cesnimda
17 KiB
P6-04 — Performance and component reuse review
Scope, baseline and evidence method
Measure representative existing workflows and audit real reuse/fetch/save/preview owners before proposing changes. The Phase5 final scope supplies historical finite Board200/1,000 membership and render observations, native CV16/36 and Settings8/24 output, and retained reliability debt. Its1,040/1,612ms Board render measurements are historical single-run data, not a budget or directly comparable current result. No supported-load/CPU/memory/latency budget was previously approved.
Fresh native measurement is owned by the responsive audit performance case. The current post-Career-spacing-repair run records Chromium 151.0.7922.34, a1280×720 CSS-pixel viewport, Windows win32, production static export, Release API, isolated synthetic SQLite, prebuiltExport: true and serverReuseAllowed: false from the actual invocation. It records hard-navigation time to the stated assertion, network requests/resource transfer/duration, long-task/layout-shift observations and CDP heap/DOM/layout/task metrics. Counts and exact API-to-rendered ID parity are asserted at200 and1,000; the separate unchanged diagnostics lane checks its seeded collection. Filtering never silently caps the Board.
The performance case covers initial Jobs, Dashboard, Analytics6/24months, job details, long-CV initial settled preview, CV-only edit including existing800ms autosave, Board API/render at200/1,000, corresponding Table navigation and filtered Board. A heading-only ready milestone measures route presentation, not complete report/list data settling; CV and Board milestones wait for the stated actual preview/count. Hard navigation includes boot/auth and is distinct from client route transitions, search debounce and React rendering counts. Single-run CDP heap samples cannot establish a memory leak or physical-device capacity.
The extended diagnostic measures actual client Board→Table→Board activation while retaining994 matches, typed search through the existing debounce to exactly one result, and Jobs→Dashboard navigation to the real1,000 tracked-job value. An unchanged document performance.timeOrigin proves these are client transitions. Request paths include query strings for these interactions. The retained first153/157 ordinary attempt loaded an earlier diagnostic before that test-only extension. The fresh post-CSS standalone run includes the extension; the final159/159 ordinary run also passes with the same extended case and zero skips/retries/flaky results. Extension adds evidence within the existing browser case, not an application change or another case count. Separate native/security/fidelity evidence and conditional integrated acceptance are recorded in P6-05.
Two preliminary diagnostic failures are retained: the Dashboard's determinate onboarding progress indicator was incorrectly treated as a loading indicator, and identical profile bullets normalized to a one-page document. The fixture now supplies eight distinct roles with four uniquely identified bullets each and asserts actual multi-page readiness. These were diagnostic milestone/fixture corrections, not product defects, renderer changes or relaxed pagination assertions.
The fresh post-CSS standalone diagnostic passes1/1, with zero skips/retries/flaky results. Performance manifest and exact command/results record the2026-10-11 08:46:49.646–08:47:46.501 UTC invocation, raw report SHA-256 53d65dfd5a7352a90bddc788114fb2d7ae3fd8aba9e1a53375a5f4bffe1c8f8a. Its browser report started08:46:50.191 UTC and lasted56.276seconds. Integrated regression is separate in the aggregate validation. No comparative optimization or supported-load budget is claimed.
| Measured workflow / milestone | Observed milliseconds |
|---|---|
| Initial Jobs heading plus successful list response | 673 |
| Dashboard tracked-job link plus stats response | 905 |
| Analytics6 /24months heading plus stats response | 767 /770 |
| Job workspace heading plus owned workspace response | 923 |
| Four-page CV: all visible page frames ready, preview not busy | 1,326 |
| CV-only edit persisted and Saved, including existing autosave delay | 996 |
| Complete Board200 /1,000 API-to-DOM membership | 1,026 /1,657 |
| Table200 /1,000 heading plus list response | 571 /589 |
| Filtered Board194 /994 cards | 958 /1,611 |
| Client Board→Table /Table→Board, retaining994 matches | 319 /695 |
| Typed search through debounce to exactly one matching card | 1,490 |
| Client Jobs→Dashboard with1,000 tracked-job metric | 184 |
Board200 response is237,273 bytes/24ms; Board1,000 is1,181,982 bytes/21ms. Board render long tasks are73/92/96ms versus63/397/264ms, with zero observed eligible layout shift at the Board milestones. Initial Jobs records a137ms long task and sampled shift0.0546968994; long-CV initialisation records a52ms long task and shift0.0128848606. Dashboard performs two reminder reads and two auth/me reads from separate shell/onboarding/report owners and records shift0.0867047106 at its ready milestone. Recorded shifts vary by route and sampling milestone; they remain observations rather than a full-session CLS certification. Automated action times include control activation/assertion overhead. These finite results preserve uncapped membership but justify separately scoped Board capacity and report-read/layout investigations.
The third preliminary failure was a diagnostic literal expecting 1000 rather than the existing English-formatted 1,000 accessible Dashboard metric. It is corrected without changing the metric or formatter; that failure's complete output folder/report/log is archived. The two earliest diagnostic failures retain exact reports/logs but some old attachment paths were overwritten before complete-folder archival was added.
Static exported bundle snapshot
Source-review inspection of the existing local export found buildID nlP_a9RD_LvG_Wn2riffj,53 JavaScript files and 3,233,467 aggregate raw bytes under out/_next/static. This export predates current Phase6 corrections and has no verified current-source provenance. The five largest chunks were432,558;417,647;257,695;228,034;207,255raw bytes; individually gzip-compressing the first produced128,684bytes. These are static artifact observations, not initial-route transfer, effective server compression or current-build measurements. Summing lazy chunks does not describe first load.
The actual .next/build-manifest.json is ignored/generated and must be sampled again after the final production build rather than committed or treated as a stable documentation link. App already lazily imports the main private domain routes, AddJobModal/commands and other substantial surfaces; JobTable separately lazily imports KanbanBoard. Suspense boundaries exist. No evidence currently justifies another framework/chunking/dependency rewrite. Fresh bundle/network observations belong in the manifest, with actual SHA/build identification.
Current verified production build 2LDblja_XnSBWSuY6ys_V has53 exported JS files:3,234,867 aggregate raw bytes and940,065 summed individual gzip bytes. The manifest records every chunk SHA256; these include initial and lazy assets. Local browser routing blocks non-local network and affects caching, so hard-navigation resource transfers do not certify deployed compression or warm-cache behavior. The older unverified export above is attributed history and not a matched before/after comparison.
ResourceTiming reports zero transferSize for the cross-origin API entries because that measurement is unavailable; it does not imply zero-byte API responses. The separate Board bytes field counts actual response-body bytes. Exported JavaScript totals and same-origin resource timing have their own stated scope. CDP heap/DOM values include retained navigation/iframe documents and are context snapshots, not isolated route footprints, allocation totals or a leak diagnosis. Layout-shift observations end at the stated milestone, not the full session lifetime; no unsupported pass/fail budget is attached to them.
Fetch, computation and preview source audit
| Actual owner | Existing efficiency/protection | Candidate or legitimate repeated work; required evidence before change |
|---|---|---|
| Jobs JobTable/KanbanBoard/jobsQuery | Shared normalized committed filters;250ms debounced Board query; Table loader disabled in Board; page-owned Table read; memoized row models/selection/group derivation and latest-resource sequence | Board is complete-array O(N) response/render plus grouped array sorting; no virtualization/cap. Native request/membership/timing/DOM/heap observations at200/1,000 establish finite envelope. Bigger backend projection/virtualization is a separately scoped compatible design, not implied by a slow-looking screenshot |
| DashboardView/Analytics | One report owner; summary stats/overview/reminders parallel, independent trends analytics/tags/tag-trends parallel; charts derived via useMemo; Analytics skips reminder feed and uses URL months | Shell independently requests the same reminder feed for its badge on mount/60second interval. This source overlap serves two responsibilities; measure actual requests/bytes and error/lifetime needs before sharing a cache. Month changes should not be miscounted as initial duplicates; no fabricated new metrics |
| useViewResource | Latest-request sequence prevents superseded results, keeps existing data during refresh, normalizes401/403 versus unavailable errors, enabled loader boundary | Dependency changes rerun load intentionally; loaderref prevents callback identity alone from being the dependency. No transport-abort/memory-leak claim follows from this hook; retain existing shared Axios session/owner cancellation. Replacing all loaders is unsupported |
| useCompanies | Owner-keyed module cache and shared inflight request, auth-user event invalidation, explicit cache-bust reload | Other dedicated Companies/command screens have distinct reads. Do not create shared cross-account cache or assume every company loader can be merged; compare freshness/invalidation/command cancellation consumers first |
| Job ApplicationWorkspacePage/ApplicationAssets/ApplicationIntelligence | Overview load by job; sections own real artifact/intelligence reads; component boundaries keep drafts/history/provider operations separate | Source-level duplicate-like attachment/CV reads can have legitimate consumer/lifetime semantics. Legacy useJobWorkspaceBaseData is not evidence of mounted production work without an actual caller. Measure the current routed workspace; do not optimize unreachable historical hooks |
| CvBuilderEditor | Variant/themes/outline load in parallel;300ms preview debounce/request sequence, retained latest good HTML/error/Retry,800ms save debounce/serialized queue/version conflict. Content stays mounted;900px pane switch changes visibility, not editor identity. Existing hidden physical ruler/paginated pages preserve rendering fidelity | Long document has one hidden measurement iframe plus page iframes with complete generated HTML. Layout/heap/long-task cost needs actual measured page count/fixture and source-equivalent exports. No needless remount/global renderer/virtualization change merely to reduce iframe count; keep original16/36+8/24 oracle |
| I18nProvider and locale-aware read consumers | Source-review baseline defines t inline, recreated whenever provider renders; locale/error text intentionally refreshes on locale change |
PF-01 tested hypothesis: feedback toast triggers another locale-dependent load. The actual-provider/CompaniesTable regression observes one initial GET, no extra GET after toast, then exactly one locale-driven GET after switching to NB. The tested trigger is disproven; no translator memoisation/source change is justified |
| ToastProvider | Stable useCallback toast and memoized context value; measured bottom clearance prevents controls being obscured | Real-provider stability test confirms its toast update does not repeat the tested Companies read. Do not infer an automatic fetch loop from inline t alone or generalize one route to every possible provider rerender |
The source baseline's exact [t] request/resource consumers include CompaniesTable, RemindersView, NotificationsPopover, SessionsSettingsCard, SettingsView, TwoFactorSettingsCard, AdminAuditPage, AdminSystemPage, CareerProfilePage, CorrespondenceInboxPage, CvBuilderPage, OfferComparisonPage, OperationsPage and ProfilePage. Dashboard resources, JobTable/Board and job asset/intelligence/workspace callbacks also include translator dependencies alongside real state. ImportExportJobs has a harmless [t] memo, not a network effect. These are a caller map, not measured redundant requests. The tested toast trigger adds none; retain current dependencies and genuine locale/error-text refresh behavior.
Component reuse and maintainability
| Pattern | Reuse decision / preservation boundary |
|---|---|
| Page hierarchy, section surfaces, controls and feedback | Retain PageHeader/PageSection, semantic theme, MUI Button/TextField/Table/Dialog/Drawer, ViewStateNotice and existing toast/confirm/prompt. P6 scoped company error/language feedback and Settings semantics use those owners. No new presentation framework or duplicate notification stack |
| Resource/state loaders | useViewResource and domain-local useAsset/useIntelligence overlap superficially; their retry/cancellation/setData/save consumers differ. No consolidation solely to reduce line count. Fresh evidence of a repeated failure or real shared behavior is required before extraction |
| Dashboard/Analytics reports | Existing shared DashboardView reporting mode is already reuse; no duplicate chart fetching owner/new chart library |
| Jobs/Table/Board | Shared query semantics and density remain, with table versus full-board data shape intentionally distinct. Sharing state never means one API response or capped Board membership |
| CV editor/library/templates | Existing CvTemplateGallery/CvCustomizeTab/settings client are reused; one editor/save/history/renderer remains. Printed templates keep their own physical tokens; app density does not change PDF layout |
| AI review | Domain differences preserve explicit source/destination, draft versus persisted item versus autosave and stale token protection. Share layout only after proving all consumers; no universal Apply or automatic model invocation |
| Settings/account/providers/admin | Existing independent subcard/API scopes and roles remain. Remaining legacy page presentation migrates in Phase7; aggregate admin metrics/new Notifications/Documents projections require real contracts first |
Bounded outcome and remaining work
This checkpoint supplies source findings and the executed reproducible native baseline. Runtime timings, large-result bytes, resource observations, long tasks/layout shifts and heap snapshots belong to the fresh manifests. No invented performance budget, unbounded capacity claim, React rerender count, comparative gain or leak diagnosis is supplied.
PF-01 has a meaningful read-count/localisation regression using the actual provider/Companies owner; its toast-trigger hypothesis is disproven and existing source is retained. It is included in the final521-case frontend run (85 suites), separately from the native diagnostic. Measured shell/report request overlap is retained for bounded owner/freshness-aware investigation rather than an unqualified cache. Leave major backend/Board/document optimization to an explicit supported-workload design.
Retain Board/reliability/manual/rendering debt, historical language/notification/Rules500 causes, IF-CV-UNSAVED-01 and production font/device/ingress/recovery gates. Original document/page/source/owner/proposal contracts are protected by the 57-group matrix. Current commit/push verification is the Phase6 ledger. Production stays NO-GO; no Phase7 or Phase8 implementation occurs here.