fix(client): filter oneOf branch sub-errors from Ajv allErrors output

[bokelley]

Closes #1111

Summary

• Adds filterOneOfSubErrors() in src/lib/validation/schema-validator.ts, applied in both validateRequest and validateResponse before error formatting.
• Filter rule: drop any AJV error whose schemaPath is a strict descendant of a failing oneOf node's path. oneOf node errors are always kept (including nested ones via an explicit keyword === 'oneOf' guard). anyOf sub-errors are intentionally NOT filtered — anyOf allows partial matches and its per-branch errors carry useful disambiguation signal.
• Adds a regression test in test/lib/schema-validation.test.js that asserts oneOf sub-branch errors are absent from validateRequest output and the node error is preserved.
• Adds .changeset/oneof-suberror-filter.md (patch bump for @adcp/sdk and @adcp/client).

Root cause

schema-loader.ts initialises Ajv with allErrors: true. When Ajv evaluates a oneOf and no branch matches it:

1. Emits the top-level oneOf node error (actionable).
2. Emits one required/type error per property mismatch in every branch it evaluated (noise).

These sub-branch errors surfaced as spurious "must have required property" failures in storyboard CI, causing false-negative reports for downstream SDKs.

What was tested

• Pre-existing test/lib/schema-validation.test.js suite (build/unit — schema-dependent assertions are vacuous without synced schemas but not misleading).
• test/validation-enum-enrichment.test.js — unaffected by this change.
• npx prettier --write and npx tsc --noEmit pass clean on the branch.

Pre-PR reviews

Two expert reviews were completed before this PR was opened:

• code-reviewer — approved after blocker fix (nested oneOf guard added; test assertion hardened from silent-vacuous to assert.ok).
• ad-tech-protocol-expert — approved; filter is mechanically correct, anyOf exclusion justified, no impact on context_value_rejected hint diagnostics.

Out of scope / follow-ons

• NODE_ENV-sensitive defaultResponseMode() in client-hooks.ts (separate CI vs local discrepancy vector — distinct issue).
• anyOf flooding in get_adcp_capabilities responses (separate scope, different keyword).

---

Triage-managed PR — opened by the automated triage agent per .agents/routines/triage-prompt.md. Assign a human reviewer before merging.

https://claude.ai/code/session_01Sv61kPXqT5b8bLYZ62muHV

---

Generated by Claude Code

[bokelley]

Triage update — source issue resolved as not-a-bug.

@bokelley closed #1111 after tracing the actual root cause to adcp-client-python: examples/seller_agent.py::seed_product was producing skeleton products missing schema-required fields (publisher_properties: minItems 1, format_ids[].agent_url, available_reporting_frequencies: minItems 1). The @adcp/client@5.21.1 validator was reporting those fields correctly, not over-reporting from non-matching oneOf branches. Fixed downstream in adcontextprotocol/adcp-client-python#321.

The filterOneOfSubErrors() change in this PR is based on the original (incorrect) hypothesis and is no longer needed. Recommend closing this PR without merging. If the Ajv oneOf error-aggregation improvement is still wanted on its own merits — not as a bug fix but as a UX polish — that would be a fresh, scoped PR with a different rationale.

---

Triaged by Claude Code. Session: https://claude.ai/code/session_01WVfLw36MJBxN4oMZGuPUWh

---

Generated by Claude Code

[bokelley]

Superseded by #1113 (merged: "compact oneOf cascade so real residuals reach the wire"). Issue #1111 is closed. Closing this draft.

[bokelley]

Acknowledged — noted that #1113 ("compact oneOf cascade so real residuals reach the wire") supersedes this draft and issue #1111 is closed. No further action from triage.

---

Generated by Claude Code

---

Generated by Claude Code