[Moskize91]

Provider requests currently fail immediately on recoverable connection failures, and interrupted streams can lose diagnostic causes and consumption evidence. Add a portable wiki-graph-runtime package shared by core and job: bounded transport recovery over full response consumption, safe nested error diagnostics, cancellation-aware budgets, resource leases, and bounded concurrency.

JSON repair now has one implementation and a shared deadline. Candidates publish caches only after business validation; Summary publishes the selected revision. Quality gates distinguish legitimate empty output from all-failed fallback. Usage records real attempts independently of acceptance, and progress has explicit scope/phase/unit/attempt boundaries. Existing core retry configuration and public entry points remain compatible; timeouts now bound the full request including retries. POST retries may duplicate work and billing.

The runtime is included in workspace builds, portable checks, release order and isolated package-install verification. Host-specific credentials, cloud storage and state persistence remain outside this public package. Adapter contracts and limitations are documented in docs/provider-runtime.md.

Validation:

• pnpm lint:fix, pnpm verify, workspace build and pnpm smoke:pack-install passed.
• Full regression: 225 files, 1520 passed, 4 skipped.
• Final adapter checks: runtime/core/local executor 74 passed; SDK worker/artifact/local executor 27 passed.
• Coverage includes interrupted SSE, permanent provider failures, cancellation/backoff, retained capacity, first failure, rejected cache candidates, unknown/known consumption and valid empty output.
• Isolated Windows SDK packing includes runtime; CLI end-to-end installation uses exact local tarballs. Local CLI E2E: 7 files, 18 passed.