fix(honcho): flatten multimodal list content in sync_turn (#30252)

[Bartok9]

Problem

HonchoMemoryProvider.sync_turn errors when called with OpenAI-style multimodal list content (the shape used for vision turns):

user_content = [
    {"type": "text", "text": "what colour is this?"},
    {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
]

The call sanitize_context(user_content or "") receives a list, raising:

TypeError: expected string or bytes-like object, got 'list'

The exception is caught silently by the _sync() wrapper and the turn is silently dropped from Honcho's memory — the user representation never learns about anything visual the assistant was shown.

Closes #30252.

Fix

Add a _flatten_content() static method that normalises list-shaped content to a plain string before sanitisation:

• Plain strings pass through unchanged (zero regression risk)
• text parts are joined with a space
• image_url / image parts become the literal placeholder [image]
• Unknown block types fall back to [<type>]

sync_turn now accepts str | list | None for both arguments.

Tests

5 new tests in TestSyncTurnMultimodalContent:

• multimodal user content is recorded (text + image placeholder)
• image-only list uses [image] placeholder
• plain string passes through unchanged
• None content does not raise
• _flatten_content static method handles all known block types

All 120 existing honcho plugin tests pass.

[alt-glitch]

Competing fix with #30264, #22768, and #26484 — all add multimodal list content flattening to Honcho sync_turn. This PR adds _flatten_content() static method; #30264 and #22768 add equivalent _content_to_text() normalizers. #26484 also includes conclude fallback. Recommend consolidating into a single PR.

[Bartok9]

Closing to stay within contributor PR limit. Will resubmit with fresh rebase if the issue remains open in main.

[teknium1]

This appears to be fixed on current main by the centralized memory-sync normalization work.

Evidence from this automated hermes-sweeper review:

• run_agent.py:3103 flattens completed-turn user and assistant content with _summarize_user_message_for_log(..., sep="\n") before calling MemoryManager.sync_all, so Honcho no longer receives the raw OpenAI multimodal list shape in the reported gateway/vision-turn path.
• agent/codex_responses_adapter.py:130 documents and implements the shared flattening behavior: text parts are joined, and image_url / input_image parts become [N image(s)] markers so visual turns are still represented in memory.
• tests/agent/test_memory_provider.py:983 covers the regression invariant that multimodal list content must be flattened before reaching providers because raw lists crash regex sanitization with expected string or bytes-like object, got 'list'.
• Commit 87893fe4cb7160c45553a6eb70fae2d2f3ed8370 added the fix, and 15439bee4700dc378777846481afe9218cd0e624 refactored it to reuse the shared summarizer. Both are included in v2026.6.19.

Thanks for the focused Honcho patch and the tests. The final implementation landed at the shared memory boundary, so the same class of multimodal sync issue is handled for all memory providers rather than only Honcho.