Returns all distinct inspect_host topics detected in a user prompt.
Used by the harness to pre-run multiple topics when the user asks for several at once,
so the model receives all results and only needs to synthesize rather than orchestrate.
Returns true when the user’s query involves computation that must be exact —
checksums, financial math, statistics, date arithmetic, algorithmic verification, etc.
Used by the harness to inject a pre-turn nudge toward run_code instead of model memory.