[JakubMrozek]

Summary

Fixes a hard crash (EXC_BREAKPOINT / SIGTRAP) reported from production (Polybuzz / PolySpeak — NetworkInfoProvider.observeNetworkType() in the crash log), by removing the crashing code path rather than guarding it.

NetworkInfoProvider resolved network.type by bridging NWPathMonitor.pathUpdateHandler into a CheckedContinuation:

monitor.pathUpdateHandler = { path in
    ...
    monitor.cancel()
    continuation.resume(returning: type)   // ← fires twice → trap
}

pathUpdateHandler is not one-shot — it re-fires on every path change, and monitor.cancel() does not retract callbacks already dispatched to the monitor's (concurrent) queue. On a flapping/constrained network a second callback resumes the same continuation twice, which Swift's checked continuation turns into a fatal _assertionFailure. The reporter's crash frame is nw_path_is_constrained — a second path update carrying a constrained-status change.

This ran inside DeviceCollector.collectAsync() → on every /preload (every chat message), so at scale it produced a steady, high-volume crash cluster concentrated on mobile users. It shipped in every release since 1.0.5 (Sept 2025) — not a v4 regression; v4 just relocated the code here.

Why remove instead of guard

We confirmed against the ad server (apps/ad-server) that network.type is not used for ad selection, targeting, filtering, or pricing — validation has every network field optional, and type only maps to OpenRTB device.connectiontype forwarded to DSPs. So there is nothing to preserve. A continuation that no longer exists cannot be double-resumed → this eliminates the crash class instead of mitigating it.

What changes

• type detection removed — the entire NWPathMonitor path (observeNetworkType / currentNetworkType) is deleted. type is now always nil.
• carrier and detail dropped — carrier was already always nil on iOS 16+; detail (cellular RAT) was only meaningful gated on type, which is gone. The mapRadioAccess table + CoreTelephony/Network imports go with them.
• userAgent unchanged — the one field the server actually consumes (tracked in the preload_requests analytics table).
• type/carrier remain in NetworkInfo as always-nil fields so the shape stays stable and a value can be reintroduced later without a breaking change.

Tests

Rewrote NetworkInfoProviderTests for the simplified contract: type/carrier always nil, bridge dict omits nil fields, userAgent surfaced verbatim. Full suite green (166 tests) on the simulator; the network suite dropped from ~127s to ~31s (no more NWPathMonitor waits).

Release

This PR includes the 0.1.0 release bump (CHANGELOG + podspec s.version). Minor bump (0.0.5 → 0.1.0) because removing NetworkInfo.detail and making type optional is a breaking API change for direct consumers — flagged with a ### Breaking note. After merge: tag 0.1.0 and pod trunk push per RELEASING.md.

Follow-up (sequenced after this release)

sdk-swift must bump KontextKit (currently pinned exact: "0.0.4") and adapt its consumer: NetworkInfo.type is now String? and NetworkInfo.detail is gone, so DeviceCollector won't compile until updated. To actually send network.type = null on the wire, sdk-swift's own NetworkDTO.type should become optional and forward nil (today it would coerce to .other). Then release a new KontextSwiftSDK — that's the version Polybuzz updates to get the fix. Same for the RN/Flutter iOS halves.

🤖 Generated with Claude Code