[jarrad]

Summary

• Stop GPU-create candidate fallback only for HTTP 400 with Brev's WORKSPACE_ALREADY_EXISTS reason.
• Keep fallback for capacity/provider failures, unknown reason codes, and unrelated text containing the old duplicate phrase.

Proof

• TestRunGPUCreateRetryDecisions failed on the baseline: a code-only duplicate retried the second type; an unrelated provider message with duplicate workspace stopped early.
• go test ./pkg/cmd/gpucreate -run '^TestRunGPUCreateRetryDecisions$' -count=1 passed after the fix.
• go test ./pkg/cmd/gpucreate -count=1 passed.
• git diff --check passed.

Exact-head validation

All runs below passed for commit 598b62caf9d196f6b8ec4d90d1a408ae09449b76:

• test, vet, lint
• build, fmt, legacy
• CodeQL, CLI output compatibility

Stack and rollout

• Base: CLI error-code parser foundation #515. Review and merge feat(cli): preserve structured Brev HTTP error codes #515 before this PR. This review diff contains only BREV-13207.
• Server producer: BREV-13202's brev-deploy workspace-create contract must deploy before this CLI behavior can recognize duplicates. It emits fixed HTTP 400 with errors[].code = WORKSPACE_ALREADY_EXISTS.
• This PR changes retry control only; it does not alter server responses or generic failure handling.