Reconcile UI with v2 backend payload + drop redundant signal cards

The Swap backend data merge (fe99ffa) routed the app through the new
view-data layer where corpus-aggregates and hierarchy.json have
partially renamed shapes. Three concrete regressions from that merge
plus one duplicate-card cleanup are addressed here.

components/signals/corpus-signals-strip.tsx
- Crash fix. The strip was throwing
`TypeError: Cannot read properties of undefined (reading
'toLocaleString')` on the home / corpus routes because the new
CompletenessCorpusBlock and ComparabilityCorpusBlock no longer
expose `total_benchmarks`, `completeness_score_mean`,
`completeness_score_median`, `variant_divergence_rate`,
`cross_party_eligible_groups`, `variant_eligible_groups`, or
`cross_party_divergence_rate`. The renamed fields (total_triples,
completeness_avg/min/max, variant_divergent_count,
groups_with_variant_check, groups_with_cross_party_check, etc.)
are already in lib/backend-artifacts.ts post-merge but the strip
still read the old names.
- Each block read now goes through `aggregates.{block}?.overall`
with optional chaining on every nested field. When a value is
absent the tile shows a non-numeric headline ("mean across
reported score triples." instead of "mean across X reported …")
rather than crashing the route.
- source_type_distribution shares (tpShare/fpShare) only compute
when both totalReports > 0 and the corresponding bucket exists.

components/eval-detail.tsx
- Drops the standalone rounded shadcn-style <CompletenessPanel>
and <ComparabilityPanel> cards from the technical-details
accordion. The compact BenchmarkSignalsStrip introduced in
32864b0 already conveys both signals in the paper-aligned 4-tile
framing; the standalone cards were a duplicate that visually
re-introduced the old design language ("Comparability — Groups
where reported scores diverge…").
- Removes the now-unused imports (CompletenessPanel,
ComparabilityPanel) and the dead `benchmarkComparability` local.

components/family-table.tsx
- Hardens collectLeafEntries against hierarchy snapshots that ship
leaves without an explicit `eval_summary_ids` array (the v2
snapshot doesn't always populate it). When that field is empty
we now synthesise the id as `${family_key}_${leaf_key}` (the
pipeline's standard naming) and fall back to the bare leaf key.
Without this guard, every leaf was silently dropped and the
inline "Benchmarks in this family · N" grid disappeared
completely on rows like Helm classic / LLM-Stats.

components/eval-detail.tsx

@@ -4,8 +4,6 @@ import { useAudienceMode } from "@/components/audience-mode-provider"
 import { Fragment, useEffect, useMemo, useState } from "react"
 import Link from "next/link"
 import { BenchmarkSignalsStrip } from "@/components/signals/benchmark-signals-strip"
-import { CompletenessPanel } from "@/components/signals/completeness-panel"
-import { ComparabilityPanel } from "@/components/signals/comparability-panel"
 import { SignalsRowBadges } from "@/components/signals/signals-row-badges"
 import { getCompletenessPopulatedCount } from "@/components/signals/signal-utils"
 import { Collapsible, CollapsibleContent, CollapsibleTrigger } from "@/components/ui/collapsible"
@@ -551,7 +549,6 @@ export function EvalDetail({ summary }: EvalDetailProps) {
       ? "Averaged model results across the suite's component benchmarks, with drill-down to each component score."
       : "Model results with benchmark context, source dataset detail, and optional instance-data links."
   const reportingCompleteness = summary.evalcards?.annotations?.reporting_completeness
-  const benchmarkComparability = summary.evalcards?.annotations?.benchmark_comparability
   const documentationPopulatedCount = reportingCompleteness
     ? getCompletenessPopulatedCount(reportingCompleteness)
     : null
@@ -732,11 +729,10 @@ export function EvalDetail({ summary }: EvalDetailProps) {
               </dl>
             </div>
 
-            <CompletenessPanel completeness={reportingCompleteness} />
-            <ComparabilityPanel
-              comparability={benchmarkComparability}
-              summary={summary.comparability_summary}
-            />
+            {/* The compact BenchmarkSignalsStrip above already covers
+             *  completeness and comparability with paper-aligned framing,
+             *  so the standalone rounded shadcn cards that used to live
+             *  here are intentionally dropped. */}
 
             {!hasMultiMetricLeaderboard && (summary.root_metrics?.length || summary.subtasks?.length) ? (
               <section

components/family-table.tsx

@@ -73,7 +73,19 @@ function collectLeafEntries(
 ): LeafEntry[] {
   const out: LeafEntry[] = []
   for (const leaf of fam.leaves ?? []) {
-    const ids = leaf.eval_summary_ids ?? []
+    // Backends differ in whether leaves carry an explicit
+    // `eval_summary_ids` array. When absent, fall back to the
+    // pipeline's standard `${fam.key}_${leaf.key}` naming, then to the
+    // bare leaf key — both are stable enough for the detail page to
+    // resolve. This stops the inline benchmarks grid from disappearing
+    // on a backend that ships hierarchy.json without leaf eval ids.
+    const explicit = leaf.eval_summary_ids ?? []
+    const ids =
+      explicit.length > 0
+        ? explicit
+        : leaf.key
+        ? [`${fam.key}_${leaf.key}`, leaf.key]
+        : []
     if (ids.length === 0) continue
     // Domain sources, in order of trust:
     //   (1) hierarchy `leaf.tags.domains` — sometimes absent

components/signals/corpus-signals-strip.tsx

@@ -23,29 +23,38 @@ export function CorpusSignalsStrip({
 }: {
   aggregates: CorpusAggregates
 }) {
-  const repro = aggregates.reproducibility.overall
-  const comp = aggregates.completeness.overall
-  const prov = aggregates.provenance.overall
-  const cmp = aggregates.comparability.overall
+  // Each block can be partly missing if the backend hasn't computed it
+  // for the snapshot in use; guard every read so a partial payload
+  // renders as "—" instead of crashing the route.
+  const repro = aggregates.reproducibility?.overall ?? null
+  const comp = aggregates.completeness?.overall ?? null
+  const prov = aggregates.provenance?.overall ?? null
+  const cmp = aggregates.comparability?.overall ?? null
 
   // Invert the gap rate to read as "documented", which matches reader intuition.
-  const reproDocumented =
-    repro.reproducibility_gap_rate == null
-      ? null
-      : Math.max(0, 1 - repro.reproducibility_gap_rate)
-  const reproDetail = topMissingFields(repro.per_field_missingness, 2)
-
-  const totalReports = prov.total_triples
-  const tpShare = totalReports > 0 ? prov.source_type_distribution.third_party / totalReports : 0
-  const fpShare = totalReports > 0 ? prov.source_type_distribution.first_party / totalReports : 0
-
-  const multiSourceRate = rate(prov.multi_source_triples, prov.total_triples)
-  const cmpRate = rate(cmp.variant_divergent_count, cmp.groups_with_variant_check)
+  const reproGapRate = repro?.reproducibility_gap_rate ?? null
+  const reproDocumented = reproGapRate == null ? null : Math.max(0, 1 - reproGapRate)
+  const reproDetail = topMissingFields(repro?.per_field_missingness ?? {}, 2)
+
+  const totalReports = prov?.total_triples ?? 0
+  const sourceDist = prov?.source_type_distribution
+  const tpShare = totalReports > 0 && sourceDist?.third_party != null
+    ? sourceDist.third_party / totalReports
+    : null
+  const fpShare = totalReports > 0 && sourceDist?.first_party != null
+    ? sourceDist.first_party / totalReports
+    : null
+
+  const multiSourceRate = rate(prov?.multi_source_triples, prov?.total_triples)
+  const cmpRate = rate(cmp?.variant_divergent_count, cmp?.groups_with_variant_check)
   const crossPartyRate = rate(
-    cmp.cross_party_divergent_count,
-    cmp.groups_with_cross_party_check
+    cmp?.cross_party_divergent_count,
+    cmp?.groups_with_cross_party_check,
   )
-  const crossPartyAvailable = cmp.groups_with_cross_party_check > 0
+  const crossPartyAvailable = (cmp?.groups_with_cross_party_check ?? 0) > 0
+  const variantDivergent = cmp?.variant_divergent_count ?? null
+  const variantEligible = cmp?.groups_with_variant_check ?? null
+  const completenessTotal = comp?.total_triples ?? null
 
   return (
     <div className="signals-grid">
@@ -56,17 +65,21 @@ export function CorpusSignalsStrip({
         headline="of reported scores have a complete setup recorded — the rest cannot be independently re-run."
         detail={
           reproDetail
-            ? `${formatPct(repro.reproducibility_gap_rate)} have at least one undocumented field. Most often missing: ${reproDetail}.`
-            : `${formatPct(repro.reproducibility_gap_rate)} have at least one undocumented field.`
+            ? `${formatPct(reproGapRate)} have at least one undocumented field. Most often missing: ${reproDetail}.`
+            : `${formatPct(reproGapRate)} have at least one undocumented field.`
         }
         asks="Can someone else run this evaluation and get the same number?"
       />
       <SignalTile
         id="completeness"
-        statValue={pctNum(comp.completeness_avg)}
+        statValue={pctNum(comp?.completeness_avg)}
         statUnit="%"
-        headline={`mean across ${comp.total_triples.toLocaleString()} reported score triples.`}
-        detail={`Observed range: ${formatPct(comp.completeness_min)} to ${formatPct(comp.completeness_max)}.`}
+        headline={
+          completenessTotal != null
+            ? `mean across ${completenessTotal.toLocaleString()} reported score triples.`
+            : "mean across reported score triples."
+        }
+        detail={`Observed range: ${formatPct(comp?.completeness_min)} to ${formatPct(comp?.completeness_max)}.`}
         asks="Is the benchmark itself documented well enough to interpret a score on it?"
       />
       <SignalTile
@@ -81,7 +94,11 @@ export function CorpusSignalsStrip({
         id="comparability"
         statValue={pctNum(cmpRate)}
         statUnit="%"
-        headline={`of setup-eligible groups diverge across variants (${cmp.variant_divergent_count.toLocaleString()} of ${cmp.groups_with_variant_check.toLocaleString()}).`}
+        headline={
+          variantEligible != null && variantDivergent != null
+            ? `of setup-eligible groups diverge across variants (${variantDivergent.toLocaleString()} of ${variantEligible.toLocaleString()}).`
+            : "of setup-eligible groups diverge across variants."
+        }
         detail={
           crossPartyAvailable
             ? `Cross-party divergence: ${formatPct(crossPartyRate)}.`