fix(runtime): include entrypoint in the handler cache key

[rlemke]

Problem

RegistryDispatcher._load_handler cached the resolved handler callable keyed by (module_uri, checksum) — dropping the entrypoint. Two facets registered from the same module (same module_uri + checksum) but with different entrypoints both received the first-cached callable, silently mis-routing every facet after the first to a single handler.

Domain packages register one dispatch entrypoint per module (routing on the facet name internally), which masked this. A handler module that exposes N distinct entrypoints — one per facet — hits it directly: e.g. a BuildMap facet dispatched to a sibling FetchGasPrices handler and returned the wrong return schema ({layer: …} instead of {html_path: …}), which then broke a downstream map.html_path reference.

Found while standing up a multi-facet handler module on the runtime.

Fix

Key the cache by (module_uri, checksum, entrypoint). One-line change in _load_handler; the _module_cache type annotation updated to the 3-tuple with an explanatory comment.

Tests

• New regression test_distinct_entrypoints_same_module_do_not_collide — two entrypoints in one module must each dispatch to their own handler, checked in both dispatch orders. Verified it fails on the old key and passes on the new.
• Updated the existing test_checksum_change_evicts_cache assertion for the new 3-tuple key.
• Full suites green: test_dispatcher.py, test_registry_runner.py, test_runner_service.py, test_inline_dispatch.py — 253 passed.

Risk

Low. The cache becomes strictly more specific; domains using a single entrypoint per module are unaffected (their key just gains a constant third element). No behavior change except correct routing for multi-entrypoint modules.

🤖 Generated with Claude Code

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[rlemke]

Note on the lint check: the failure is pre-existing on main, not introduced by this PR.

The 3 files changed here pass both linters cleanly:

• ruff check facetwork/runtime/dispatcher.py tests/runtime/test_dispatcher.py tests/runtime/test_registry_runner.py → All checks passed!
• ruff format --check on those files → 3 files already formatted

Running CI's repo-wide commands locally shows the failures are in unrelated files (36 would-reformat + 25 ruff check errors, e.g. unused pytest imports in test_parser.py / test_validator.py) that this PR does not touch. Kept the PR focused on the dispatcher fix rather than bundling a repo-wide reformat.

[rlemke]

Update — the red lint / test / typecheck checks are pre-existing on main, not introduced here.

Evidence: main's own latest CI run is failure (gh run list --branch main → CI: failure). This PR's changes are clean against all three:

• lint — the 3 changed files pass ruff check and ruff format --check; the repo-wide failures are in ~36 unrelated files.
• typecheck — mypy facetwork/runtime/dispatcher.py → 0 errors (I fixed the one annotation this PR touched, the module_cache property return type). Remaining mypy errors are in unrelated files (validator.py, block_execution.py, evaluator.py).
• test — the CI test job fails on a collection error in examples/osm-equity (ModuleNotFoundError: No module named 'tests…'), which interrupts the whole run before it gets to the change. Locally, the entire tests/runtime/ suite passes: 1257 passed, 24 skipped, including the new regression test.

So this PR is verified clean; the red checks are pre-existing CI-health debt (tracked separately). Happy to open a follow-up PR to green the CI (repo-wide ruff format, the mypy fixes, and the osm-equity collection error) if wanted — kept out of here to keep the dispatcher fix reviewable.

[rlemke]

Rebased onto main (picks up the pinned ruff + reformat from #33). lint ✅, test-go ✅, dispatcher suites pass (84 tests). The only red is typecheck — the pre-existing mcp version-drift tracked in #34, not from this change. Landing the verified dispatcher fix.