Fix timeline gap frame conditioning

[Paul (nuclearspike)]

Summary

• pass the selected timeline boundary frame into the existing image-to-video generation path
• reverse the generated video and linked audio when the end frame is selected
• use the real filesystem path for replacement frames while keeping data URLs preview-only
• preserve text-to-video behavior when frame conditioning is disabled and preserve explicit source-image precedence

Root cause

The editor extracted the clips' boundary frames, but used them only for prompt suggestions and previews. The modal's Start/End selection was local UI state, and handleGapGenerate only forwarded a separately uploaded gapImageFile, so ordinary gap fills omitted imagePath and ran as text-to-video. Generated gap clips also hardcoded reversed: false, despite the UI promising that End-frame generation would be played in reverse.

Scope

This implements the behavior already documented in the UI: one conditioning frame at a time. Start conditions the beginning directly; End is used as the generation's start frame and the result is reversed.

This does not add simultaneous first-and-last-frame interpolation. The current local and cloud request contracts accept one conditioning image, so true two-boundary conditioning would require a separate backend/API capability change.

Addresses #89.

Validation

• pnpm typecheck
• pnpm build:frontend
• pnpm openapi:check
• pnpm backend:test — 567 passed
• independent control/data-flow QA for Start, End, disabled, replacement-frame, and explicit-source-image paths
• git diff --check