A mapping can be confirmed without a person #504

Asked for: let exploration's proposals take effect without a person confirming them.

0011 wrote the opposite — "the agent proposes, a person decides what takes effect" — and the reason is worth restating before changing it. A wrong definition raises no error. Chat prints a number that looks entirely normal. Drop a FILTER (WHERE status = 'paid') from a GMV definition and every answer quietly includes refunds; nobody finds out until the books disagree.

The repo already has the shape for automation: auto_type_resolution (0016 C2, default on, measured at 39/41) and governance (0025, default off, with a gate, precedents and a fuse). This should copy that shape rather than invent one.

Shape

• knowledge_bases.auto_confirm_mappings, boolean, default off, beside governance. Editor turns it on in base settings.
• Ask the prompt for confidence, and keep rationale — the model already returns a rationale today and explore_mappings drops it on the floor. Store both. Below a threshold a proposal still waits for a person, so the gate has something to read and the page can answer "why is it mapped this way".
• An automatically confirmed row is recognisable by decided_by IS NULL. propose's ON CONFLICT … WHERE status = 'proposed' widens to refresh those too, so a later exploration can still improve a definition no person has touched; once someone revises or decides it, it locks. Without this the first run's guess becomes permanent, which is the opposite of the labour the switch is meant to save.
• Audit as mapping.auto_confirmed, and mark the row on the page.

Default-on needs a number first

Not before #501 produces one, and not before #502 and #503: automation multiplies whatever the explorer currently misses, and today it cannot tell a key from a quantity (#502) and cannot say what it failed to cover (#503). The number that decides it is the bench's wrong column — proposals that run and compute something other than the truth. A broken proposal is harmless; a plausible wrong one is the entire risk.

If the actual pain is the clicking rather than the reviewing, #505 answers it for a fraction of this.