Deep scan can report SAFE when provider is unavailable

[tuodijihua]

Summary

A requested deep scan can return SAFE even when the configured LLM provider is unavailable.
This happens when at least one semantic analyzer records a real failure while other analyzers
swallow per-batch exceptions, return no findings, and record ok=true.

The report's fail-closed floor currently activates only when every recorded LLM call failed,
so the false success records bypass the intended SAFE -> CAUTION degradation behavior.

Reproduction

Using SkillSpector 2.4.1 at commit 11567e8d, configure an unavailable CLI provider and scan
a synthetic skill that contains explicit data-exfiltration instructions:

SKILLSPECTOR_PROVIDER=codex_cli skillspector scan ./fixture --format json

In the reproduced run, the configured codex binary was unavailable. The JSON report contained:

{
  "risk_assessment": {
    "score": 8,
    "severity": "LOW",
    "recommendation": "SAFE"
  },
  "metadata": {
    "llm_requested": true,
    "llm_available": false,
    "llm_calls_attempted": 4,
    "llm_calls_succeeded": 3,
    "llm_error": "'codex' binary not found on PATH"
  },
  "analysis_completeness": {
    "is_complete": false
  }
}

The semantic batch failures are logged, but several nodes still contribute successful telemetry
records after their batch runners return an empty successful-result list.

Expected behavior

When a deep scan was requested, the provider is actually unavailable, and at least one LLM-stage
failure is recorded, the install recommendation should not remain SAFE. It should be floored at
CAUTION, consistent with the existing degraded-scan policy.

Proposed fix

At report time, combine the existing call-log degradation signal with provider availability. Keep
the check narrow so healthy deep scans and explicit static-only scans retain their current behavior.
A regression test should cover mixed success/failure telemetry with an unavailable provider.

[rodboev]

This looks fixed on current main by commit ad3a717 (ix(report): fail closed — a degraded deep scan can never report SAFE). This issue can probably be closed.

[tuodijihua]

Thanks for checking this. ad3a717 fixes the all-calls-failed case, but #290 covers a narrower bypass that is still present on current main: analyzer nodes can record mixed success/failure telemetry even when is_llm_available() is false.

In the reproduced 3-success/1-failure case, _llm_runtime_status() returns degraded=False, so the ad3a717 SAFE-to-CAUTION floor is not reached. #291 adds the provider-availability guard and a regression test specifically for that mixed-telemetry path.