#3652 [Engine] Withdrawn: unsubstantiated image_decode gate regression in serve 0.2.2

Descripción

## Correction and withdrawal

The original report's claimed new `image_decode` compatibility regression is not supported by the source. I am withdrawing this issue rather than asking maintainers to implement the proposed fix. Sorry for the inaccurate report.

Review of [the published v0.2.2 server](https://github.com/Niko1221/Strata/blob/v0.2.2/serve/server.py#L4606-L4638) and the local checkout at `011c9e600a93bbc9fd451c563501b5831c32bbca` found:

- `_note_unreadable_tool_images()` calls `self.vision.encode(src)` directly. The claimed `vision_engine_supports_image_decode()` / `_rpc("image_decode_supported")` gate is not present in that server. The quoted traceback pointing to that gate is not verified evidence and is withdrawn.
- The proposed exception handler did not even catch the `TypeError` quoted in the original report.
- #3174 is Lithivm's PR, not ours. At the time of this correction it is open, not closed as superseded. It adds `_vision_up()` before encoding a tool-returned image; it does not widen the exception handler.

The actual elastic-encoder problem is already described in #3171, with a proposed fix in #3174. Our local server has a change adding `_vision_up()` before the tool-image encode and updating `last_image_at`. That does not establish the separate old-engine regression alleged here.

**Environment of the original report:** Windows 11 with WSL2 interop, RTX 5090 32 GB VRAM, 96 GB RAM; serve v0.2.2 with engine 0.1.44.1 during the update window; Qwen3.8-Flash-Next UD Q4_K_XL, 256K context, elastic vision enabled.

**Verification status:** source and local diff reviewed; no fresh end-to-end reproduction was run for this correction. There is no verified reproduction of the claimed new gate failure. This closure withdraws that claim; it does not mean the existing bug in #3171 has been fixed.