chore: code-review quick-wins cleanup (core, scripts, app-runner)

[leoschwarz]

Groups six small, independent cleanups surfaced by a code review. All behavior-preserving except the logging change noted below; none touch the core read/write API.

Changes

• api inspect now fails with a clear "only supported with the SUDS engine" error instead of leaking an AttributeError from private engine internals under Zeep.
• Entity.ENDPOINT becomes a ClassVar[str] attribute instead of a deprecated, unused instance method that contradicted the str every subclass and FindMixin already assumed — clearing the reportAssignmentType suppressions this forced across ~20 subclasses.
• bfabric logging now follows loguru's library convention: bfabric emits no log records by default (logger.disable("bfabric") at import), including the version banner previously logged on every client construction. The CLI re-enables it in setup_script_logging; other apps opt in with logger.enable("bfabric").
• Input resolver replaces its hand-maintained issubclass ladder with a dict[type, resolver] registry, an exhaustiveness check validated against the InputSpecType union at construction, and an MRO-walk fallback. Drops dead spec surface (resolve_filename stubs, unused BfabricAnnotationSpecField) and renames bfabric_resource_dataset.py to the _spec convention.
• uv-based commands raise a clear error when the uv binary is missing on PATH instead of passing None into a subprocess call.

Non-obvious decisions

• Demoting the banner to DEBUG was insufficient — loguru's default handler emits DEBUG too, so it only silenced the banner in apps that raise the level, and hid it in the CLI. logger.disable/enable fixes it properly. Note this makes all bfabric loguru logging opt-in for library/server consumers, not just the banner.
• The ~547 removed basedpyright baseline lines are genuine fixes (underlying errors no longer occur), not new suppressions; two entity subclasses keep a targeted reportUnannotatedClassAttribute ignore.
• The resolver keeps issubclass-based dispatch (via the MRO-walk fallback) to stay correct if BfabricAnnotationSpec later becomes a union.

Testing: full pytest for bfabric (748), bfabric_scripts (66), bfabric_app_runner (319) pass; basedpyright clean for all three, no baseline drift; ruff clean. Pre-existing dispatch_resource_flow failures (Python 3.14 / pandera) are unrelated — identical on clean origin/main.

🤖 Prepared with assistance from Claude Opus 4.8 via Claude Code.

[github-actions]

📝 "TODO" Changes Detected

Summary: ✅ 1 "TODO" removed

✅ Removed "TODO"s (1)

• bfabric_app_runner/src/bfabric_app_runner/specs/inputs/file_spec.py:109: # TODO delete

---

This comment is automatically updated when "TODO" changes are detected.

[leoschwarz]

Superseded by #592 and #593.

This branch was cut before 26 commits landed on main and is not worth rebasing: three of its ten commits (the app-runner FindMixin migration, the EntityReader class-form endpoint inference, and the entities_by_id/present_entities helpers, the last superseded by EntityResult.by_id/.present) are already on main in some form, and the utils/console.py move was reverted within the branch itself.

Redone against current main:

• Clearer failures for api inspect and a missing uv, plus core diagnostics cleanup #592 — the api inspect non-SUDS guard, the _uv_bin fail-fast, the loguru/BfabricRequestError conventions and their AGENTS.md write-up, plus the dead spec surface
• Dispatch app-runner input resolution through a spec registry #593 — the registry-based input Resolver

Not carried over: the logger.disable("bfabric") library-silence change (a behaviour change for library consumers, and the cause of this branch's intermittent logot failures under random test order) — worth doing, but as its own PR with a decision behind it.