fix: preserve SDK error results through typed teardown exceptions

[svozza]

After SDK 0.2.160, a Bedrock API failure can emit an error ResultMessage followed by the SDK's new ResultError. Aceiro recognized only the previous bare-exception wrapper. Because ResultError inherits ProcessError, the run incorrectly reported “without a result envelope,” skipped its API-error record and retry path, and could discard an already verified artifact.

This occurred in scheduled run 36433006668: 41/42 review scenarios and 6/6 plan scenarios passed; sandbox_credential_read submitted incomplete arguments, then exhausted ten CLI retries on Bedrock 503s. The failed sample remains invalid and its evidence is retained.

Recognize the typed SDK wrapper only after capturing an error result, so the existing retry, permanent-error and verified-artifact rules can handle it. Plain process crashes, programming errors and absent/successful envelopes remain failures. No model, prompt, fixture, grader or retry-budget changes.

Validation:

• 2,376 deterministic tests passed; two fixture-dependent skips. Type checking passed.
• Regression tests failed on main before the fix, then passed for retrying 503s, preserving verified artifacts and diagnosing non-retryable permission errors.
• A custom in-memory transport exercises the installed SDK's actual message reader and exit-wrapper behavior, in addition to scripted recovery tests.
• All 42 review and six remediation scenarios passed three repetitions with unchanged inputs: 126/126 review and 18/18 plan samples using Opus 5.5/high.
• GitHub CI passed on the same commit, including 42/42 review and 6/6 plan samples (run 36436709851).
• One additional 503 recovered inside the CLI during local validation; the deterministic regression tests cover exhausted internal retries and the resulting SDK exception.

The api_error metric continues to count terminal session API failures. Internal CLI retries remain visible in the captured stream.