Analysis action-center redesign + kernel default-on via Pages proxy

docs/Features/Face-Analysis.md

@@ -37,53 +37,51 @@ Select an analyzed video clip to see its boxes, landmarks, and anonymous person
 labels over the Preview. Existing yellow face markers remain available on the
 timeline. The timeline assigns every anonymous person a stable colour: their
 sample dots and source-time range bands share that colour. The Analysis tab
-combines source-time facts in one workspace. Its compact map aligns scenes,
+combines source-time facts in one workspace. Its collapsible compact map aligns scenes,
 cuts, speech, people, motion, focus, quality, audio, and text with the
-playhead. Below it, one virtualized scene list shows larger speaker face crops,
-word-synchronized dialogue, and scene time in compact scene blobs. Clicking a
-blob expands the scene-scoped identity tools, facts, descriptions, quality
-notices, and transcript in place.
-
-Expanded scenes keep contextual People and Needs Review controls. Clicking a
-person filters the virtualized scene list; **Next appearance** advances through
-that identity's source-time appearances. Person identities and individual
-appearance crops remain draggable correction sources. Dropping one onto another
-person merges a false split or moves that appearance. Small yellow detections
-that are unsafe to identify are consolidated into short visual tracks without
-claiming an identity; their scene-scoped review crops can be clicked to seek or
-dragged onto a confirmed person to assign the track manually. These previews
-are resolved only for the bounded virtualized window and open scene, stay in a
-bounded browser-memory cache, and are also persisted as versioned JPEGs under
+playhead. Below it, one virtualized scene list fills the remaining panel height
+and shows visible face crops with separate diarized-speaker abbreviations,
+word-synchronized dialogue, and scene time in compact scene blobs. Face IDs and
+speaker labels remain distinct until an explicit mapping exists. Scene blobs
+stay compact and never expand; clicking a word or scene time seeks directly to
+that source position.
+
+Person identities and individual appearance crops remain draggable correction
+sources inside Analysis settings. Dropping one onto another person merges a
+false split or moves that appearance. Small yellow detections that are unsafe
+to identify are consolidated into short visual tracks without claiming an
+identity; their review crops can be clicked to seek or dragged onto a confirmed
+person to assign the track manually. These previews stay in a bounded
+browser-memory cache and are also persisted as versioned JPEGs under
 `Cache/face-thumbnails` so they load without video seeks after a refresh or
 project reopen.
 
-Below the virtualized scene list, the complete detected-person strip restores
-the correction surface for source-wide work: representative found-frame crops,
-all appearances, person-to-person merge, appearance reassign, and review-track
-assignment by drag and drop. It uses the same durable face correction services
-as scene cards; no second people state is stored. Face crops are requested only
-once their tile is near the viewport. Equal crop requests share the existing
-pending/blob cache, so scrolling does not enqueue a fresh video seek for every
-rendered card.
-
-The top of the Analysis tab exposes separate action rows for **Focus & Motion**,
-**Faces**, **Scene Cuts**, **Transcript**, and **AI Scenes**. Every row can be analyzed,
-reanalyzed, retried, or cancelled without clearing the other results. A
+Analysis settings contain the complete detected-person strip for source-wide
+correction work: representative found-frame crops, all appearances,
+person-to-person merge, appearance reassign, and review-track assignment by
+drag and drop. It uses the same durable face correction services as scene
+cards; no second people state is stored. Face crops are requested only once
+their tile is near the viewport. Equal crop requests share the existing
+pending/blob cache.
+
+The top of the Analysis tab exposes compact status pills for **Focus & Motion**,
+**Faces**, **Cuts**, **Transcript**, and **AI Scenes**. Every pill can analyze,
+reanalyze, retry, continue, or cancel without clearing the other results. A
 metrics-only pass preserves compatible face observations. A face-only pass
 reuses existing focus and motion samples; when no metrics exist yet, it creates
 the inexpensive metrics baseline during the same source decode. At normal
-Properties widths the actions form one compact three-column grid (with
-responsive two- and one-column fallbacks). **Analyze All** creates a coalesced
+Properties widths the pills wrap naturally beside **Analyze all** and a compact
+settings disclosure. **Analyze All** creates a coalesced
 job graph: repeated clicks observe the same run, compatible completed channels
 are reused, Focus/Faces and the cut scan serialize on the shared source decoder,
 and independent transcript work may proceed in parallel. AI scene descriptions
 remain a separate opt-in action because they can incur provider cost and share
 visual content externally.
 
-Above those rows, compact **Scope** (`Source`, `Used Ranges`, `Selection`,
-`In-Out`) and **Profile** (`Quick`, `Balanced`, `Deep`, `Custom`) controls show
-uncached duration, cache reuse, known frame/sample work, relative cost, and a
-benchmark-derived time range when one exists. **Quick** (1 fps) and
+The settings disclosure separates shared visual-analysis controls from
+transcript controls. Compact **Scope** (`Source`, `Used Ranges`, `Selection`,
+`In-Out`) and **Profile** (`Quick`, `Balanced`, `Deep`, `Custom`) choices stay
+visible without estimate or explanatory rows. **Quick** (1 fps) and
 **Balanced** (2 fps) are execution settings for local Focus/Motion and Faces:
 the runner receives the selected clipped source range and explicit sample
 cadence. Used Ranges, Selection, and overlapping In/Out therefore analyze only
@@ -216,9 +214,9 @@ to use WebGPU for rendering.
 
 Faces smaller than 36 pixels in the 640-pixel analysis frame are intentionally
 excluded from automatic identity grouping. They remain visible as yellow
-Preview overlays and as scene-scoped **Needs review** crops in the Analysis
-workspace, because title-card grids and background footage are too small for
-dependable anonymous matching.
+Preview overlays and as **Needs review** crops in Analysis settings, because
+title-card grids and background footage are too small for dependable anonymous
+matching.
 
 ## Models and licenses
 

docs/Features/UI-Panels.md

@@ -374,31 +374,35 @@ same source-handle edges.
 | **Masks** | Mask shapes with mode and feather controls |
 | **Analysis** | Shared analysis map, transcript controls, and compact scene blobs with faces, synchronized dialogue, cuts, metrics, quality, and descriptions |
 
-The Analysis workspace uses one source-time model for its graph and scene
-list. Each virtualized scene blob keeps speaker face crops on the left, larger
+The Analysis workspace uses one source-time model for its collapsible graph and
+segment list. Visual scene boundaries remain unchanged in the graph, while the
+list independently divides long dialogue into readable speech segments. Speaker
+changes and long pauses are hard natural boundaries; sentence endings nearest
+the 10-second target are preferred, and unpunctuated speech is split at a word
+boundary before 15 seconds. This keeps talking-head footage navigable even when
+cut detection returns one long scene. The list consumes all remaining panel
+height and owns its scroll. Each virtualized segment keeps visible face crops on
+the left with a separate transcript-speaker abbreviation beside them, larger
 word-synchronized dialogue in the center, and range/duration on the right.
-The flat inspector directly beneath the graph replaces the former Current
-Frame and Summary boxes with playhead metrics and clip-wide counters. Clicking
-a blob expands its scene-scoped people, appearances, identity correction drop
-targets, review detections, camera/focus/motion facts, description provenance,
-coverage, quality notices, OCR, and transcript in place. Person chips filter
-the scene list, while **Next appearance** advances in source time. Face crops
-remain lazy because only the bounded virtualized window and the open scene
-resolve thumbnails. Contextual People and Needs Review controls stay in their
-expanded scene; the source-wide correction strip follows the list.
-
-The scene list is followed by a compact source-wide people/review strip for
-the complete correction workflow. It deliberately reuses the same person and
-review identities as scene cards: drag a person to merge, an appearance to
-move it, or a review track to assign it. Crop loading is viewport-lazy while
-the virtualized scene window and crop cache keep scrolling bounded.
-
-The Action Center above the map uses the same flat 2px control treatment as
-the rest of Properties. Its analysis actions use three compact equal-height
-cards per row at normal panel widths, with integrated status and action
-controls instead of full-width button bars. Scope and profile selectors update
-a read-only estimate before any work starts, including cached reuse and known
-frame/sample counts. A matching real-device benchmark adds a time range.
+Speaker labels are not silently equated with anonymous face identities. Segments
+remain compact and never expand. Clicking a word seeks to that exact word;
+clicking the time seeks to the speech-segment start (or the scene start when no
+speech exists). Redundant playhead and clip-summary statistics are omitted. Face
+crops remain lazy because only the bounded virtualized window resolves
+thumbnails.
+
+The source-wide people/review correction strip lives in Analysis settings so
+the normal scene list stays focused. It deliberately reuses the same person
+and review identities as scene cards: drag a person to merge, an appearance to
+move it, or a review track to assign it. Crop loading remains viewport-lazy.
+
+The Action Center above the map is a single compact row of status pills using
+the same flat control treatment as Export. Each pill exposes its channel state
+and triggers analyze, reanalyze, continue, retry, or cancel as appropriate.
+**Analyze all** and a settings button complete the row. The settings disclosure
+uses separate compact areas for shared visual-analysis controls and transcript
+controls. It keeps only control titles, scope/profile choices, language/mode,
+and actions visible; estimates and explanatory copy do not push the map down.
 Quick/Balanced Focus/Motion and Faces execute with the selected source interval
 and sampling cadence; frame-accurate cuts remain source-wide. Deep/Custom stay
 blocked without qualifying evidence. **Analyze All** deliberately excludes AI
@@ -407,10 +411,11 @@ provider cost and share visual content externally.
 
 Analysis contains the transcript workspace header for provider and language
 selection, start/continue/cancel/clear, fusion progress, coverage, and shared
-search. The same search filters the virtualized scene list. Words are grouped
-into timestamped speaker turns; clicking a compact or expanded word seeks the
-playhead to that source position. During playback, the currently spoken word
-and speaker turn are highlighted and the active scene transcript follows them.
+search. The same search filters individual rows in the virtualized segment list.
+Words are grouped into timestamped speaker turns and readable timed segments;
+clicking a word seeks the playhead to that source position. During playback, the
+currently spoken word and active speech segment are highlighted and the list
+follows them when that segment is visible under the current search filter.
 Scrubbing updates the same highlight without animated lag.
 
 Best Quality uses a fixed provider split: Deepgram supplies every displayed word

functions/api/kernel/[[path]].ts

@@ -0,0 +1,70 @@
+import { json, methodNotAllowed } from '../../lib/db';
+import type { AppContext, AppRouteHandler } from '../../lib/env';
+
+const DEFAULT_KERNEL_ORIGIN = 'https://fassandra.de';
+// Story compiles include provider planning rounds plus the director run.
+const FORWARD_TIMEOUT_MS = 240_000;
+
+interface AllowedRoute {
+  methods: string[];
+  pattern: RegExp;
+  requiresUser: boolean;
+}
+
+const ALLOWED_ROUTES: AllowedRoute[] = [
+  { methods: ['GET'], pattern: /^health$/, requiresUser: false },
+  { methods: ['POST'], pattern: /^compile$/, requiresUser: true },
+  { methods: ['POST'], pattern: /^runs\/[A-Za-z0-9._:-]+\/complete$/, requiresUser: true },
+];
+
+function resolvePath(context: AppContext): string {
+  const raw = (context.params as Record<string, unknown>).path;
+  if (Array.isArray(raw)) {
+    return raw.map(String).join('/');
+  }
+  return typeof raw === 'string' ? raw : '';
+}
+
+export const onRequest: AppRouteHandler = async (context: AppContext): Promise<Response> => {
+  const path = resolvePath(context);
+  const route = ALLOWED_ROUTES.find((candidate) => candidate.pattern.test(path));
+  if (!route) {
+    return json({ error: 'Unknown kernel route.' }, { status: 404 });
+  }
+  if (!route.methods.includes(context.request.method)) {
+    return methodNotAllowed(route.methods);
+  }
+  if (route.requiresUser && !context.data.user) {
+    return json({ error: 'Sign in to use the kernel service.' }, { status: 401 });
+  }
+
+  const token = context.env.KERNEL_AUTH_TOKEN?.trim();
+  if (!token) {
+    return json({ error: 'Kernel service is not configured.' }, { status: 503 });
+  }
+
+  const origin = context.env.KERNEL_ORIGIN?.trim().replace(/\/+$/, '') || DEFAULT_KERNEL_ORIGIN;
+  const headers: Record<string, string> = {
+    Accept: 'application/json',
+    Authorization: `Bearer ${token}`,
+  };
+  const init: RequestInit = {
+    method: context.request.method,
+    headers,
+    signal: AbortSignal.timeout(FORWARD_TIMEOUT_MS),
+  };
+  if (context.request.method !== 'GET') {
+    headers['Content-Type'] = 'application/json';
+    init.body = await context.request.text();
+  }
+
+  try {
+    const upstream = await fetch(`${origin}/kernel/${path}`, init);
+    return new Response(upstream.body, {
+      status: upstream.status,
+      headers: { 'Content-Type': 'application/json', 'Cache-Control': 'no-store' },
+    });
+  } catch {
+    return json({ error: 'Kernel service is unreachable.' }, { status: 502 });
+  }
+};

functions/lib/env.ts

@@ -54,6 +54,8 @@ export interface Env {
   ELEVENLABS_API_KEY?: string;
   GOOGLE_CLIENT_ID?: string;
   GOOGLE_CLIENT_SECRET?: string;
+  KERNEL_AUTH_TOKEN?: string;
+  KERNEL_ORIGIN?: string;
   KIEAI_API_KEY?: string;
   KIEAI_GENERATION_RATE_LIMITER?: AppDurableObjectNamespace;
   KV: AppKVNamespace;