#3257 [GUI] Desktop install: MTP download shows no progress ("0.0 / 0.0 GB"), and cancelling leaves mtp_fetch.py running as an orphan

Descrição

> **Note:** This issue was identified and analyzed with AI assistance (Claude). The root-cause observations below were reviewed by the reporter, but the AI-assisted analysis may contain inaccuracies, so please verify against logs before acting on it.

**Versions:** Strata Desktop 0.2.0-beta, Control/Engine 0.1.45.3 (windows-x86_64-cuda13), Windows 11 26300, RTX 3060 12 GB
**Model:** Swift 1.5 IQ2_XS, `--data-dir G:\StrataLLM`

**Steps**
1. Install Swift 1.5 IQ2_XS from the Desktop app. Model download and iq_pack finish (`[ok] model prepared`).
2. Step 6 runs `tools/mtp_fetch.py fetch`. The app shows "Preparing your model for Strata… IQ2_XS: 0.0 / 0.0 GB" with no progress for the ~5 GB MTP download.
3. Assuming it is stuck, cancel (or close the app), then retry.

**Actual**
- `native-jobs.json` reports the model part as `done_gb:0,total_gb:0` for the whole MTP download.
- After cancel, setup.py exits 1 (`strata-mcp: setup ended with exit code 1`), but its child `mtp_fetch.py` keeps writing. setup.log shows progress lines after the exit line, and `mtp-manifest.json` was written 4 minutes after the job was marked failed.
- On retry, a new mtp_fetch appends to the same tensor file while the orphan is still running. The orphan then dies with
  `OSError: [Errno 9] Bad file descriptor` at `mtp_fetch.py:285/288` (`open(path, "ab")` / `f.write`).
- The job never reaches step 7, so no `strata-*.json` run config is written and Settings → Storage shows "No downloaded models" even though all GGUF shards are verified.

**Expected**
- MTP download progress is shown (it already prints per-tensor % to stderr).
- Cancel ends the whole process tree (`proc_kill_tree` uses `taskkill /T /F`, yet the grandchild survived here), or mtp_fetch holds a lock so two instances cannot append to the same file.
- Storage indicates "downloaded, preparation incomplete" instead of "No downloaded models".

**Workaround:** Run the install again without cancelling. Shards (`.done`) and the pack are skipped, and the MTP tensors pass `mtp_fetch.py verify`.