fix(#2866): persist LlmMessage.images / ImagePart in file-backed snap…

[Skobeltsyn]

…shot

Pre-#2866 SnapshotJson.encodeMessage serialised only role / content / toolCalls per message. Any vision attachments on the conversation (LlmMessage.images: List<ImagePart>? from #2470) were dropped on encode and decodeMessage always rehydrated images = null. File-backed snapshot/resume silently lost image context; SSE-style apps that checkpointed mid-conversation could not resume with the original images attached.

Fix in core/Snapshot.kt:

• encodeMessage now writes an images array when non-null, with per-part {base64, mime}. Wire shape stays byte-identical for messages without images (the images key is omitted), so pre-#2866 snapshots and current image-less calls produce the same JSON.
• decodeMessage rehydrates each ImagePart via a decodeWireMime helper that maps the four shipped MIMEs (PNG / JPEG / GIF / WEBP) back to the closed WireMime sealed type. Unknown MIMEs are skipped defensively (a hand-edited or future-extended snapshot doesn't crash resume — the unknown part drops, the rest of the message rehydrates).

Encoding choice: base64 directly inside the message rather than ContentRef.hash + BlobStore re-fetch on resume. Self-contained snapshots — no BlobStore dependency at restore time — match the audit story better. Cost is the snapshot file size (~5KB per image vs. ~80 bytes for a ref), accepted as the right trade for the 0.6.x line.

Regression coverage in new SnapshotImagesRoundTripTest:

• single image round-trips byte-for-byte
• mixed-MIME ordered list round-trips
• no-image message back-compat (the images key is omitted)
• legacy pre-#2866 snapshot decodes with null images
• unknown-MIME entries skip on decode (defensive)

Full ./gradlew test + detekt green. detekt baseline regenerated to include the new test file's package-naming entry (same as the other agents_engine.* packages).