its fallback composes the page's own measured idiom pieces: the container mx-auto py-8 px-4 max-w-7xl shell, its RefreshCw animate-spin in-flight motif, and text-muted-foreground secondary text.

9.2 Noise floor (shared concurrent tree)

399 app routes in both tables — none added or removed. Of the 397 routes this slice did not touch, 389 moved exactly 0 kB and 8 moved exactly +1 kB (/blog, /blogs, /command-center/approvals, /email-agent, /evaluations, /invest, /unauthorized, /vision) — a display-rounding tick consistent with "other shared chunks" moving 3.6 → 3.64 kB while both named shared chunks kept identical content hashes (31255-c7c904d6581c47f1.js, 4bd1b696-bad92808725a934a.js). The floor is therefore ±1 kB; the two changed-route deltas below are 147–156× that floor.

9.3 Changed routes — before / after (measured local build, concurrent tree)

The after column is a real measurement of a completed exit-0 build (2026-08-07, ~20:29–20:35 UTC) taken while both sibling streams were still editing their disjoint zones; §9.2 quantifies exactly how quiet the rest of the table stayed. The wave-close quiet-tree confirmation build (§9.6) is the authoritative close-out number.

Shared chunks (all routes):

The recharts weight did not disappear — it moved into async chunks fetched when each dashboard renders. Budget rule (§5) extended: these two routes must not regress above their after-numbers without a measured justification.

9.4 Remaining candidates — §7 re-ranked (this slice's after build)

The §7 baselines drifted +1 kB on several routes between slice 1 and today's builds (intervening session-8/9 merges); the numbers below are from this slice's after build and supersede §7's ranking. The two recharts direct importers are done; the rest need ANALYZE=1 attribution before splitting.

9.5 Gates at completion of this slice

• npx tsc --noEmit — exit 0, run twice (pre-build and post-slice on the still-concurrent tree); zero errors in any zone both times.
• npm run build — exit 0 (lint + TS build gates enforced by next.config.mjs remained on; compile warnings identical class to baseline — pre-existing OTel winston + exhaustive-deps warnings only).
• npm test -- --run (mid-edit shared tree): 90 files / 549 tests → 542 passed, 7 failed — every failure in sibling-zone files (app/command-center/kg-live/page.test.tsx 5, app/api/boss-proxy/[...path]/route.test.ts 2), zero failures touching app/operations/** or anything this slice imports. Wave-close quiet-tree rerun deferred to the orchestrator's go signal (§9.6).

9.6 Wave-close quiet-tree confirmation (PENDING)

Reserved for the single authoritative build after BOTH sibling streams land and the tree is quiet (orchestrator-signaled). Before numbers are the orchestrator's baseline log; §9.3's concurrent-tree measurements stand as measured but the cells below are the close-out record.

Quiet-tree full-suite rerun: npm test -- --run exit 0 — 91 files / 575 tests, all passed (2026-08-07 wave close; build exit 0, 402 routes).

10. Cache policy (measured, 2026-08-07)

Claim labels: "measured (code)" = read from the tree; "measured (local serve)" = observed on the wire from this slice's built standalone artifact served at 127.0.0.1:3999 (E2E_SKIP_BUILD=1 scripts/e2e-prepare.sh, then node .next/standalone/server.js). No live-infrastructure claims.

10.1 Client cache — React Query (measured, code)

The sole production QueryClient is lib/query-client.tsx, mounted app-wide via app/layout.tsx → components/providers.tsx → QueryProvider (app/journey-intelligence/page.tsx additionally re-mounts the same provider nested — pre-existing, harmless, noted). Global defaults:

Representative per-hook overrides (measured): hooks/api/useKgReads.ts polls both KG summaries at refetchInterval 60 s with retry: 1; hooks/api/useGovernanceReads.ts polls governance reads at 120 s default with retry: 1; hooks/api/useBOSS.ts sets staleTime per surface from 10 s (queue-ish reads) to 120 s; hooks/agent-ide/useAgentIDE.ts uses 30–60 s staleTime + 60 s polls. Pattern: hooks override freshness per surface; the 5-min global default is the quiet fallback.

10.2 Server emission — route code vs next.config (measured, code + local serve)

• Code: 28 of 253 app/api/** route files set a Cache-Control of their own — mostly SSE streams (no-cache[, no-transform]) plus two routes that declare cacheable values: app/api/kg/stats/route.ts (public, max-age=30, stale-while-revalidate=60; stale path public, max-age=10) and app/api/integrations/health/route.ts (public, max-age=60). 0 of the 31 BFF route files set any Cache-Control; each emits NextResponse.json inline — there is no single central success emitter (only the error path shares bffErrorPayload).
• Config: next.config.mjs headers() sends Cache-Control: no-cache, no-store, must-revalidate on the /:path* catch-all (which matches /api/**) and public, max-age=31536000, immutable on /_next/static/:path*.
• Wire (measured, local serve): pages (/operations/email-intelligence, /operations/unified-revenue), a BFF route (/api/bff/approvals/pending, 401 tenant_required), /api/kg/stats (403 upstream-mirror path), and /api/config/a2a-gateway (200) all emitted the config triple no-cache, no-store, must-revalidate; the static chunk emitted public, max-age=31536000, immutable. The a2a-gateway data point is decisive: its only code path sets Cache-Control: no-store, yet the wire carried the config value — the config catch-all overrides route-handler-set Cache-Control on the served response. Consequences (documented, not changed): kg/stats' and integrations/health's public, max-age never reach the client today; SSE routes' no-transform is likewise overridden; and unit tests that assert route-set headers (e.g. app/api/config/a2a-gateway/route.test.ts expecting no-store) pass at unit level while the wire differs.
• Static assets: fingerprinted /_next/static immutable-for-1-year is the correct posture and is confirmed on the wire.

10.3 Decision: DOCUMENT-ONLY (criteria measured)

An explicit Cache-Control: no-store at a central BFF emitter was NOT implemented, because both instruction criteria fail: (1) no single central point emits BFF successes — all 31 BFF routes call NextResponse.json inline, so the change would touch >3 files, including the sibling-zone app/api/bff/kg/*; (2) it would change zero wire bytes today — the config catch-all already serves no-cache, no-store, must-revalidate on every API/BFF response (measured, local serve).

Recommendation (recorded for a future slice): the config catch-all is currently the single enforcement point for the no-store posture — do not relax it without first revisiting the 28 route-level values (especially the two public, max-age routes, which would silently become shared-cacheable — a tenant-data risk once auth arms — and the SSE no-transform loss). When a central BFF success emitter is eventually created (One Boss consolidation), bake Cache-Control: no-store into it with a colocated test so the BFF envelope's posture stops depending on the page-cache catch-all.