The package root is the stable API surface. Deep imports under `inire.router.*` and `inire.geometry.*` remain accessible for advanced use, but they are unstable semi-private interfaces and may change without notice.
`RoutingResult.locked_geometry` stores canonical physical geometry only. The next run applies its own clearance rules when treating it as a static obstacle.
### Initial paths with `PathSeed`
Use `RoutingProblem.initial_paths` to provide semantic per-net seeds. Seeds are materialized with the current width, clearance, and bend collision settings for the run, and partial seeds are retried by normal routing in later iterations.
| `bend_collision_type` | `"arc"` | Bend collision/proxy model: `"arc"`, `"bbox"`, `"clipped_bbox"`, or, for backward compatibility, a custom polygon. A legacy custom polygon here is treated as both the physical bend and its proxy unless overridden by the split fields below. |
| `bend_proxy_geometry` | `None` | Optional explicit bend proxy geometry. Use this when you want a custom search/collision envelope that differs from the routed bend shape. Supplying only a custom polygon proxy warns and keeps the physical bend as the standard arc. |
| `bend_physical_geometry` | `None` | Optional explicit bend physical geometry. Use `"arc"` or a custom polygon. If you set a custom physical polygon and do not set a proxy, the proxy defaults to the same polygon. |
-`congestion_presence_cache_hits` / `congestion_presence_cache_misses`: Reuse of cached per-span booleans indicating whether a move polygon could overlap any other routed net at all.
-`congestion_presence_skips`: Number of moves that bypassed full congestion evaluation because the presence precheck found no other routed nets in any covered dynamic-grid span.
-`congestion_candidate_precheck_hits` / `congestion_candidate_precheck_misses`: Reuse of cached conservative per-span booleans indicating whether any candidate nets survive the net-envelope and grid-net broad phases.
-`congestion_candidate_precheck_skips`: Number of moves that bypassed full congestion evaluation because the candidate-net precheck found no surviving candidate nets after those broad phases.
-`congestion_candidate_nets`: Total candidate net ids returned by the dynamic net-envelope broad phase during routing.
-`congestion_net_envelope_cache_hits` / `congestion_net_envelope_cache_misses`: Reuse of cached dynamic net-envelope candidate sets keyed by the queried grid-cell span.
-`congestion_grid_net_cache_hits` / `congestion_grid_net_cache_misses`: Reuse of cached per-span candidate net ids gathered from dynamic grid occupancy.
-`congestion_grid_span_cache_hits` / `congestion_grid_span_cache_misses`: Reuse of cached dynamic-path candidate unions keyed by the queried grid-cell span.
-`congestion_lazy_resolutions`: Number of popped nodes whose pending congestion was resolved lazily.
-`congestion_lazy_requeues`: Number of lazily resolved nodes requeued after a positive congestion penalty was applied.
-`congestion_candidate_ids`: Total dynamic-path object ids returned by the congestion broad phase before exact confirmation.
Lower-level search and collision modules are semi-private implementation details. They remain accessible through deep imports for advanced use, but they are unstable and may change without notice. The stable supported entrypoint is `route(problem, options=...)`.
The current implementation structure is summarized in **[docs/architecture.md](docs/architecture.md)**. The committed example-corpus counter baseline is tracked in **[docs/performance.md](docs/performance.md)**.
Use `scripts/diff_performance_baseline.py` to compare a fresh local run against that baseline. The counter baseline is currently observational and is not enforced as a CI gate.
- Use `search.bend_physical_geometry` and `search.bend_proxy_geometry` together when you need a real custom bend shape plus a different conservative proxy.
