Honor AI_AGENT and pass raw values through

[renaudhartert-db]

This PR mirrors parallel changes in databricks-sdk-go, databricks-sdk-py, and databricks-sdk-java, keeping agent detection in sync across SDKs.

Summary

Honors the Vercel AI_AGENT=<name> convention as a secondary fallback for AI-agent detection in the user-agent header, and surfaces the raw fallback value instead of coercing unrecognized names to agent/unknown.

Why

The SDK appends the detected AI coding agent to the user-agent header so Databricks can understand which agents invoke the SDK. Today, when no explicit product-specific env var fires, detection consults only the agents.md AGENT=<name> standard, and any value it does not recognize is coerced to the literal "unknown".

Two gaps follow from that:

• AI_AGENT=<name> is the parallel convention pushed by Vercel's @vercel/detect-agent library. Tools that integrate with the Vercel ecosystem set it for the same purpose as AGENT, under a different name, and the SDK ignores it.
• Coercing unrecognized values to "unknown" buries useful signal. A tool reporting claude-code_2-1-141 shows up as agent/unknown, so versioned variants and not-yet-listed tools are indistinguishable. Bucketing arbitrary tool names is an ETL concern, not the SDK's.

This PR adds AI_AGENT as a fallback after AGENT and passes the raw value through instead of categorizing it.

What changed

Interface changes

None. The exported functions (lookupAgentProvider, agentProvider, clearAgentCache) keep their signatures.

Behavioral changes

• AI_AGENT=<name> is now honored as a fallback. AGENT is checked first; AI_AGENT only when AGENT is unset or empty. Empty is treated as unset for both. Explicit product-specific matchers (e.g. CLAUDECODE) still win over both.
• The fallback value is no longer coerced to "unknown". The raw value is passed through, sanitized to satisfy the user-agent allowlist ([-0-9A-Za-z_.+], other chars become -) and capped at 64 characters to keep the header bounded.

Internal changes

• packages/core/src/clientinfo/agent.ts: Add AI_AGENT_ENV_VAR and MAX_AGENT_FALLBACK_LEN constants; rewrite agentEnvFallback to consult AGENT then AI_AGENT, route the value through the existing sanitize() helper, and cap its length. Update doc comments.
• packages/core/tests/clientinfo/agent.test.ts: Update the two AGENT-fallback cases for raw passthrough; add cases for sanitization, the length cap, AI_AGENT detection, and AGENT/AI_AGENT precedence.
• packages/core/tests/clientinfo/default.test.ts: Update the integration case that previously expected agent/unknown.

How is this tested?

• npx vitest run packages/core/tests/clientinfo/agent.test.ts passes (40 tests).
• npm test --workspace=packages/core passes (367 tests); npm run test:browser --workspace=packages/core passes (266 tests).
• npm run format:check, npm run lint, and npm run typecheck are clean.

NO_CHANGELOG=true