docs(site): 3 arch_category updates in the field explorer (cassia-2, #165) (#237)

Fold in cassia-2's three architecture-category corrections (grounded in the
compass_artifact + 2026-05-24-true-memory-comparison repo docs, not guessed):
- supermemory: unspecified -> extraction
- letta_memgpt: unspecified -> hybrid
- cognee:       unspecified -> compile-upstream

unspecified drops 13 -> 10 (Celiums et al. stay unspecified — the docs discuss
their benchmark methodology, never their architecture). Updated in both the
canonical baselines/cross_system_multipass_matrix_2026-05-30.json field_roster
AND the embedded #fieldData copy the explorer renders, so they agree.

Follow-up to #236 (the explorer merged before these landed — they were lost
when the worktree was cleaned up mid-session; re-applied here off main).

Headless-validated: html.parser clean, embedded JSON 33 rows, node --check PASS,
arch_category filter chips now bucket the 3 systems correctly.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: jphein

baselines/cross_system_multipass_matrix_2026-05-30.json

@@ -1252,7 +1252,7 @@
         "source": null,
         "source_type": "not_benched"
       },
-      "arch_category": "unspecified"
+      "arch_category": "extraction"
     },
     "everos_evermind": {
       "status": "published_field_only",
@@ -1681,7 +1681,7 @@
         "source": null,
         "source_type": "not_benched"
       },
-      "arch_category": "unspecified"
+      "arch_category": "hybrid"
     },
     "cognee": {
       "status": "no_published_data",
@@ -1714,7 +1714,7 @@
         "source": null,
         "source_type": "not_benched"
       },
-      "arch_category": "unspecified"
+      "arch_category": "compile-upstream"
     },
     "caviraoss_openmemory": {
       "status": "no_published_data",
@@ -1878,4 +1878,4 @@
     "baseline-control": "SME-internal control / diagnostic arm, not a product",
     "unspecified": "the survey states no architecture; not assigned (no fabrication)"
   }
-}
+}