fix(vision): accept pixel-correct resize when bytes grow (#48013)

[kyssta-exe]

Fixes #48013. Removed the byte-size comparison that discarded valid dimension-reduced resizes of PNG images. The _resize_image_for_vision function already handles dimension reduction; the caller should trust its output.

[alt-glitch]

Duplicate of #48035 — both PRs fix #48013 in agent/conversation_compression.py by relaxing the same byte-growth rejection in _shrink_data_url (len(resized) >= len(url)) so a pixel-correct downscale is accepted even when the re-encoded PNG grows in bytes. #48035 (filed earlier, ~20:33Z) is the more complete version of the same approach — it tracks which constraint triggered the shrink and adds a regression test — so it's the canonical one. Marking this as duplicate.

[teknium1]

Merged via #49140 (commit 990273d) — closes #48013. This was a duplicate fix for the same image-shrink brick. The merged PR keys the accept gate on the binding constraint (pixels vs bytes) against current main; the one-line change here predated the max_dimension plumbing from #45979 and would have reverted that cap parsing. Thanks for flagging the issue!