feat(feature-flags): turn insights-visual on for all users

docs/feature-flags.md

@@ -57,7 +57,8 @@ it can never ship a flag that silently resolves to `undefined`.
 > **Every `example-*` key in this document is illustrative, not a real flag.**
 > Grep for one and you will find only this file. The registry holds exactly the
 > flags that gate live behaviour — today four rollout gates, all plain booleans
-> (`insights-visual` also runs a 50/50 experiment):
+> (`retention-policies` and `insights-visual` default on and are kill switches;
+> `insights-visual` ran a 50/50 experiment first):
 > `retention-policies`, `lore-explorer-instruments`,
 > `lore-explorer-duplicate-clusters` and `insights-visual`. It has also held entries since removed: the
 > rollout gate for the `/insights` route (deleted once that page became
@@ -260,7 +261,8 @@ nothing is the correct, expected symptom of that gap, not a broken query.
 Per-user splits need no storage: a signed-in user is bucketed by their Supabase
 user id (see [Targeting key resolution](#targeting-key-resolution)), so they see
 the same arm on every device for as long as the weights stand still.
-`insights-visual` runs this way and is the worked example.
+`insights-visual` ran this way (50/50, before it was turned on for every user)
+and is the worked example.
 
 ### 1. Declare the arms, dark
 

packages/feature-flags/generated/flags.manifest.json

@@ -57,35 +57,20 @@
     },
     {
       "key": "insights-visual",
-      "description": "Insights — the visual first version of `/insights`: ONE page-wide window, the Learn → Recall → Promote → Prune loop strip, the Pulse timeline (lessons written and records read on a shared time axis, chosen vs bulk), the recall flow (client → tool → scope type), and runs drawn as a weekday × hour rhythm plus swimlanes — with the existing 'earning its place' grid, runs list and operational health kept below. Gates the whole page body through the `InsightsView` resolver; with it off `/insights` renders exactly as before. Reads only routes that already ship (`/memories/activity`, `/read-activity`, `/usage`, `/usage/runs`, `/utility`), so there is no backend prerequisite and flipping it is a pure UI decision. RUNNING as a 50/50 experiment: each signed-in user is assigned an arm by a hash of their user id (the same arm on every device), a developer override still wins, and setting `enabled: false` sends everyone back to `off`. `on` is listed first so ramping its weight up never moves a user from `on` back to `off`; lowering it or reordering the arms reassigns users mid-experiment. Exposure is measured by `feature_flag.exposed`, emitted by the `InsightsView` resolver — see docs/feature-flags.md § Running an experiment.",
+      "description": "Insights — the visual first version of `/insights`: ONE page-wide window, the Learn → Recall → Promote → Prune loop strip, the Pulse timeline (lessons written and records read on a shared time axis, chosen vs bulk), the recall flow (client → tool → scope type), and runs drawn as a weekday × hour rhythm plus swimlanes — with the existing 'earning its place' grid, runs list and operational health kept below. Gates the whole page body through the `InsightsView` resolver; with it off `/insights` renders exactly as before. Reads only routes that already ship (`/memories/activity`, `/read-activity`, `/usage`, `/usage/runs`, `/utility`), so there is no backend prerequisite and flipping it is a pure UI decision. DEFAULTS ON for every user: it ran as a 50/50 experiment (users bucketed by a hash of their user id, exposure measured by `feature_flag.exposed` from the `InsightsView` resolver — see docs/feature-flags.md § Running an experiment), and that split has ended. The flag is now a kill switch for turning the visual page back OFF, not a rollout gate; a developer override still wins. Retiring it is the copy-and-suffix removal: promote `InsightsView.on.tsx`, delete `InsightsView.off.tsx` and the `useFeatureFlagExposure` call, drop this entry.",
       "type": "boolean",
       "variants": {
         "off": false,
         "on": true
       },
-      "defaultVariant": "off",
-      "experiment": {
-        "enabled": true,
-        "variants": [
-          {
-            "key": "on",
-            "description": "The visual page",
-            "weight": 50
-          },
-          {
-            "key": "off",
-            "description": "The original page (control)",
-            "weight": 50
-          }
-        ]
-      },
+      "defaultVariant": "on",
+      "experiment": null,
       "owner": "@lorekit/web",
       "tags": [
         "dashboard",
         "web",
         "insights",
-        "rollout",
-        "experiment"
+        "rollout"
       ]
     }
   ]

packages/feature-flags/src/registry.ts

@@ -81,19 +81,12 @@ const REGISTRY_INPUT = [
   {
     key: 'insights-visual',
     description:
-      "Insights — the visual first version of `/insights`: ONE page-wide window, the Learn → Recall → Promote → Prune loop strip, the Pulse timeline (lessons written and records read on a shared time axis, chosen vs bulk), the recall flow (client → tool → scope type), and runs drawn as a weekday × hour rhythm plus swimlanes — with the existing 'earning its place' grid, runs list and operational health kept below. Gates the whole page body through the `InsightsView` resolver; with it off `/insights` renders exactly as before. Reads only routes that already ship (`/memories/activity`, `/read-activity`, `/usage`, `/usage/runs`, `/utility`), so there is no backend prerequisite and flipping it is a pure UI decision. RUNNING as a 50/50 experiment: each signed-in user is assigned an arm by a hash of their user id (the same arm on every device), a developer override still wins, and setting `enabled: false` sends everyone back to `off`. `on` is listed first so ramping its weight up never moves a user from `on` back to `off`; lowering it or reordering the arms reassigns users mid-experiment. Exposure is measured by `feature_flag.exposed`, emitted by the `InsightsView` resolver — see docs/feature-flags.md § Running an experiment.",
+      "Insights — the visual first version of `/insights`: ONE page-wide window, the Learn → Recall → Promote → Prune loop strip, the Pulse timeline (lessons written and records read on a shared time axis, chosen vs bulk), the recall flow (client → tool → scope type), and runs drawn as a weekday × hour rhythm plus swimlanes — with the existing 'earning its place' grid, runs list and operational health kept below. Gates the whole page body through the `InsightsView` resolver; with it off `/insights` renders exactly as before. Reads only routes that already ship (`/memories/activity`, `/read-activity`, `/usage`, `/usage/runs`, `/utility`), so there is no backend prerequisite and flipping it is a pure UI decision. DEFAULTS ON for every user: it ran as a 50/50 experiment (users bucketed by a hash of their user id, exposure measured by `feature_flag.exposed` from the `InsightsView` resolver — see docs/feature-flags.md § Running an experiment), and that split has ended. The flag is now a kill switch for turning the visual page back OFF, not a rollout gate; a developer override still wins. Retiring it is the copy-and-suffix removal: promote `InsightsView.on.tsx`, delete `InsightsView.off.tsx` and the `useFeatureFlagExposure` call, drop this entry.",
     type: 'boolean',
     variants: { off: false, on: true },
-    defaultVariant: 'off',
-    experiment: {
-      enabled: true,
-      variants: [
-        { key: 'on', weight: 50, description: 'The visual page' },
-        { key: 'off', weight: 50, description: 'The original page (control)' },
-      ],
-    },
+    defaultVariant: 'on',
     owner: '@lorekit/web',
-    tags: ['dashboard', 'web', 'insights', 'rollout', 'experiment'],
+    tags: ['dashboard', 'web', 'insights', 'rollout'],
   },
 ] satisfies unknown[];