fix(slack): stop inlining PDFs as DataContent on vision-capable models

[Aaronontheweb]

Summary

Fixes the urgent production incident where Slack PDF uploads to ArdyBot fail with LLM provider error (500): Invalid url format: data:application/pdf;base64.

Root cause: SlackThreadBindingActor was computing inlinePdfs = InputModalities.HasFlag(ModelModality.Image) — conflating "vision-capable" with "accepts inline PDFs". A PDF dropped into a Slack DM against a llama.cpp / vLLM / Ollama endpoint via OpenAiCompatibleChatClient got wrapped as DataContent(application/pdf), which ToMessage unconditionally serialized as an image_url content part with a data:application/pdf;base64,... URL. No OpenAI-compatible upstream accepts that.

Fix: PDFs are now unconditionally path-only on all providers. The file still lands in {sessionDir}/inbox/ and the announcement line already points the agent at AttachmentNotes.ModelMissingPdf — "current model has no native PDF support; use shell_execute (e.g., pdftotext) to extract text". The agent reads the file via tool calls instead of the runtime shoving bytes through a wire format the endpoint doesn't support.

Images are unchanged: vision-capable models still inline as image_url, because that's the only way they can be "seen" on OpenAI-compatible wire formats.

Reproduces on session D0AC6CKBK5K/1776091017.523089 (2026-04-13).

Changes

• src/Netclaw.Channels.Slack/SlackThreadBindingActor.cs — delete inlinePdfs from live ingress and historical backfill paths; simplify ResolveInlineDecision to two parameters.
• src/Netclaw.Actors.Tests/Channels/SlackAttachmentIngressTests.cs — flip Pdf_in_dm_on_vision_capable_model_* into a regression lock asserting no DataContent(application/pdf) ever reaches the chat client.

Three prior PRs (#601, #607, #626) plumbed PDFs through ingress → scanner → inbox → LLM input assembly, which is what unmasked this far-end gating bug. None of them touched the modality gate or the provider client.

Out of scope (deliberate)

• Enabling native inline PDFs on providers that actually support them (Anthropic document blocks via DataContent through the Anthropic SDK, OpenAI file content parts) — those go through different client paths and need their own plumbing. Separate PR if/when someone wants that feature.
• Revisiting whether Qwen3.5-27B-UD-Q4_K_XL.gguf's detected modalities are accurate. Orthogonal.

Test plan

• dotnet build clean
• dotnet test — full suite, 2500+ tests pass including the flipped Pdf_in_dm_is_saved_to_inbox_path_only_and_never_inlined regression
• dotnet slopwatch analyze — 0 issues
• Manual end-to-end: rebuild netclawd, re-send the same PDF to ArdyBot in Slack. Expected: announcement line arrives with inlined="false" and the "current model has no native PDF support" note; agent runs shell_execute pdftotext ... against inbox/Logo Design Services.pdf and answers the user. No 500.