Swift Performance

Finding

The optimized Swift facade is neutral or faster for the shared crypto workloads. Identity import is 7.38% faster after removing the generic combined record path; the cached public-key getter is 91.90% faster. Parse-and-verify is 6.96% slower at the median and 12.49% slower at p95, within the descriptive 10%/15% latency budgets and not reproduced as a native Rust regression.

The audit used three paired optimized processes on a GitHub-hosted arm64 macos-15 runner with Swift 6.1.2. Values are nanoseconds per authored SoftchatKit call. Negative deltas mean the current SDK is faster.

Operation	Bytes	Old median ns	Current median ns	Delta	Current p95 ns	p95 delta
Import identity	32	17,871	16,552	-7.38%	17,979	-1.18%
Read cached public key	0	247	20	-91.90%	20	-92.22%
NIP-44 encrypt	32	45,979	44,054	-4.19%	44,490	-5.50%
NIP-44 decrypt	32	43,587	43,153	-1.00%	43,286	-6.57%
NIP-44 encrypt	1,024	53,399	48,900	-8.43%	49,421	-13.90%
NIP-44 decrypt	1,024	52,602	48,059	-8.64%	48,727	-13.77%
NIP-44 encrypt	64,000	376,964	366,781	-2.70%	379,782	-8.74%
NIP-44 decrypt	64,000	386,037	366,391	-5.09%	406,401	+4.27%
Sign NIP-01 event	32	95,549	92,454	-3.24%	97,601	+0.01%
Parse and verify event	374	39,696	42,460	+6.96%	45,655	+12.49%

The 20-ns getter is an API-shape check, not a useful absolute latency claim.

Current-only baseline

Operation	Bytes	Median ns	p95 ns
Create NIP-59 rumor	15	4,897	4,947
Create NIP-17 text message	15	15,485	15,689
Create durable gift wrap	15	292,357	295,934
Authenticate and unwrap	15	173,322	173,632
Stream-encrypt attachment	64,000	681,080	687,838

Optimizations retained

• SoftchatIdentity imports through the lightweight native constructor, caches its validated public key, and avoids the Android-only registry state.
• Encrypt and decrypt return or accept direct payload fields instead of round-tripping an aggregate generated record.
• Immutable Swift values own received events, messages, and metadata; native records do not leak into application layers.
• The identity lock protects handle lifetime and early erasure without making every public getter an FFI call.
• Streaming media uses bounded chunks and delays exposing decrypted output until authentication succeeds.

Footprint

The performance harness links three unstripped dynamic libraries:

Harness library set	Old bytes	Current bytes	Delta
Rust + generated FFI + SoftchatKit	2,425,336	4,895,304	+101.84%

This is a link-input measurement, not a shipped iOS application size. The Swift package carries a static XCFramework so the final application linker can dead-strip unused code. A release archive and dSYM measurement on physical iOS is still required before setting an Apple footprint gate.

Developer guidance

• Create one SoftchatIdentity for the authenticated session and retain its cached publicKey; do not recreate it per message.
• Call synchronous crypto and parsing from a bounded task or executor rather than the main actor. Do not add an async wrapper that hides unbounded work.
• Keep the identity as an ordinary ARC-owned reference. Call erase() only when authority must end early.
• Pass complete events, envelopes, relay frames, and attachment chunks through SoftchatKit. The generated FFI module is internal.
• Keep Keychain/Secure Enclave policy, URLSession, GRDB, background transfer, and application lifecycle in Swift.
• For extensions, create an independent session owner and close it explicitly; do not share callbacks or secret-derived state across process boundaries.

Run the shared and current-only audit with scripts/run-swift-performance.sh, or dispatch the manual performance workflow. Any facade, FFI, package, or native-profile change requires the macOS package consumer and affected performance audit.

Open work

The macOS source, DocC, XCFramework/SwiftPM build, and package consumer are green. Remaining Apple evidence is an external application/archive/extension integration, physical-iOS latency, startup and memory, and final archive/dSYM footprint.