Implement the following plan:

# Reframe Script Generation: Psychoanalyst-Director, Not Storyboard Artist

## Context

Now that the sequential chain pipeline handles visual continuity mechanically (reference frames from each completed video feed into the next scene's image generation), the system prompt is doing dead work — ~40% of its tokens teach Claude how to fake visual continuity through words. That energy should go into what Claude is actually good at: reading the unconscious material in the writing.

Anky is spiritual technology for the exponential age. Virgil in the Divine Comedy. The guide into the self. The singularity is coming — and it's not just AGI, it's the collective awakening of humanity. Anky guides you into it by guiding you into yourself. The script generation should reflect that.

## Changes

### 1. `src/pipeline/video_gen.rs` — System Prompt Rewrite

**Reframe the identity**: From "story weaver creating a cinematic video" to psychoanalytic director. Anky is Virgil — the guide through the underworld of the self.

**Kill dead continuity instructions**: All visual handoff language is gone. The pipeline handles this now.

**Deepen the psychoanalytic reading**: Claude reads the writing the way an analyst reads dreams — not for content but for structure, repetition, avoidance, breakthrough.

**Reframe the 5-act structure** through a depth/spiritual lens:
1. SURFACE → the mask, the conscious presentation
2. TENSION → what's underneath pulling, the return of the repressed
3. DESCENT → Anky appears as Virgil, guiding into the real material
4. REVELATION → the moment of seeing — not understanding, *seeing*
5. INTEGRATION → not resolution but integration, the surface returns changed

**New scene fields in the prompt**:
- `psychoanalytic_note`: Internal reading — what this scene reveals about the writer. Never sent to APIs. Forces Claude to do the deep reading before directing.
- `sound_direction`: What the scene sounds like. Grok already generates audio; we just start directing it.
- Keep `narration` but reframe as `spoken_words` or `[silence]` with a reason

**Simplified existing fields in the prompt**:
- `image_prompt`: Emotional direction, not pixel specs. "The moment before admission. A threshold. Anky watches." Let Gemini + reference frames handle the visual thread.
- `video_prompt`: Emotional movement + sound direction. "From containment to release."
- Remove `transition_from_previous` from the prompt output

### 2. `src/pipeline/video_gen.rs` — VideoScene Struct

- Add `psychoanalytic_note: String` (`#[serde(default)]`)
- Add `sound_direction: String` (`#[serde(default)]`)
- Keep `transition_from_previous` in struct with `serde(default)` for backward compat with old DB records, but stop asking for it in the prompt

### 3. `src/pipeline/video_gen.rs` — Video Prompt Composition

In `generate_sequential_chain()`, when sending video_prompt to Grok, append sound_direction:
- `"{video_prompt}. Sound: {sound_direction}"` if non-empty

### 4. `templates/video.html` — Scene Overlay & Script Panel

- Show `psychoanalytic_note` in scene overlay with distinctive styling (this is the per-scene "the system sees you" moment)
- Show `sound_direction` in scene overlay
- Remove `transition_from_previous` from overlay (graceful: only show if present for old projects)
- Show `psychoanalytic_note` in the script panel's expanded detail view