A Pathfinder outage is indistinguishable from an empty result set

Found by the review of PR #242 (Phase 2, code search).

PathfinderClient.search() collapses every failure to [] with a console.error, so the pipeline cannot tell "the code index is down" from "there are no code hits for this question". Both arrive as an empty array.

The consequence is the silent-no-retrieval mode that #242 exists to remove, one level up: during a code-index outage the answer is built from docs alone, at full confidence, with nothing marking it degraded. Nobody finds out except by reading worker logs.

This matters more now than it did with one retrieval source, because with two the failure is partial and therefore invisible — a docs-only answer still looks like a normal answer.

Shape

Returning { results, degraded } from the search path, or otherwise surfacing the failure to the confidence signal, would let the pipeline mark an answer as degraded when a source it expected was unavailable. ConfidenceAssessment already carries a degraded flag for the scorer-unavailable case, so there is a precedent to follow.

Related: #231 covers the same class of problem in the generator (an error and a guard trip landing on identical output), and #197 is why console.error is the only signal today.

Start by tracing PathfinderClient.search() through the retrieval pipeline and compare its failure handling with the degraded flag on ConfidenceAssessment. Follow the precedent in the scorer-unavailable path, then verify that a Pathfinder outage is distinguishable from an empty result set and causes the resulting answer to be marked degraded.