docs: custom json output extension, add frontmatter, SEO optimization

fix: update all repo references to NVIDIA/NemoClaw

- Fix recovery scripts in agent-runtime.ts that still used curl -sf
  on / instead of the new HTTP status code pattern on /health (#3)
- Add device-auth-health-e2e to scorecard.needs (#8)
- Use openshell-${SANDBOX_NAME} SSH host alias in E2E test (#7)

Run NVIDIA#7 failed at the 'brev ls' sanity check in the Install Brev CLI
step with:

  API key saved for org ***
  rpc error: code = Internal desc = context deadline exceeded

Confirmed from local testing: the Brev platform's workspaces/list
endpoint is currently returning 'context deadline exceeded' across
MULTIPLE orgs (not just Nemoclaw CI/CD), even with working
credentials. Other endpoints (brev org ls, brev login) work fine.
This is a transient platform issue, not a credential or code issue.

Add a 5-attempt retry loop with 15s backoff around the brev ls
sanity check. On final failure, emit a ::warning:: and proceed \u2014
the test harness's own waitForSsh / refreshAndWaitForSsh calls
retry brev refresh on their own, so downstream steps will still
validate auth works. We only want to fail here if auth is
genuinely broken (e.g., wrong key), not because Brev's backend
is having a bad hour.

Last gap from the pre-flight audit: the budget accumulator was reading
brev_cost_usd from a per-candidate metadata.json field that the skill
doesn't currently emit. The cost would have silently stayed at zero,
making BREV_BUDGET_USD a no-op.

Switched to orchestrator-owned cost calc that doesn't depend on the
skill writing anything:

- New DOGFOOD_BREV_HOURLY_USD env var (default $3/hr — conservative,
  covers mid-tier GPU SKUs). Operators tune up for H100 candidates or
  down for CPU-only.
- Per candidate: record start_epoch / end_epoch around the agent
  invocation, compute cost = ceil(wallclock_sec / 3600 × hourly_rate),
  add to .spent-usd accumulator. Conservative-by-design — over-estimate
  beats under-estimate when the var is a hard budget gate.
- Write per-candidate cost.json (sibling to metadata.json so we don't
  clobber anything the skill emits): wallclock_sec, cost_usd, rate,
  agent_exit, started_at, ended_at, cost_method.
- Snapshot `brev ls --json` filtered to verify-stale-* before AND after
  each candidate (brev-pre.json / brev-post.json). Not used to gate —
  the Phase 7 straggler sweep is the actual cleanup — but surfaces
  candidates that leak boxes when reading the per-candidate logs.
- run-summary.json aggregator reads cost.json (was: metadata.json's
  missing brev_cost_usd field). Total adds up from the accumulator file.
- Preflight check #7 uses the new conservative ceiling
  (DOGFOOD_BREV_HOURLY_USD × 1 hour wallclock cap) instead of the
  hardcoded $1 floor that bore no relation to actual SKU pricing.

Trade-off acknowledged: the cost number is not actual Brev billing —
it's a deterministic estimate. The cost_method field in cost.json /
run-summary.json names this explicitly so anyone reading the run log
knows it's a budget gate, not an invoice.

Signed-off-by: Prekshi Vyas <prekshiv@nvidia.com>