feat(agent-sdk): resolveContext orchestrator + ticket classification (E1/E5/E6/E4)

[officialCodeWork]

Step 5.1 — resolveContext orchestrator (opens Phase 5)

Failure modes E1 / E5 / E6 / E4. New @coderadar/agent-sdk package (bundled into ui-lineage), the agent-facing entry point.

• classifyTicket(ticket) — rule-based + keyword lexicons, no LLM (deterministic, G8). Routes to:
  • visual (screenshot attached)
  • textual (UI terms in prose — E1)
  • behavioral ("clicking X does nothing" — E5)
  • out-of-domain (backend/infra/perf → decline — E6)
  • unsupported (video/GIF → decline — E4)
  • Precedence: unsupported → visual → out-of-domain → behavioral → textual.
• resolveContext(graph, ticket) — classify, then run the matcher on terms pulled from the prose (quoted phrases, else capitalized runs). OOD/unsupported get a structured decline instead of a forced match.

Eval

Hand-written ticket suite (eval/tickets/tickets.json, 17 cases: 5 visual, 4 textual, 3 behavioral, 3 out-of-domain, 2 unsupported). The harness reports and gates entry-point accuracy (≥ 0.90) and OOD rejection (≥ 0.95) — both 1.000.

CLI

$ ui-lineage resolve "The 'Billing summary' card is wrong"
entry point: textual — UI terms in prose — match on rendered text
BillingSummaryCard  confidence=high (1.00)

$ ui-lineage resolve "Redis latency spiked and the cron worker is backed up"
entry point: out-of-domain — backend/infra/perf vocabulary — not a UI ticket
declined (out-of-scope): … route it to a human or a backend tool.

Verification

41 core + 9 agent-sdk + 3 vision + 105 parser tests pass; pnpm eval green; bundle has no leaked @coderadar imports.

🤖 Generated with Claude Code