fix: align API clients with the current v3 API and add CI tests

[digizeph]

Updates the API clients that no longer match the live v3 responses and adds CI coverage:

• Broker: latest() raised TypeError: BrokerItem.__init__() got an unexpected keyword argument 'delay'. Latest items carry a per-item delay (seconds since the latest catalog update), now modeled on BrokerItem. peers() and collectors() parsed outdated field shapes, and _paginate could loop forever on endpoints that ignore pagination parameters (it now pages until the reported total and also stops when a page adds no new items).
• AsnLookup: v3 items are nested (as2org, hegemony, peeringdb, population); AsnInfo models them with org_name/org_id lifted from as2org, and Pagination matches the API object (total_pages, total_count, has_next, has_prev).
• CommunityLookup: /communities is per-ASN with a nested community list; query() returns one CommunityEntry per community, and sources() reads the sources key instead of silently returning an empty list.
• Roas: targets the v3 ROAS endpoint, paginates to the reported total, and uses the endpoint's 0-indexed page parameter.
• Shared plumbing (bgpkit/_utils.py): from_item (dataclass parsing that ignores unknown fields), request_json (HTTP error surfacing with a configurable timeout), and BGPKITApiError. Requests now fail loudly instead of returning empty lists.
• Broker base URL: BGPKIT_BROKER_URL is honored (the same variable the Rust broker client uses); an explicit api_url argument still wins.

Unit tests cover recorded response shapes for every client and a new Tests workflow runs them on push and pull request. The suite is green (17 tests), and all clients were also exercised against the live API.

[Copilot]

Copilot review overview

🟡 Changes recommended

The ASN client skips page zero by default, and transport failures bypass the new public API exception.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Default v3 ASN pagination to page 0 · New
• Wrap transport failures in BGPKITApiError · New

What changed in this PR

Aligns API clients with current v3 response formats and adds automated unit-test coverage.

Changes:

• Updates Broker, ASN, Community, and ROAS response parsing and pagination.
• Adds shared HTTP/error-handling utilities.
• Adds recorded-response unit tests and a CI workflow.

File	Description
.github/​workflows/​test.yml	Runs unit tests on Python 3.9 and 3.12.
bgpkit/​_utils.py	Adds shared parsing and request helpers.
bgpkit/​__init__.py	Exports BGPKITApiError.
bgpkit/​bgpkit_asn.py	Parses nested v3 ASN responses.
bgpkit/​bgpkit_broker.py	Updates Broker models, pagination, and configuration.
bgpkit/​bgpkit_community.py	Flattens nested communities and parses sources.
bgpkit/​bgpkit_roas.py	Migrates ROAS queries to the v3 endpoint.
tests/​test_asn_unit.py	Tests v3 ASN parsing and errors.
tests/​test_broker_unit.py	Tests Broker response shapes and pagination.
tests/​test_community_unit.py	Tests community flattening and sources.
tests/​test_integration.py	Expands live API integration coverage.
tests/​test_roas_unit.py	Tests ROAS pagination and errors.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🔵 Needs a closer look

The Broker full-feed property misclassifies peers exactly at either documented threshold.

Review effort: Balanced
Findings: None

Resolved since last review (2)

• Default v3 ASN pagination to page 0
• Wrap transport failures in BGPKITApiError

Previously missed (1)

In code that hasn't changed since last review

Correct inclusive full-feed threshold classification

bgpkit/​bgpkit_broker.py:46

The full-feed thresholds are inclusive in BGPKIT's peer model (>= 700_000 IPv4 or >= 100_000 IPv6), so peers exactly at either threshold are currently misclassified as partial-feed. This can also contradict the result of peers(full_feed=True) for those boundary values.

[digizeph]

On the full-feed boundary note for bgpkit_broker.py:46: keeping the comparison strict is intentional — the property mirrors the endpoint's own full-feed filter, which applies num_v4_pfxs > 700_000 || num_v6_pfxs > 100_000 for full_feed=true (the parameter documentation uses the same strict wording). An inclusive property would disagree with the server-side classification for peers exactly at a threshold. No peer in the current dataset sits exactly at either threshold (3,406 rows checked; nearest v6 counts 99,833 / 100,484), so this is a definitional edge with no behavioral impact today. If the API definition should become inclusive, the property can flip in the same change.