biome check has been failing repo-wide since the calibration/admin PRs
landed with unformatted files, so every branch inherits a red CI check.
Formatting-only; no behavior change.

Co-Authored-By: Zypher Agent <zypher@corespeed.io>

* fix(graph): traverse from relevant hubs, not the newest pages

The graph rendered as all-isolated dots ("全散了") — every node degree 0,
uniform size, no edges — even though the brain has a dense link graph.

Root cause: #45 picks the deep-traversal roots as
`[...titles.keys()].slice(0, TRAVERSE_ROOTS)`. `titles` is filled
list_pages-first in updated_desc order, so the roots are just the 8
NEWEST pages — which are frequently freshly-created and not yet linked.
A depth-5 traversal from an unlinked page reaches nothing, and the rich
hubs (entities/haas, companies/corespeed) sit at position #9+, excluded
from the roots. Result: 0 edges, deterministically.

- Seed the traversal roots from the seed-query hits (relevance-ranked,
  round-robin across queries so each contributes its top hit), which
  reliably surface the hubs; fall back to recent pages only to fill
  TRAVERSE_ROOTS. On the live brain this restores 0 → 87 edges.
- edgeRows now reports ok/failure instead of collapsing both to []. When
  the graph ends up edgeless AND a traversal actually errored, buildGraph
  throws so /api/graph returns 502 (uncached) rather than caching a
  misleading scattered graph for the full 1h TTL.
- Add a test for the fail-loud path; update AGENTS.md graph bullet (which
  still described the pre-#43 get_links implementation).

Co-Authored-By: Zypher Agent <zypher@corespeed.io>

* chore: format admin files left unformatted by #42/#44

biome check has been failing repo-wide since the calibration/admin PRs
landed with unformatted files, so every branch inherits a red CI check.
Formatting-only; no behavior change.

Co-Authored-By: Zypher Agent <zypher@corespeed.io>

* fix(graph): close the fail-loud guard's blind spots + serve stale over 502

Review hardening on top of the hub-roots fix (multi-model review of
this PR):

- Every upstream read feeds ONE failure signal: seed queries and
  list_pages now track failures (a total outage used to cache an EMPTY
  graph as healthy for the 1h TTL; an all-seed-query failure silently
  reverted roots to recency order — the exact regression this PR
  fixes). edgeRows treats MCP isError results, non-array payloads, and
  edge-row-less arrays (schema drift) as failed reads instead of
  "healthy, zero edges".
- The edgeless guard gates on the PRE-filter edge set, so a brain
  whose every edge touches a hash-titled page doesn't 502.
- Degraded builds (some reads failed, edges survived) are served but
  NOT cached, so a partial graph can't get pinned for the full TTL.
- Rebuilds are single-flighted, and a failed rebuild serves the last
  good expired graph stale instead of a 502 — no full-pipeline retry
  storm against an already-degraded gbrain.
- The route logs the failure reason before the generic 502.
- Overview renders the link stat as "—" and TopHubs says the data is
  unavailable when the graph read failed, instead of a misleading 0.
- Hash-titled mem0 imports no longer consume traversal-root slots.
- Tests: root selection is pinned (reverting to recency-ordered roots
  fails the suite — mutation-verified), plus coverage for the
  uncached-throw contract, isError, total outage, partial failure,
  stale-serve, single-flight, round-robin order, and the route 502.

Co-Authored-By: Zypher Agent <zypher@corespeed.io>

---------

Co-authored-by: Zypher Agent <zypher@corespeed.io>