v1.0.2 crashes (0xC0000005) loading qwen3.6-moe:35b-a3b weights downloaded with v0.9.45 — `flm pull` doesn't detect outdated weights

Environment

• FLM: upgraded v0.9.46 → v1.0.2 (Windows, fastflowlm_1.0.2_windows_amd64.zip)
• Model: qwen3.6-moe:35b-a3b (weights downloaded under v0.9.45/0.9.46, model dir version metadata says 0.9.45)
• Hardware: AMD Ryzen AI NPU (XDNA 2)

Steps to reproduce

1. Upgrade FLM v0.9.46 → v1.0.2 (zip overwrite install)
2. flm serve qwen3.6-moe:35b-a3b -c 65536

Actual behavior

On startup, FLM detects config.json as corrupted/outdated, removes it, and re-downloads the v1.0.2 config successfully:

[WARNING]  Local model qwen3.6-moe:35b-a3b version: 0.9.45 < 1.0.2
[FLM]  Checking file: config.json...
[FLM]  Fail!
[FLM]  Removing corrupted file: config.json...
[FLM]  Downloading 1/1: config.json
[FLM]  All files verified successfully.
[FLM]  Loading model: C:\Users\Administrator\.flm\models\Qwen3.6-35B-A3B-NPU2

Then the process crashes with 0xC0000005 (Access Violation) while loading the model — the remaining weight files (model.q4nx etc., from the 0.9.45 era) pass file verification but are apparently incompatible with the v1.0.2 runtime.

flm pull qwen3.6-moe:35b-a3b reports "Model already downloaded. Use --force to re-download." — the pull completeness check does not catch the weight/runtime version mismatch. flm pull --force (full 22.6 GB re-download) is the only fix.

Expected behavior

Either:

• flm pull should detect that existing weights were produced for an older runtime version and re-download them (or at least warn), or
• flm serve should fail gracefully with a clear "model files outdated for runtime vX, please run flm pull --force" message instead of crashing with 0xC0000005.

Workaround

flm pull --force qwen3.6-moe:35b-a3b (full re-download).

Notes

The proxy env injection (HTTPS_PROXY) works fine for the built-in downloader — thanks for adding standard curl env support.

Start by reproducing the upgrade path with flm serve qwen3.6-moe:35b-a3b -c 65536 and compare it with flm pull qwen3.6-moe:35b-a3b. Trace the version metadata and completeness checks for config.json and model.q4nx, then inspect the model-loading entry point. Done means outdated weights are detected or the load fails with a clear recovery message instead of 0xC0000005.