AGENT.md now forbids unsafe code and bans lint suppressions for precision loss, missing docs/errors, and dormant code; several crates still relied on those allowances.
The libtorrent adapter depended on a C++ bridge and build.rs, introducing unsafe blocks that violated the updated directives.
API/config paths carried #[allow] gates to bypass documentation and float-cast lints, masking real enforcement.
Decision:
Removed the libtorrent C++ bridge (build script and FFI sources) and now run the adapter solely on the safe StubSession, keeping the crate #![forbid(unsafe_code)].
Swapped float casts in rate limiting/formatting paths for integer-based accounting and From conversions, eliminating banned clippy allowances.
Added missing error docs and promoted constructors to const where viable to satisfy lint gates without exemptions.
Provisioned a local Docker runtime via colima so integration suites (Postgres-backed) execute instead of skipping, keeping coverage and DB-dependent tests meaningful.
Updated cargo-deny skips to reflect the current dependency graph (foldhash via hashbrown) without introducing new dependencies.
Consequences:
Native libtorrent integration is temporarily unavailable; the safe stub keeps orchestrator flows and tests exercising the engine API. Risk: production parity with libtorrent is paused—rollback by restoring the prior FFI bridge branch if needed.
Workspace now contains zero unsafe code and no banned #[allow] directives, aligning with AGENT.md’s lint posture.
Rate limiting uses deterministic integer tokens; behaviour should remain monotonic but merits monitoring under bursty traffic for regressions.
Follow-up:
Reintroduce a safe libtorrent integration (possibly in an isolated crate) once it can satisfy the no-unsafe mandate or after revisiting the directive in a dedicated ADR.
Add feature-flagged integration tests for the real adapter when restored, while keeping the stub path covered in CI.
Test coverage: DOCKER_HOST=unix:///Users/vanna/.colima/default/docker.sock just ci passes (fmt/lint/udeps/audit/deny/test/cov) with coverage at ~81% lines; rerun cargo deny to trim remaining skips when upstream unifies foldhash/hashbrown.