Knowledge retrieval on the GitHub Copilot harness suddenly slow (~1:25 min/turn)

Since about a week we're seeing severe knowledge-retrieval latency on a Copilot Studio agent (GitHub Copilot harness, SharePoint sources) — ~1:25 min per answer. No config changes on our side. Same source on the standard harness: ~20 s.

We captured a HAR of a slow turn and found this in the SSE stream:

• knowledge_search_sharepoint completes in ~11 s (durationMs: 10774) — search is not the problem.
• But instead of snippets, the tool result says: "The full documents have been downloaded to /app/uploads/" — 9 complete files, including large PDFs. 8 of 10 results contain no snippet at all.
• It then instructs the model: "Read each relevant file IN FULL before responding — every part, from the start to the end of the file."
• The remaining ~70 s is the model reading those files before answering. More sources = more full documents = longer turns.

Oddly, the tool's own description promises the opposite: "Results are search snippets … call KnowledgeRetrieve with the ReferenceId to get the full document."

Anyone else seeing this since ~mid-August? Known issue or intended GA behavior?