feat(hub): expose 14 more DuckDB tables + fix edge-proxy healthcheck

DuckDB browser
--------------
GME 2.1.0rc1 ships 14 high-value tables the hub didn't surface yet:

  - openalex.work_github_urls (1.1k rows) — per-paper GitHub link
    that bridges OpenAlex works to repos already in our github index.
  - openalex.work_authors / work_institutions / work_references —
    the relational joins that let the row browser pivot from a work
    to its co-authors, host institutions, or downstream citers.
  - snsf.output_publications / output_datasets / output_academic_events
    / output_knowledge_transfers / output_public_communications /
    output_collaborations / output_use_inspired — every grant output
    keyed by grant_number so consumers can chain back to the
    funding source.
  - snsf.persons (146k) — the researcher registry every grant + output
    references.
  - swissubase.persons + swissubase.institutions — the author / org
    sides of the studies table, browsable independently.
  - epfl_graph.category_concepts (40k) — the discipline → concept
    edges underlying the EPFL Graph taxonomy.

All sit in _AUTO_TABLES + _AUTO_SEARCH_EXAMPLES; the heuristic
_AUTO_STAT_PATTERNS picks contextual stats per table automatically.

Container limits
----------------
HUB_MEM_LIMIT bumped from 512m to 2g. Default 512m was enough for the
original six DuckDBs but the new inventory keeps ~14 more files open
in DuckDB connection state; the cumulative working set blew past 512m
and the kernel OOM-killed the hub mid-request on the first sweep of
the new collections. 2g matches the openalex working set and leaves
headroom for the SQLite history db + Jinja2 cache.

Edge-proxy healthcheck
----------------------
wget --spider http://localhost:80/ followed Caddy's permanent
HTTP→HTTPS redirect into a real TLS handshake and choked on SNI
(localhost ≠ openpulse.epfl.ch). Every check failed for hours; the
container was marked unhealthy even though external traffic worked
fine. Hit Caddy's admin API on :2019 instead — it answers in plain
HTTP and only when the server process is alive.

Author: caviri

infra/open-pulse-stack/docker-compose.yml

@@ -461,7 +461,13 @@ services:
       - ${OPEN_PULSE_DATA_DIR:-./data}/edge-proxy/data:/data
       - ${OPEN_PULSE_DATA_DIR:-./data}/edge-proxy/config:/config
     healthcheck:
-      test: ["CMD-SHELL", "wget -q --spider http://localhost:80/ || exit 1"]
+      # Caddy serves a permanent 308 HTTP→HTTPS redirect on :80 which
+      # wget --spider follows into a real TLS handshake and chokes on
+      # SNI (localhost ≠ openpulse.epfl.ch) — every check failed even
+      # though external traffic worked fine. Hit Caddy's admin API on
+      # :2019 instead: it returns the running config as JSON over plain
+      # HTTP and only answers when the server process is alive.
+      test: ["CMD-SHELL", "wget -q -O- http://localhost:2019/config/ >/dev/null || exit 1"]
       interval: 15s
       timeout: 5s
       retries: 5

src/open_pulse/gui/hub/knowledge/duckdb_browser.py

@@ -321,11 +321,39 @@ class Backing:
         _DATA_ROOT / "index/openalex/duckdb/openalex.duckdb",
         "works",
     ),
+    # OpenAlex relational + GitHub-extracted tables. ``work_github_urls``
+    # is the highest-signal addition for cross-corpus joins — every row
+    # is a (research paper, GitHub repo) edge our pipeline can use to
+    # anchor a Person's contributions to their published work.
+    "openalex_work_authors": (
+        _DATA_ROOT / "index/openalex/duckdb/openalex.duckdb",
+        "work_authors",
+    ),
+    "openalex_work_institutions": (
+        _DATA_ROOT / "index/openalex/duckdb/openalex.duckdb",
+        "work_institutions",
+    ),
+    "openalex_work_references": (
+        _DATA_ROOT / "index/openalex/duckdb/openalex.duckdb",
+        "work_references",
+    ),
+    "openalex_work_github_urls": (
+        _DATA_ROOT / "index/openalex/duckdb/openalex.duckdb",
+        "work_github_urls",
+    ),
     # EPFL Graph
     "epfl_graph_disciplines": (
         _DATA_ROOT / "index/epfl_graph/duckdb/epfl_graph.duckdb",
         "categories",
     ),
+    # 39k discipline → concept edges. Browses the same EPFL discipline
+    # taxonomy as ``epfl_graph_disciplines`` but at the concept level
+    # (Wikipedia-grounded sub-topics), so the row count matches the
+    # number of (category, concept) pairs.
+    "epfl_graph_concepts": (
+        _DATA_ROOT / "index/epfl_graph/duckdb/epfl_graph.duckdb",
+        "category_concepts",
+    ),
     # ETH-Z Research Collection
     "ethz_research_collection_articles": (
         _DATA_ROOT
@@ -429,11 +457,57 @@ class Backing:
         _DATA_ROOT / "index/snsf/duckdb/snsf.duckdb",
         "grants",
     ),
+    # SNSF — research outputs broken out per type. Every row is keyed
+    # by ``grant_number`` so the row browser can pivot from any output
+    # back to the funding grant that produced it. ``persons`` is the
+    # 146k researcher registry every grant + output references.
+    "snsf_persons": (_DATA_ROOT / "index/snsf/duckdb/snsf.duckdb", "persons"),
+    "snsf_output_publications": (
+        _DATA_ROOT / "index/snsf/duckdb/snsf.duckdb",
+        "output_publications",
+    ),
+    "snsf_output_datasets": (
+        _DATA_ROOT / "index/snsf/duckdb/snsf.duckdb",
+        "output_datasets",
+    ),
+    "snsf_output_academic_events": (
+        _DATA_ROOT / "index/snsf/duckdb/snsf.duckdb",
+        "output_academic_events",
+    ),
+    "snsf_output_knowledge_transfers": (
+        _DATA_ROOT / "index/snsf/duckdb/snsf.duckdb",
+        "output_knowledge_transfers",
+    ),
+    "snsf_output_public_communications": (
+        _DATA_ROOT / "index/snsf/duckdb/snsf.duckdb",
+        "output_public_communications",
+    ),
+    "snsf_output_collaborations": (
+        _DATA_ROOT / "index/snsf/duckdb/snsf.duckdb",
+        "output_collaborations",
+    ),
+    "snsf_output_use_inspired": (
+        _DATA_ROOT / "index/snsf/duckdb/snsf.duckdb",
+        "output_use_inspired",
+    ),
     # SwissUBase — studies is the largest non-empty table.
     "swissubase_entities": (
         _DATA_ROOT / "index/swissubase/duckdb/swissubase.duckdb",
         "studies",
     ),
+    # SwissUBase author + institution registries. Cross-walked into the
+    # studies via ``study_persons`` / ``study_institutions`` join
+    # tables — kept here for direct lookup ("which Swiss social-sciences
+    # ORCIDs appear in our corpus?", "which institutions submit studies
+    # via SwissUBase?").
+    "swissubase_persons": (
+        _DATA_ROOT / "index/swissubase/duckdb/swissubase.duckdb",
+        "persons",
+    ),
+    "swissubase_institutions": (
+        _DATA_ROOT / "index/swissubase/duckdb/swissubase.duckdb",
+        "institutions",
+    ),
     # Zenodo — ``zenodo_records`` is hand-tuned in ``_BACKING`` above
     # (it joins ``record_communities`` to expose a per-record community
     # list). ``communities`` + ``creators`` are exposed here so the
@@ -461,8 +535,13 @@ class Backing:
     "institutions": ("EPFL", "ETH", "MIT", "Switzerland"),
     "sources": ("Nature", "Science", "IEEE", "ACM"),
     "topics": ("artificial intelligence", "genetics", "climate", "robotics"),
+    "openalex_work_authors": ("W2", "A5", "first", "corresponding"),
+    "openalex_work_institutions": ("EPFL", "ETH", "MIT", "I"),
+    "openalex_work_references": ("W2", "W3", "W4", "W5"),
+    "openalex_work_github_urls": ("torvalds", "tensorflow", "pytorch", "epfl"),
     # EPFL Graph
     "epfl_graph_disciplines": ("physics", "computer", "biology", "mathematics"),
+    "epfl_graph_concepts": ("machine", "neural", "protein", "quantum"),
     # ETH-Z Research Collection
     "ethz_research_collection_articles": (
         "deep learning",
@@ -501,8 +580,28 @@ class Backing:
     "snsf_epfl": ("EPFL", "machine learning", "physics", "Lausanne"),
     "snsf_ethz": ("ETH", "Zurich", "robotics", "quantum"),
     "snsf_switzerland": ("Swiss", "professor", "biology", "chemistry"),
+    "snsf_persons": ("Patrick", "Müller", "EPFL", "ETH"),
+    "snsf_output_publications": (
+        "machine learning", "quantum", "nature", "epfl",
+    ),
+    "snsf_output_datasets": ("dataset", "zenodo", "10.5281", "swiss"),
+    "snsf_output_academic_events": ("conference", "workshop", "ICML", "NeurIPS"),
+    "snsf_output_knowledge_transfers": (
+        "patent", "spin-off", "industry", "transfer",
+    ),
+    "snsf_output_public_communications": (
+        "media", "talk", "press", "interview",
+    ),
+    "snsf_output_collaborations": (
+        "Switzerland", "Germany", "United States", "industry",
+    ),
+    "snsf_output_use_inspired": ("application", "industry", "use", "EPFL"),
     # SwissUBase
     "swissubase_entities": ("survey", "FORS", "Switzerland", "households"),
+    "swissubase_persons": ("Müller", "Schmidt", "Patrick", "Anna"),
+    "swissubase_institutions": (
+        "FORS", "Lausanne", "Bern", "Switzerland",
+    ),
     # Zenodo
     "zenodo_records": ("dataset", "epfl", "10.5281", "machine learning"),
     "zenodo_communities": ("epfl", "swiss", "open", "research"),