The practical version of this question arrives when you take over a site someone else has been optimising and need to decide what you have just inherited.
Audit before you commit
Before signing, check the backlink profile for network patterns, the site for doorway pages, and the content for scaled generation. Discovering a manual action in month two, having promised results in month three, is a position with no good exits.
Separate penalty risk from devaluation risk
A manual action is visible in Search Console and recoverable through a defined process. Algorithmic devaluation is invisible, has no reconsideration process, and is far more common. Most inherited problems are the second kind, which means there is nothing to appeal — only work to redo.
Cleanup competes with growth for the same budget
Every hour spent disavowing links or removing thin pages is an hour not spent earning new visibility. Be explicit with whoever is paying that the first quarter may show no growth because it is being spent undoing damage. That conversation goes far better before the quarter than during it.
Where the line actually sits
The test that holds up is intent: are you making the page genuinely better for the person who lands on it, or making it appear better to a crawler? Most grey-area tactics resolve cleanly once framed that way, and it is a defensible position to state to a client.