v5.17.1 - ClawScan follow-up

The ClawHub LLM review of 5.16.2 found two real issues in the bundled Python scripts. The npm package's TypeScript engine was not affected.

ClawScan follow-up (ClawHub Python skill)

The ClawHub LLM review of 5.16.2 found two real issues in the bundled Python scripts. The npm package's TypeScript engine was not affected.

• Low-confidence results no longer pass review (swarm_guard.py):
• A result below the 0.7 confidence threshold is now usable: false, gets a REVIEW REQUIRED recommendation, and exits with code 1.
• The warning is stored with the validation, and supervisor-review returns BLOCKED until that agent's latest result meets the threshold.
• Non-numeric or out-of-range confidence values are rejected.
• Blackboard commits re-check for conflicts (blackboard.py):
• Propose, validate, commit, and abort each run as one locked operation.
• Commit re-checks that the key hasn't changed and refuses a stale change without writing anything, and a change can't be committed twice.
• Reloads replace the in-memory cache, so deleted keys no longer reappear, and reads never rewrite the file.
• The blackboard file is replaced atomically, and the "all-or-nothing" claim was removed from the docs.

Regression checks cover each finding: 20 checks, 13 of which fail on 5.17.0.

Includes everything in v5.17.0: the opt-in Perpetual Runtime.