fix(sec-core): degrade prompt scan to L1 when ML model not downloaded

[jfeng18]

Summary

Fixes prompt injection scanning being non-functional on any host where the ML model was never downloaded (the default after install). Previously every prompt_scan errored with ModelLoadError instead of degrading to L1.

Fixes #790.

Root Cause

MLClassifier inherited is_available()->True, so the scanner treated L2 (ML) as mandatory-and-available even when the Llama-Prompt-Guard-2 model was absent. At scan time detect() raised ModelLoadError, and because ml_classifier was not in _OPTIONAL_DETECTORS, the whole scan errored — no fallback to L1 (regex).

(torch/transformers ARE installed; only the model files were missing. Distinct from #680, which changes the cosh hook's fail-open→fail-ask behavior but not this model-availability gap.)

Changes

File	Change
detectors/ml_classifier.py	Add is_available() override: probe torch/transformers importability + model presence
models/model_manager.py	New is_model_downloaded() predicate (reused by _resolve_local_model_path)
scanner.py	Add ml_classifier to _OPTIONAL_DETECTORS → skip to L1 when unavailable
scanner.py	warmup() bypasses the is_available() gate (builds detectors from config) so it can download a currently-unavailable model

Coupling note: the is_available() override and the _OPTIONAL_DETECTORS change must land together — changing only the former makes a False detector raise LayerNotAvailableError in the constructor (eager crash) instead of degrading.

What's NOT changed

• cosh hook / prompt_scanner_hook.py — that's fix(sec-core): prompt scanner fail-ask on error #680's domain
• Daemon preload (feat(sec-core): route prompt scan to daemon and add prompt model preload #786) — separate auto-warmup path
• No install-time auto-warmup added (out of scope; tracked separately)
• Pre-existing isort disorder in model_manager.py (pydantic import) left untouched — exists on main, not in scope

Verification (ECS, kernel 6.6.102+, Python 3.11)

Scenario	Before	After
Model absent, --mode standard	verdict=error, ModelLoadError	L1 fallback, warns + skips L2, no error
scan-prompt warmup	silently no-op (chicken-and-egg)	downloads model (1.04G)
Model present, injection text	n/a	DENY, jailbreak, 99.9% confidence
Model present, benign text	n/a	PASS
Unit tests	—	287 passed
black 26.3.1 (CI version)	—	clean

Discriminating signal

# Before (ml_classifier mandatory): verdict=error "Detector 'ml_classifier' is not available"
# After: PASS via L1, log "Detector 'ml_classifier' ... will be skipped"

[jfeng18]

Hi @haosanzi, this PR degrades to L1 (regex) scanning when the ML model is unavailable, instead of skipping the scan entirely. As the prompt scan module owner, could you take a look when you have time?