perf: finish smooth navigation follow-up

docs/architecture.md

@@ -107,7 +107,7 @@ CI and staging must not upload authenticated browser captures. Logging must not
 
 ## PWA and offline behavior
 
-The service worker and manifest provide installable web assets and an offline app shell. It precaches only the navigation shell and static identity assets; route chunks enter the runtime cache when a user visits them. Updates wait for the normal service-worker lifecycle instead of reloading an active page. Offline support does not make unvisited routes, authenticated Supabase, or AI operations available without a network.
+The service worker and manifest provide installable web assets and an offline app shell. It precaches only the navigation shell and static identity assets; route chunks enter the runtime cache when a user visits them. On first registration, the client also reports same-origin `/assets/` resources that loaded before the worker took control so the first visited route works offline without a forced reload. Updates wait for the normal service-worker lifecycle instead of reloading an active page. Offline support does not make unvisited routes, authenticated Supabase, or AI operations available without a network.
 
 ## iOS shell
 

docs/performance/smooth-navigation-final/README.md

@@ -0,0 +1,201 @@
+# Smooth navigation final evidence
+
+Captured on 2026-09-21 from `31549f99cccedba6f4e6c482ada42512520270cc`
+with Node `22.23.2`, npm `12.0.2`, Playwright `1.62.1`, Chromium
+`151.0.7922.34`, Vite `7.3.6`, and the same unthrottled loopback profiles as
+the baseline. This is local lab evidence. Stub sessions and blocked external
+requests prove client rendering and request sequencing, not authenticated
+backend, staging, field, or production performance.
+
+## Artifacts and method
+
+- [`task-11-audit.json`](task-11-audit.json) is the pre-change run at
+  `6d8c9178bb2ebfff65b1c44a934890cfc727a21f`. Its old `sourceCommit` field
+  followed the moving `origin/main`; `harnessCommit` is the authoritative code
+  revision. This run used the corrected pre-pointer resource boundary, so its
+  first-visit request and byte values replace the stale PR #51 values.
+- [`final-1.json`](final-1.json) and [`final-3.json`](final-3.json) are the
+  accepted repeated runs at `31549f9`. An intermediate run had one 58.2 ms
+  warm-navigation outlier, so a third complete run was taken instead of hiding
+  it. Runs 1 and 3 stayed within the documented tolerance: maximum timing
+  variance was 23.94% with only a 6.8 ms absolute difference, request variance
+  was 0%, transferred-byte variance was below 0.001%, and CLS variance was 0.
+- [Desktop delayed video](desktop-delayed.webm),
+  [desktop filmstrip](desktop-delayed-filmstrip.png),
+  [Pixel 5 delayed video](pixel-5-delayed.webm), and
+  [Pixel 5 filmstrip](pixel-5-delayed-filmstrip.png) show the 750 ms stalled
+  Attack chunk. The shell remains mounted and the route-shaped skeleton stays
+  inside the main content region.
+
+Each run has 38 records: 16 cold refreshes, 10 first-visit clicks, 10 warm
+clicks, and two delayed clicks. The final harness also records same-origin
+resource URLs, route-chunk completion, post-chunk render time, long tasks after
+the chunk, controlled Event Timing, and CDP script/task/layout duration. CDP
+`ScriptDuration` covers all JavaScript execution inside the measured boundary;
+it is not presented as isolated evaluation time for one module. The
+`postChunkRenderMs` field covers module execution plus React rendering after
+the route response completes.
+
+## Task 11: page data and rendering audit
+
+The first audit showed route files finishing in 9-32 ms while first real
+content often arrived 145-455 ms after click. Most of that local-development
+gap was module execution and React work, not route-file transfer. Empty stub
+data did not reproduce expensive list, Markdown, chart, or storage hydration
+that justified virtualization or a new rendering abstraction.
+
+One backend cause reproduced deterministically. Generate -> Attack remounted
+`ModelSelector` and repeated the subscription settings and daily-usage reads.
+The browser regression failed three times with four reads instead of two.
+`useSubscription` now uses the existing TanStack Query client with a 60-second
+freshness window and an explicit refresh path. Desktop and mobile regressions
+now record one settings read and one usage read across both routes.
+
+No private data or AI generation is prefetched.
+
+## Task 12: public rendering audit
+
+The landing page renders one image asset twice: `/favicon.png` was 1024x1024,
+transferred 279,765 bytes in the production browser audit, and lacked intrinsic
+dimensions while displayed at 32-36 px. The new 72x72 display asset transfers
+4,274 bytes, a 275,491-byte reduction, and both callers declare their rendered
+width and height. Desktop and Pixel 5 CLS remained `0`.
+
+Production CSS coverage marked 29,400 of 116,213 bytes used on desktop and
+27,330 bytes used on Pixel 5. This does not by itself prove that the unused
+rules can be removed because the stylesheet is shared by lazy application
+routes.
+
+A `content-visibility` prototype for all below-fold sections cut some desktop
+long-task time but overestimated initial scroll height by 36% on desktop and
+22% on mobile. Limiting it to the Features section gave no repeatable mobile
+benefit and still changed initial scroll height. It was rejected. The current
+SEO DOM, anchor navigation, find-in-page behavior, and visual layout remain
+unchanged.
+
+## Task 13: HTTP and service-worker audit
+
+Vercel already sends no-cache headers for `index.html`, `theme-init.js`, and
+`sw.js`, plus one-year immutable caching for hashed assets. Nginx previously
+let its generic one-year JavaScript rule catch `sw.js`; exact no-cache locations
+now cover both `index.html` and `sw.js`.
+
+The first controlled page could load route assets before the newly installed
+worker claimed it. The client now reports only same-origin `/assets/` resources
+for five seconds after registration. The worker validates and caps that list,
+then backfills those already visited assets into the runtime cache. It still
+does not call `skipWaiting`, listen for `controllerchange`, or reload the page.
+
+The committed production-build smoke in
+[`scripts/test-service-worker-offline.mjs`](../../../scripts/test-service-worker-offline.mjs)
+provides direct browser proof:
+
+- a seeded v2 worker cache is removed when v3 activates and claims the page;
+- first-visited Generate reloads offline;
+- an unvisited Attack route shows the existing route error offline;
+- an intent-prefetched Attack route opens offline after its prefetch finishes;
+- runtime cache inspection found zero `/auth/v1`, `/rest/v1`, `/functions/v1`,
+  or `/storage/v1` entries;
+- stale or unavailable chunks return the existing `504 Offline` response and
+  route-level recovery UI rather than a blank document.
+
+## Complete matrix
+
+Cold rows use `requests/transferred/decoded/real-content-ms/CLS/long-tasks`.
+Final values are accepted run 1 and run 3.
+
+```text
+profile route          before                                  final 1                                 final 3
+desktop /              133/7274710/7234810/3049.5/0/3         133/7004559/6964659/2715.8/0/2         133/7004547/6964647/2925.0/0/2
+desktop /login         121/6750638/6714338/711.8/0/1          121/6755453/6719153/582.5/0/1          121/6755441/6719141/564.3/0/1
+desktop /app           119/6960389/6924689/865.8/0/1          119/6689713/6654013/773.4/0/1          119/6689701/6654001/654.9/0/1
+desktop /app/generate  162/8586542/8537942/922.1/.0029/2      162/8312240/8263640/863.6/.0029/2      162/8312228/8263628/893.3/.0029/2
+desktop /app/attack    163/8182519/8133619/1186.9/.0110/2     163/7908217/7859317/880.9/.0110/2      163/7908205/7859305/906.0/.0110/2
+desktop /app/flowchart 129/7732752/7694052/1064.1/0/2         129/7462076/7423376/792.4/0/2          129/7462064/7423364/708.0/0/1
+desktop /app/history   128/7190734/7152334/959.4/.0105/3      128/6920058/6881658/705.3/.0105/1      128/6920046/6881646/739.0/.0105/1
+desktop /app/admin     135/8830912/8790412/972.0/0/2          135/8560236/8519736/820.5/0/2          135/8560224/8519724/835.7/0/2
+Pixel 5 /              133/7274710/7234810/975.9/0/3          133/7004559/6964659/777.2/0/2          133/7004547/6964647/679.8/0/2
+Pixel 5 /login         121/6750638/6714338/632.5/0/1          121/6755453/6719153/527.3/0/1          121/6755441/6719141/530.6/0/1
+Pixel 5 /app           119/6960389/6924689/715.1/0/1          119/6689713/6654013/693.4/0/1          119/6689701/6654001/582.5/0/1
+Pixel 5 /app/generate  162/8586542/8537942/1466.0/0/3         162/8312240/8263640/973.4/0/2          162/8312228/8263628/868.7/0/2
+Pixel 5 /app/attack    163/8182519/8133619/970.2/0/3          163/7908217/7859317/888.7/0/2          163/7908205/7859305/883.4/0/2
+Pixel 5 /app/flowchart 129/7732752/7694052/1261.3/0/4         129/7462076/7423376/858.4/0/2          129/7462064/7423364/713.4/0/2
+Pixel 5 /app/history   128/7190734/7152334/1409.7/.0464/3     128/6920058/6881658/800.0/.0464/1      128/6920046/6881646/663.2/.0464/1
+Pixel 5 /app/admin     135/8830912/8790412/1322.3/.0825/3     135/8560236/8519736/801.8/.0825/2      135/8560224/8519724/705.9/.0825/2
+```
+
+First-visit rows use
+`requests/transferred/decoded/real-content-ms/route-chunk-ms/CLS/long-tasks`.
+
+```text
+profile transition           before                               final 1                              final 3
+desktop Landing -> app       2/36896/36296/94.3/9.4/0/0          2/36896/36296/118.6/17.2/0/1        2/36896/36296/113.2/11.4/0/1
+desktop Home -> Generate     44/1652826/1639626/396.5/9.0/0/1    44/1649200/1636000/279.2/11.2/0/1   44/1649200/1636000/315.2/10.1/0/1
+desktop Generate -> Attack   8/221821/219421/323.5/17.4/0/3      8/221821/219421/154.1/9.7/0/1       8/221821/219421/154.4/8.8/0/0
+desktop Home -> Flowchart    13/811783/807883/198.2/12.2/0/1     11/799036/795736/203.4/9.8/0/0      11/799036/795736/169.0/11.0/0/0
+desktop Home -> History      12/269765/266165/144.6/15.8/0/0     10/257018/254018/118.3/12.4/0/0     10/257018/254018/136.8/10.4/0/0
+Pixel 5 Landing -> app       3/242473/241573/168.8/27.2/0/1      2/36896/36296/99.0/10.7/0/0         2/36896/36296/81.8/7.1/0/0
+Pixel 5 Home -> Generate     44/1652826/1639626/455.0/21.2/0/2   44/1649200/1636000/301.7/10.3/0/1   44/1649200/1636000/314.2/6.5/0/1
+Pixel 5 Generate -> Attack   8/221821/219421/320.2/19.2/0/3      8/221821/219421/146.9/12.6/0/0      8/221821/219421/150.2/8.1/0/0
+Pixel 5 Home -> Flowchart    11/799036/795736/328.1/32.3/0/2     11/799036/795736/154.9/11.6/0/0     11/799036/795736/143.9/10.8/0/0
+Pixel 5 Home -> History      10/257018/254018/205.8/15.5/0/0     10/257018/254018/125.6/10.1/0/0     10/257018/254018/103.3/12.3/0/0
+```
+
+Warm rows use `real-content-ms/controlled-event-ms/script-ms/long-tasks`.
+Every warm navigation issued zero requests and transferred zero bytes.
+
+```text
+profile transition           before real   final 1                 final 3
+desktop Landing -> app       32.5          40.2/96/64.3/0          45.4/160/71.1/0
+desktop Home -> Generate     91.4          86.4/136/127.1/1        79.7/96/104.9/1
+desktop Generate -> Attack   67.9          75.7/96/85.3/1          63.0/96/75.3/1
+desktop Home -> Flowchart    52.2          48.4/88/72.5/1          50.6/104/86.9/1
+desktop Home -> History      26.0          29.0/56/38.5/0          25.7/40/39.6/0
+Pixel 5 Landing -> app       37.4          29.9/48/46.2/0          30.6/56/52.5/0
+Pixel 5 Home -> Generate     72.9          78.3/88/112.9/1         71.3/80/97.7/1
+Pixel 5 Generate -> Attack   102.3         65.4/80/87.8/1          62.5/72/84.8/1
+Pixel 5 Home -> Flowchart    76.2          47.9/72/83.3/0          41.7/64/66.6/0
+Pixel 5 Home -> History      28.6          28.4/40/40.9/0          21.6/32/31.3/0
+```
+
+## Bundle and budget result
+
+Static graph gzip changed as follows:
+
+```text
+route          before       final       delta
+index.html     322,213      322,480       +267
+/              374,629      374,938       +309
+/app           324,010      324,288       +278
+/app/generate  453,060      456,162     +3,102
+/app/attack    406,960      410,059     +3,099
+```
+
+The largest critical gzip increase is 0.76%, below the 10% startup budget.
+The build gate still excludes PDF, DOCX, chart, ReactFlow, Admin, and unrelated
+page modules from initial routes.
+
+Accepted-run budget result:
+
+- warm real-content maximum: 86.4 ms desktop and 78.3 ms Pixel 5;
+- ordinary controlled Event Timing maximum: 160 ms;
+- first-visit and warm CLS: `0` for every tested transition;
+- intent-warmed and warm transitions show no generic route loader;
+- constrained-connection tests issue no speculative route request;
+- the delayed chunk keeps shell identity, shows the route skeleton, and records
+  CLS `0.0110` desktop and `0` Pixel 5.
+
+The delayed Pixel 5 recording reports 272-280 ms Event Timing because the
+test captures screenshots synchronously around an artificial 750 ms stalled
+request. It is filmstrip instrumentation, not the ordinary controlled INP
+result. All non-instrumented first-visit and warm interactions remain below
+200 ms.
+
+## Limits
+
+- The baseline did not record CDP script duration or Event Timing, so those
+  fields are reported for the final runs without inventing before values.
+- Local Vite timings include transform and shared-host noise. Request and byte
+  evidence is more stable than absolute cold timing.
+- Stub sessions do not prove Supabase authorization, RLS, billing, or real
+  account data. No production or staging target was used.