test(ai-proxy): retry transient provider errors in LLM integration test

[nbouliol]

Problem

The LLM Integration Tests (ai-proxy) job on main goes red whenever a provider returns a transient error. The all models should support tool calls test treated an Anthropic HTTP 529 overloaded_error as a model capability failure:

AIProviderUnavailableError: Anthropic server error (HTTP 529): ... "type":"overloaded_error" ... model: claude-opus-4-5

The dispatcher is built with maxRetries: 0 by design (passthrough proxy), so nothing absorbs the blip, and expect(failures).toEqual([]) fails on a purely transient provider availability issue that says nothing about the model.

Fix

• Retry transient errors (429 / 5xx / 529 / overloaded / network timeouts) with a small backoff (up to 3 attempts).
• If a model still errors transiently after retries, skip it (logged) instead of counting it as a failure — a provider blip is not a capability signal.
• Genuine capability failures still fail the test.
• Hard infra errors (401/403/auth/DNS) still fail loudly — those mean the test setup is broken.

This job already has continue-on-error: true and is not in the release needs, so it never blocked publishing; this just stops the noise and makes the signal meaningful.

🤖 Generated with Claude Code

Note

Retry transient provider errors in LLM integration tool-call test

• Adds queryModelToolSupport, a helper that issues an ai-query with a function tool and retries up to 3 times (MAX_MODEL_QUERY_ATTEMPTS) with incremental backoff on transient errors.
• Classifies provider errors into hard (auth/infra), transient (rate-limit/5xx/network), or capability using two new regex patterns; hard errors fail immediately, transient errors are retried then skipped with a warning, and capability errors fail the test.
• Updates the 'all models should support tool calls' test in llm.integration.test.ts to collect skipped and failed models separately, so intermittent provider outages no longer cause false test failures.

^{Macroscope summarized fcc6648.}