Fail interrupted groundings at boot and bound hand grounding

At boot, after the job runner fails the jobs a previous process left
running, FailInterruptedGroundings sets every dinner still pending with
no ground_dinner job queued or running to failed (interrupted). A crash
mid-job, or between saving a dinner and enqueuing its job, no longer
leaves the dinner pending forever.

A search failure on recipe-candidates or ground now logs only the reason
code search_failed, never the wrapped error, which may quote the
provider; the response stays redacted.

When the grounding job cannot be enqueued, POST /dinners answers 201
with the saved dinner failed (grounding_error) and grounding_job_id
null, instead of a 500 that hid the dinner it had created.

Grounding a candidate or link by hand runs under a 90-second deadline
(fetch, photo, normalize_recipe); past it the request answers 504
ground_timeout and the dinner is unchanged.

Maestro task: 0cbaa5bf-8be4-4cbd-aab2-891e7f1bc2b5

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

migrations/queries/dinners.sql

@@ -109,3 +109,20 @@ SET grounding_status = 'failed',
 	grounding_error = ?,
 	updated_at = ?
 WHERE id = ? AND grounding_status = 'pending';
+
+-- name: FailOrphanedPendingDinnerGroundings :execrows
+-- Fails every dinner still pending whose grounding job is no longer queued
+-- or running: the process stopped before the job, or before its handler
+-- recorded a result. Run at boot, after the job runner fails the jobs a
+-- previous process left running.
+UPDATE dinners
+SET grounding_status = 'failed',
+	grounding_error = sqlc.arg(grounding_error),
+	updated_at = sqlc.arg(updated_at)
+WHERE grounding_status = 'pending'
+	AND NOT EXISTS (
+		SELECT 1 FROM jobs
+		WHERE jobs.kind = sqlc.arg(kind)
+			AND jobs.status IN ('queued', 'running')
+			AND json_extract(jobs.payload, '$.dinner_id') = dinners.id
+	);