feat(findings): grouped duplicate suggestions, ranked candidates, relaxed merged acceptance (Codex rewrite)

Authored entirely by Codex CLI (codex exec, reasoning effort high), one session
resumed across capacity fallbacks:
- gpt-6.1-sol (schema + initial db work, stopped at capacity)
- gpt-6-astra (db/server/agent/frontend, stopped at capacity)
- gpt-6-sol (remaining items, tests, fixes; stopped at capacity after final checks)

Schemes 1+2+3: prompt/acceptance (2..20, evidenced differences justify
independent, merged no longer blocked by unified report), group suggestions with
member table, superset reuse/subset superseded, veto-aware reject/exclusion,
anchor re-anchoring trigger, sweep after merge/reject/remove, frontend group card
with stale-response guard, candidate scoring with tier/hit_reasons/strong_total.

web/src/lib/api.ts

@@ -645,8 +645,11 @@ export const api = {
     get<FindingCaseSuggestion[]>(
       `/exploration/finding-case-suggestions${task ? `?task_id=${encodeURIComponent(task)}` : ""}`,
     ),
-  resolveFindingCaseSuggestion: (id: string, accept: boolean) =>
-    post<{ case_id: string }>(`/exploration/finding-case-suggestions/${id}`, { accept }),
+  resolveFindingCaseSuggestion: (id: string, accept: boolean, excludeIds?: string[]) =>
+    post<{ case_id: string }>(`/exploration/finding-case-suggestions/${id}`, {
+      accept,
+      ...(excludeIds && excludeIds.length > 0 ? { exclude_ids: excludeIds } : {}),
+    }),
   reviewFindingCases: (ids: string[]) =>
     post<{ runs: { conversation_id: number; task_id: string }[] }>("/exploration/finding-case-review", {
       finding_ids: ids,