feat(providers): claude on azure (#1009 final cell)

[chaholl]

Summary

Closes the final cell of the provider × platform matrix. Best-endeavors:
the URL pattern follows Microsoft's standard AIServices deployment shape
(mirrors openai+azure exactly except /messages instead of
/chat/completions). No live verification has been done — the
Foundry Anthropic quota for our subscription is currently 0 in eastus2
and swedencentral and quota requests are portal-only.

Wire format (planned, to verify on first live run)

URL	{endpoint}/openai/deployments/{deployment}/messages?api-version={v}
Auth	Entra Bearer via credentials.AzureCredential (same as openai+azure)
Body	Standard Anthropic Messages API verbatim — model in body, stream:true for streaming. Distinct from Bedrock/Vertex which use the partner shape.
API version	PlatformConfig.AdditionalConfig.api_version → defaults to 2024-12-01-preview (same as openai+azure)
Headers	No anthropic-version (api-version query parameter pins the contract)

If the Foundry deployment turns out to use a vendor-prefixed pattern
({endpoint}/anthropic/v1/messages?api-version=...) instead, the fix
is a one-line change to messagesURL/messagesStreamURL.

Code structure

• azurePlatform constant + defaultAzureAPIVersion in claude.go,
  plus azureAnthropicDeploymentEndpoint(endpoint, deployment).
• isAzure() predicate. New usesCredentialAuth() predicate
  covering Bedrock + Vertex + Azure (the auth-side superset of
  isPartnerHosted, which stays Bedrock+Vertex only because Azure
  does not use the partner body shape).
• azureAPIVersion() reads PlatformConfig.AdditionalConfig with
  defaultAzureAPIVersion fallback.
• messagesURL() / messagesStreamURL() branch on Azure to produce
  {baseURL}/messages?api-version={v}. Both return the same path —
  Azure uses one endpoint signalled by stream:true in body.
• NewProviderWithCredential derives the deployment URL from
  PlatformConfig.Endpoint when BaseURL == "" and Platform="azure".
• makeClaudeHTTPRequest skips the anthropic-version header for
  Azure (api-version pins it). Errors route through
  providers.ParsePlatformHTTPError.
• applyToolRequestHeaders switched to usesCredentialAuth so Azure
  goes through credential.Apply (Bearer) rather than x-api-key.
• PredictStream direct branch now calls messagesStreamURL()
  (Azure-aware) and skips the anthropic-version header for Azure.
• Registry skips https://api.anthropic.com default for
  Platform == "azure".

Tests

• azure_unit_test.go — predicates, URL builder, factory URL
  derivation, api-version selection, header omission, full
  PredictStream wire-format assertion against an httptest server.
• registry_extended_test.go — claude+azure regression test
  matching the openai+azure / gemini+vertex / claude+vertex pattern.

Bedrock / Vertex / direct paths unchanged — confirmed via existing
unit tests (all pass) and not regressed by the predicate refactor.

Pre-commit

• lint: 0 issues
• coverage on changed files: claude.go 83.9%, claude_streaming.go
  88.1%, claude_tools.go 84.1%, registry.go 93.5% (all ≥80%)

Test plan

• go test -race -count=1 ./runtime/providers/claude/... ./runtime/providers/... — green
• New Azure unit tests pass (URL/auth/body wire format verified via httptest)
• Registry regression test pass
• Existing claude+vertex, claude+bedrock, direct-API tests pass — refactor didn't regress
• Live verification — deferred: requires Anthropic quota allocation in your Foundry subscription (portal-only flow). First live PredictStream call will validate the URL shape; if the model returns 404 the fix is a one-line URL pattern swap to /anthropic/v1/messages per the inline comment.

Matrix status after this PR

	bedrock	vertex	azure
claude	✅	✅	✅ (this PR, unverified)
openai	✅	⛔ rejected	✅
gemini	⛔ rejected	✅	⛔ rejected

9 of 9 cells delivered. #1009 is closeable on the code side; one cell flagged as "URL-shape unverified pending live deployment".

Closes #1009

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud