Reviewed end-to-end against main. CI is fully green across 3.10/3.11/3.12/3.13 + lint + typecheck + codecov, and the surface closes out the group-DM coverage cleanly. The new transport infrastructure (multipart upload + binary GET on both clients) is the most interesting part — flagging a few cross-helper asymmetries below.

Strengths

• RFC 6266 §4.2 filename escape pinned at the byte level. test_upload_message_attachment_escapes_quote_in_filename asserts b'filename="weird\\"name.png"' in body — not just "has escape" but the exact backslash-quoted form FastAPI's UploadFile parser expects. The right depth for a hand-rolled multipart envelope.
• Emoji URL encoding pinned by exact wire bytes — %F0%9F%91%8D (sync) and %F0%9F%8E%89 (async), not just "contains percent encoding". Catches a future regression where someone passes safe=':/' and the multi-byte sequence sneaks through.
• Multipart wire shape inspected directly — test_upload_message_attachment_builds_multipart_body parses out the boundary token from the Content-Type header and confirms it appears in the body as --{boundary}--. Verifies the round-trip rather than trusting it.
• Error envelope parity through _build_api_error — 413 + structured {detail: {message, code: "LIMIT_EXCEEDED"}} unwraps as ColonyAPIError with .code == "LIMIT_EXCEEDED"; 403 on bytes path becomes ColonyAuthError. Same shape as the JSON path. Pinned on both helpers, both clients.
• ColonyNetworkError parity — URLError (sync) and httpx.ConnectError (async) both wrap with status=0 so callers branch on type rather than parsing message strings. Pinned on both helpers, both clients.
• Lazy-token mint coverage — both new transport paths trigger _ensure_token() when the client has no token in memory. Pinned on sync + async.
• Hook fan-out coverage — on_request + on_response callbacks fire on both multipart-upload and bytes-GET paths. Easy to forget when copy-pasting _raw_request plumbing; pinned explicitly.
• Mock records byte-length, not raw bytes. test_upload_message_attachment_records_size_not_bytes confirms this — sensible call-shape choice for large uploads, and grep-friendly. The bytes-returning getters yield a deterministic sentinel by default, overridable via responses={"get_message_attachment": b"..."} — pinned by test_get_message_attachment_custom_bytes_response.
• os.urandom(16).hex() for the sync boundary token prevents collisions across concurrent uploads from the same client. Async delegates to httpx for the same reason.
• PR scope discipline — author explicitly held the version bump for a dedicated release PR consolidating all three changelog entries. Matches the plan committed to in #56.

Minor observations (none blocking)

1. Async multipart + bytes helpers skip self.last_rate_limit tracking. The sync counterparts both do self.last_rate_limit = RateLimitInfo.from_headers(...), and the existing async _raw_request (src/colony_sdk/async_client.py:389) also tracks it on every response. The new async helpers don't. Callers reading client.last_rate_limit after an await client.upload_message_attachment(...) would see stale data on async but fresh data on sync. Worth a follow-up to re-sync the shared client state.
2. Circuit breaker bypassed by both helpers on both clients. Sync _raw_request and async _raw_request both check _circuit_breaker_threshold before issuing and update _consecutive_failures on failure / reset on success; the new multipart + bytes helpers do none of this. The PR description explains the retry skip ("uploads + downloads are rarely safe to retry blindly"), but the circuit breaker is a pre-flight refusal that wouldn't retry — bypassing it means a tripped breaker still lets uploads through. Probably intentional but the rationale isn't documented; worth a note or a small if breaker_open: raise at the top of each helper.
3. Idempotency-Key header not supported on multipart. Sync _raw_request threads Idempotency-Key (per the #56 review); the new multipart helpers don't accept one. Uploads are exactly where idempotency matters — network blip during a 7 MB POST and a naive retry creates two attachments. Server-side dedup-by-content_hash (the deduped: bool response field) mitigates this case, so it's defensible to defer, but worth a note that the same hash will dedupe a retry rather than the SDK threading the header.
4. field_name kwarg is wired but always "file". Both helpers pass field_name="file" for every callsite. Fine — leaving the param exposed makes a future endpoint with a different field name a one-line change rather than a refactor.

Approval note

Approving — the three observations above are all follow-up work, not blockers. Holding the merge button for a human reviewer per the TheColonyCC/* convention.