fix(agent): retry transient provider failures

[Alix-007]

Summary

• retry transient single-provider LLM failures inside the agent loop instead of aborting immediately
• reuse provider error classification so HTTP 500/502/503 style failures are treated like temporary outages
• add a regression test proving the agent recovers after one transient provider failure

Fixes #629

Test plan

• go test ./pkg/agent -run TestAgentLoop_ContextExhaustionRetry -count=1
• go test ./pkg/agent -run TestAgentLoop_TransientProviderRetry -count=1

[Alix-007]

Pushed an updated branch for this PR. It keeps the transient single-provider retry fix, resolves the current pkg/agent/loop drift against main, and keeps the targeted regression in place.

Local verification:

• go test ./pkg/agent -run TestAgentLoop_ -count=1

[Alix-007]

Pushed a follow-up lint-only fix on top of the retry update to satisfy golines in pkg/agent/loop.go and reran go test ./pkg/agent -run TestAgentLoop_ -count=1 locally.

[Alix-007]

Refreshed this branch onto the current main and force-updated the PR head.

During the refresh I explicitly dropped an old conflict-resolution commit that had accumulated unrelated registry/provider reload changes, so the PR is back to the intended narrow scope: transient retry handling for single-provider calls plus the focused regression test.

Local verification after the refresh:

• go test ./pkg/agent -run 'TestAgentLoop_' -count=1
• go run github.com/golangci/golangci-lint/v2/cmd/golangci-lint@v2.10.1 run ./pkg/agent/...

GitHub PR checks are green again.

[Alix-007]

Closing this to reduce queue noise. The branch is conflicting again, the issue lane already has another open carrier (#866), and this PR has aged without maintainer review despite several newer focused fixes being merged first. If we revisit this bug, it should come back as a fresh clean-base follow-up rather than continue on this stale branch.