feat: forward ACP prompt content to Muse

[obviyus]

Problem

The adapter currently keeps only ACP text blocks. Images, resource links, audio, and embedded resources are silently discarded. A prompt can therefore appear to succeed while Muse never receives the user's attachment.

Fix

• Advertise ACP image support.
• Forward PNG, JPEG, GIF, and WebP blocks through Muse's repeatable --image flag.
• Stage image bytes in one private turn directory with 0700 directory and 0600 file permissions.
• Remove staged files after success, failure, spawn failure, or cancellation.
• Preserve baseline ACP resource links as deterministic prompt context.
• Reject audio, embedded resources, unsupported image types, invalid base64, and image-only prompts with actionable ACP errors before spawning Muse.
• Reserve the session before asynchronous staging so concurrent prompts cannot corrupt one Muse session.

The adapter does not duplicate Muse's image-size policy. Muse 1.0.1 rejects oversized files before model execution with its exact 20 MiB limit.

Live proof

• Real ACP image request → adapter → native Muse → Meta provider read CAT from the attached PNG.
• Native Muse read the same CAT image as PNG, JPEG, GIF, and WebP and returned 4.
• Native Muse rejected a 50,490,138-byte PNG with image file is too large (... > 20971520 bytes).
• Native Muse rejected an image-only headless request with missing prompt; the adapter now reports the required text before spawn.

Verification

• npm run check
• npm run build
• npm run test:run — 110 passed, 2 live-only skipped
• Focused real-provider image integration — passed
• P1 autoreview — clean
• Fable architecture and implementation reviews — clean

Dependency

This draft includes the test-only commit from #2 so the Muse 1.0 suite is green. Merge #2 first; GitHub will then reduce this PR to the content implementation.