Summary

Two gaps left the docs build weaker than it looked. After this, mkdocs build --strict passes with zero warnings and nothing suppressed.

1. The nav rule was unenforceable

docs/README.md (from #1525) states that a new doc needs a nav entry in mkdocs.yml. I tested that claim and it was not enforced: a page absent from nav built cleanly under --strict, surfacing as a single INFO line, so a missing entry silently dropped the page out of the site's navigation.

validation.nav.omitted_files: warn now makes --strict fail on it. Verified both ways — a page left out of nav fails the build, and removing it passes again.

26 existing pages are out of nav on purpose and are declared in not_in_nav rather than switching the check off:

Listing every entry would bury the rest of the tree; declaring them keeps the check strict for everything else.

2. The griffe suppression is now unnecessary

#1543 shipped docs/_hooks/griffe_filter.py purely to keep one message — "No type or annotation for parameter" — from failing --strict. Its own docstring said "Remove this hook once the annotations land." All nine are annotated here, so the hook is deleted. Warning count is 0 with no filter in place.

Bonus: the generated signatures now carry types — device_id: int, log_level: int | None, config: CallConfig | None — instead of bare names.

Two annotations needed a decision, not a transcription

This is why I did not do it in #1543 as a mechanical pass:

• ChipWorker.init's bins → Any. The docstring accepts "any object exposing host_path / aicpu_path / …", and the body reads dispatcher_path through getattr. Naming RuntimeBinaries would be both narrower than the actual contract and a dependency from simpler onto simpler_setup — the wrong direction per project-layout.md.
• ChipWorker.run's handle → CallableHandle, imported under TYPE_CHECKING. The runtime import is deliberately lazy at its use site (task_interface.py:1157, with a # noqa: PLC0415), so a module-level import would undo that on purpose-avoided grounds. Both files already have from __future__ import annotations, so the annotations cost no import at all.

Verification

• mkdocs build --strict → exit 0, 0 warnings, 98 pages, with griffe_filter.py removed
• nav enforcement tested in both directions (omitted page fails; clean tree passes)
• The 9 griffe warnings went from 9 → 0
• Generated signatures confirmed to render the new types in the built HTML
• ruff check / ruff format --check clean; test_package_surface.py 5 passed / 1 skipped
• No behaviour change — annotations, one config block, and one deleted build hook