feat(client): looksLikeV3Capabilities — drop the v2-downgrade footgun on capabilities-validation failure

[bokelley]

Summary

Port of JS commit 27bd79d (#1201) to the Python SDK. Closes #461. Refs #452.

When get_adcp_capabilities fails strict schema validation but the response is structurally v3-shaped, surface the v3 validation error loudly instead of silently re-classifying the agent as v2 — which downstream tooling turns into a cascade of confusing "AdCP schema data for version v2.5 not found" errors that have nothing to do with the original wire-shape bug.

• looks_like_v3_capabilities(data) (in src/adcp/capabilities.py, also re-exported from top-level adcp) checks for any one v3 signal: the adcp envelope, supported_protocols array, or any v3 protocol-level block (account, media_buy, signals, creative, brand, governance, sponsored_intelligence, compliance_testing).
• Arrays are explicitly rejected via a small _is_plain_object helper so adcp: [] / media_buy: [] don't false-positive (mirrors the JS isPlainObject defense).
• ADCPClient.refresh_capabilities distinguishes parse failure (the _parse_response "Failed to parse response:" prefix) from transport failure: parse failures re-fetch the raw dict from the adapter and run the heuristic; transport failures fall straight through to the original "Failed to fetch capabilities" path so the existing failure-raises test stays green.

Test plan

• 23 unit tests for the heuristic (positive + negative shapes, including the array-rejection defense)
• 4 wire-up tests on refresh_capabilities:
  • clean v3 → parses successfully
  • broken v3 with one shape bug → loud v3-flavor error, error message mentions v3 and never v2.5
  • non-v3-shaped response → falls through to existing transport-error path
  • transport failure with no data → still raises "Failed to fetch capabilities", adapter called exactly once (no extra probe)
• ruff check clean on touched files
• mypy clean on src/adcp/capabilities.py + src/adcp/client.py
• Full tests/ suite: 3065 passed, 18 skipped, 1 xfailed (no regressions)

🤖 Generated with Claude Code