The adjudicator looks before it asks

[WaylandYang]

Closes #358. Stacked on #474 (0027) and #473 (0026); rebases onto dev as those merge.

Record: new 0028 · The adjudicator looks before it asks; the status lines of 0025 and the open questions of 0026 now point at it (and at 0027).

What the issue asked, and what this does

#358 asked for an adjudicator that can go and look, and said to build it only once a number said which kind of miss the adjudicator makes. The record is honest about that number: 0025 decision 10 measured the governed path (96.9% agreement, four wrong merges, eleven left for people on 589 pairs) and found the remaining misses are judgment and precedent-following, not evidence nobody read; this PR does not claim the loop raises agreement on that corpus. What it fixes is narrower and cheap because the loop already exists:

• With governance off there was no second look at all. The batch adjudicator now sends the pairs it cannot settle (no verdict, or below AUTO_CONF), that a hard rule would not hold anyway (type families, version tails, phrases) and that carry no revert, through the governor's loop via look_again — same tools, same LOOP_DAILY_CALLS. The loop's verdict replaces the batch's in the cache and lands through the same apply_verdict (0026's sample and 0027's gate still apply), with via: investigated in the ledger. A deferred pair is escalated as proposed, so its card carries the chip and the question.
• Every look is a row in agent_decisions, applied or proposed, with trace, calls and question — so the daily budget counts it and the Agent queue shows it whether or not governance is on. The switch turns the governor on, not the queue.
• The loop can see the boundary before it decides. New consequences tool (no arguments): Impact::describe() from 0027's gate — one-value relations where the sides hold different values, derived facts resting on either side, chat answers that named either side — plus whether the two types share a family. The prompt says a merge that would touch anything outside the graph is held whatever the confidence, so prefer to defer with the question that would settle it.
• Nothing else moves: the batch prompt, the confident verdicts, the governor's own rule and both bars are unchanged.

Verified

• cargo check, clippy and fmt clean on extract, store and server; extract's governor unit tests (tool list, step reader) updated and green; Impact::describe unit-tested; the four store DB tests (0026 rationale, ledger precedents, batch decisions, 0027 gate) green on the real database.
• The loop itself needs a model; not exercised end to end here. The number that measures it is the one 0026's sample collects, split by via.

🤖 Generated with Claude Code