Harden model store and guard ML init

[chiruu12]

Summary

• Move ML spec resolution/download inside Guard's degradation try/except so download failures honor require_ml and avoid double resolve
• Harden ModelStore: corrupt manifest handling, weight-file validation, atomic manifest writes and download swaps, fixed list_status stale-revision reporting, UNPLUG_MODEL_PATH warnings, filtered HF snapshot downloads
• Raise ConfigError for unknown active_model tiers; improve unplug-models download error messages (ImportError vs hub/network)

Test plan

• make check (ruff, mypy, pytest)
• make test-cov (82.08% total, gate 80%)
• Regression tests for all audit findings (F1, F3, F4, F5, F6, F9, F11, F12, F14)

---

Note

Medium Risk
Changes affect ML loading, cache upgrades, and config validation at Guard init; behavior is stricter (unknown tiers fail fast) but failures are more recoverable during downloads.

Overview
Hardens the local ML model cache and tightens how Guard loads active_model.

ModelStore now tolerates corrupt manifests (warn + treat as missing), requires weight files for a valid checkpoint, writes manifests atomically, and downloads into a temp dir with backup/rollback so failed upgrades do not remove a working install. list_status treats stale catalog revisions as still installed while flagging upgrade available; invalid UNPLUG_MODEL_PATH is logged and ignored. Hugging Face snapshots are limited to safetensors/tokenizer patterns.

Unknown active_model tier names raise ConfigError with the valid catalog list instead of silently skipping ML. Guard resolves and downloads the model in one try/except path (reusing the resolved spec for the provider), propagates ConfigError, and respects require_ml on download/load failures. unplug-models download separates ImportError (install [ml]) from hub/network errors with a HF hint.

Regression tests cover manifest corruption, missing weights, stale revision reporting, download rollback, env path warnings, CLI errors, and Guard degradation vs require_ml.

^{Reviewed by Cursor Bugbot for commit 25580e2. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

SDK Coverage •

File	Stmts	Miss	Cover	Missing
src/unplug
guard.py	343	28	91%	127, 194–195, 197, 213–214, 288, 301, 304, 309, 417, 449, 451, 453, 490, 509, 512, 517, 552, 599, 684–685, 687–690, 696, 715
src/unplug/cli
models.py	80	6	92%	22–25, 72, 120
src/unplug/core/runtime
model_runtime.py	96	29	69%	30, 40, 42, 44, 53, 61, 67–70, 73, 78, 85, 101, 103, 118–126, 131, 133–136
src/unplug/ml
store.py	157	14	91%	42, 74, 111, 121, 124, 157, 162–163, 173, 186, 190, 201, 262, 266
TOTAL	7087	1094	84%

Tests	Skipped	Failures	Errors	Time
1128	25 💤	0 ❌	0 🔥	35.350s ⏱️

[greptile-apps]

Greptile Summary

This PR hardens model download, cache, and Guard ML initialization behavior. The main changes are:

• Corrupt model manifests are ignored with warnings.
• Checkpoints now require config and weight files.
• Model downloads use temporary directories, atomic manifest writes, and rollback on failed swaps.
• Guard ML setup now degrades or fails based on require_ml after download/load errors.
• CLI download errors distinguish missing ML extras from Hugging Face or network failures.

Confidence Score: 4/5

This is close, but the model-store fixes should be tightened before merging.

• Valid sharded safetensors downloads can be rejected after the files are fetched.
• Concurrent downloads for the same tier can still leave the manifest and checkpoint out of sync.

sdk/src/unplug/ml/store.py

Important Files Changed

Filename	Overview
sdk/src/unplug/ml/store.py	Adds stricter validation, filtered downloads, corrupt manifest handling, atomic manifest writes, and rollback, but the validation and concurrent swap behavior still need fixes.
sdk/src/unplug/guard.py	Moves model preparation and loading under the Guard degradation path while keeping configuration errors fatal.
sdk/src/unplug/core/runtime/model_runtime.py	Adds explicit handling for unknown active model tiers and lets Guard reuse the resolved model spec.
sdk/src/unplug/cli/models.py	Improves download error messages for missing ML extras and Hugging Face failures.

_{Reviews (2): Last reviewed commit: "Restore backup on failed checkpoint swap" | Re-trigger Greptile}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 25580e2. Configure here.}