Add evidence-based task completion gate

[Y1fe1Zh0u]

Outcome

• replaces the product-style Agent base prompt with a concise understand, execute, verify, finish contract
• keeps Memory, Workspace, Focus, Trigger, and Directory/A2A discoverable in the stable prompt
• runs deterministic Tool ledger verification before an independent semantic task-completion check
• returns explicit completion gaps to the Agent for up to 10 repair rounds
• fails open when the semantic judge times out or emits invalid output
• delivers the latest normal model answer when semantic repair capacity is exhausted, while recording internal diagnostics

Safety boundary

• Tool ledger and reference integrity remain fail-closed
• only an explicit semantic repair decision re-enters the Agent loop
• semantic judge infrastructure errors do not suppress normal output

Verification

• 78 scoped backend tests passed
• scoped Ruff passed
• git diff check passed

Not tested

• live provider completion-gate call
• 3010 end-to-end Run