fix(sdk): classify Unavailable errors as side-effect free for retries

[detail-app]

Detail bug report: View on Detail

Closes #691

Bug

ErrorCode::Unavailable was classified as side-effecting (has_no_side_effects() == false), but every server-side use of this code maps from StreamerMissingInActionError, which is raised before the append is submitted to the streamer actor (at lease acquisition or message-send time) — so no mutation could have occurred.

The impact: clients using AppendRetryPolicy::NoSideEffects would not retry a unary append that received Unavailable after the request body had already been sent (frame signal set). The retry logic falls back to err.has_no_side_effects(), which returned false, so the safe retry was blocked. This contradicts the inline comments in lite/src/handlers/v1/error.rs ("Unavailable error code promised to be side-effect free").

Fix

Move ErrorCode::Unavailable from the false branch to the true branch of ErrorCode::has_no_side_effects() in api/src/v1/error.rs. This matches the documented intent and the streamer backend's own internal retry behavior (backend/core.rs already retries on StreamerMissingInActionError).

The contrasting RequestDroppedError (which occurs after submission, where a mutation may have taken effect) continues to map to ErrorCode::Other and remain side-effecting — unchanged.

Testing

• Added a unit test in api/src/v1/error.rs that requires every ErrorCode variant to be explicitly classified as side-effecting or side-effect free, guarding against future silent misclassifications as the #[non_exhaustive] enum evolves.
• Added an SDK test verifying an Unavailable server error is both retryable and side-effect free through the ServerError/AppendError surfaces.
• Extended the existing NoSideEffects retry-policy test to cover the core bug scenario: a signalled frame signal (body sent) + Unavailable error must now be safe to retry.
• Routine checks pass: unit tests for s2-api (49) and s2-sdk (106), s2-lite backend/streamer/handler tests (225), clippy (default features), and cargo +nightly fmt --all --check.

End-to-end verification

Started a local s2-lite server (s2 lite --port 4243) and ran the SDK integration test suite using the same scoping as CI (cargo test -p s2-sdk --all-features -- --skip access_token --skip metrics): 77 tests pass, 0 failed, confirming no regression in append/read/list flows.

The unfiltered integration run has 26 pre-existing failures (all 501 not_implemented from the lite server for metrics and access-token-management endpoints); verified identical on the baseline before this change via git stash.

Not verified

The exact race window — a streamer actor exiting mid-request after guard acquisition but before the append acks — was not reproduced end-to-end. The streamer's 60s dormancy timeout and the lack of an internal shutdown hook in the black-box SDK integration harness make it non-deterministic to trigger at a precise point. The verifiable core of the bug (the retry decision under a signalled frame signal + Unavailable) is covered by the extended unit test.

---

Automatic Fixes PRs can be configured here.

[greptile-apps]

PR author is not in the allowed authors list.