v0.8.0 — Inline rendering + auto-trigger + auto-poll

Three UX fixes shipping together:

Inline rendering for oversized outputs

• Adds sharp (^0.34.5) for image recompression
• Lowers INLINE_MAX_BYTES from 8 MB → 4 MB (Anthropic API caps individual image attachments at 5 MB)
• fetchAsContentBlock now recompresses oversized images to JPG with progressive quality (85 → 55) and dimension fallback (scale 1 → 0.55) so 4K PNG generations always render inline instead of falling back to a URL

Auto-trigger via MCP-level instructions

• The MCP Server constructor now passes an instructions string surfaced on every session init
• Lists trigger keywords (FR + EN) + tool routing, so Claude reaches for Keou tools on any media-gen request without the user having to mention "Keou"

Auto-poll without user prompts

• Tightens descriptions on keou_generate_image, keou_generate_video, keou_text_to_speech, keou_generate_sfx, keou_get_status to make explicit that the assistant must poll in a tight loop without ending its turn or asking the user between polls

No breaking API changes. Bump 0.7.0 → 0.8.0.

Test e2e validated: 4K PNG (~28 MB) → fetch → recompress JPG q85 → 1.35 MB inline content block in ~700 ms.