Once buttons and tables become structurally readable, it is easy to describe the tool as able to read anything. Unexposed structures, bounds and unsupported frames remain. A successful read needs an explanation of how much material the result actually covers.
An empty control list does not establish an empty page
A screen may contain material even when no controls are returned. It may still be loading, contain only text or use an input surface outside current recognition rules. An empty array cannot establish that the site is broken.
The observation can retain bounded page-health metadata when controls are absent, including root display state and DOM counts. These clues guide further inspection rather than diagnose all screen content. An empty result is material for narrowing a cause, not an automatic success-or-failure verdict.
Open shadow roots differ from closed structures
The roots function follows accessible shadow roots to collect candidates. Reading an open structure does not establish equivalent access to every closed shadow tree. Meaningful controls cannot be invented from content the DOM route cannot establish.
Accessibility diagnostics may identify some inputs, but turning that clue into a stable action reference is another problem. Diagnostic and ordinary observation paths have different meanings. A supporting clue must not become a broad promise of complete closed-structure support.
Capturing pixels and interpreting controls are different outputs
A canvas document or chart can appear in an image without exposing buttons, rows and cells like ordinary DOM. Saving a screenshot differs from deriving a precise operation target from its pixels. An image file does not supply the same name-and-role contract.
Later capture support in Northstar cannot erase that distinction. A screen or region capture needs file and image-scope checks. Claims about reading and operating every canvas editor would require separate reproductions and execution evidence. Different capabilities inside the same product still have different success criteria.
Frame and content bounds belong in the support statement
Frame traversal has count and depth bounds, while elements, text and tables have separate limits. A valid bounded reply cannot establish complete page coverage. Unsupported frames retain reasons, and unread material must remain distinguishable in later decisions.
A frame excluded by authentication or public-destination policy is not automatically bypassed. The original request and access boundary remain. Unable to read can be a necessary result. Relaxing protections merely to fill every gap would expand apparent coverage while losing the original contract.
Keep diagnostics limited to structural evidence
The control probe reports bounded tags, roles, dimensions, display properties and ancestor state for selected accessibility inputs. It excludes sensitive categories and is not a diagnostic collector of values, body text, names or destinations. Input and frame counts are limited too.
I read the actual probe, observation and frame code and relevant fixture and policy tests. This is not a new execution review of every nonstandard screen. Having diagnostic evidence, repairing its cause and verifying real input on the site remain separate states.
Explicit boundaries make the next improvement concrete
Documenting limits does not diminish a capability; it identifies the next reproduction and completion criterion to test. A description of reading routes, bounds and unsupported structures is more useful for real integration than an unqualified all-sites claim.
A new editor should first be compared with existing general rules, then receive a narrow adaptation for the missing structure. A passing small case cannot become universal compatibility. Retained boundaries let later articles and modules explain precisely what improved and what still remains.