Fix a fatal error when binding non-scalar values, and static analysis issues, in WP_Interactivity_API - #12725

Static analysis work on WP_Interactivity_API (Core-64898) surfaced a fatal error when data-wp-bind binds a non-scalar value to an HTML attribute (Core-65740). This fixes both, and closes the neighbouring gaps that fix brought to light.

The fatal error

Any non-scalar value reaching WP_HTML_Tag_Processor::set_attribute() hit a TypeError, because that method is typed string|bool and passes the value straight to the escaping functions. Both escaping paths failed, in different places:

This is reachable from post content alone, with no plugin code, because data-wp-context can supply the array inline. Paste either of these into the post editor's code editor (⌥⌘M) to reproduce on trunk:

<!-- wp:accordion -->
<div class="wp-block-accordion"><span data-wp-interactive="myPlugin" data-wp-context='{"list":[1,2,3]}' data-wp-bind--id="context.list">Boom</span></div>
<!-- /wp:accordion -->

<!-- wp:accordion -->
<div class="wp-block-accordion"><span data-wp-interactive="myPlugin" data-wp-context='{"list":[1,2,3]}'><a data-wp-bind--href="context.list">Boom</a></span></div>
<!-- /wp:accordion -->

The core/accordion wrapper is needed only because directives are processed for blocks whose supports.interactivity is true; any of those works. The directives have to sit on a nested element, since the accordion's own render filter overwrites data-wp-interactive and data-wp-context on its wrapper.

data_wp_bind_processor() now rejects such a value with _doing_it_wrong() and leaves the attribute unset (removing it if it was already present, consistent with how a null value is handled), rather than taking down the whole page render.

Objects

The interesting case is objects, since the value bound on the server is also serialized into the client store and re-evaluated during hydration. Those two have to agree, and only the value the client receives can decide it — so an object is resolved by round-tripping it through wp_json_encode()/json_decode(), which is exactly how that store is built. Agreement is then structural rather than checked.

Rows 1 and 3 rendered a value the client then overwrote with [object Object], since PHP coerces __toString() for the escaping functions' string parameters but JSON has no equivalent. Row 4 fataled even though the client store was already correct, purely because PHP cannot coerce an object without __toString(). Row 5 is why the resolution round-trips rather than calling jsonSerialize() once: a single call stops after one level, whereas the encoder keeps going, so the server rejected a value the browser was about to accept.

Resolving the object before the null check also means one serializing to null removes the attribute quietly, and the rest composes: a serialized number is formatted as below, a boolean keeps the boolean-attribute semantics.

Note that __toString() is never consulted. Row 3 is rejected because the object's own JSON opinion is what the browser will see.

Numbers

Numbers are formatted by the same encoder rather than cast to string, which closes two more divergences from the client. Measured on PHP 7.4:

Casting a float is locale-dependent before PHP 8.0 (RFC), and 7.4 is the minimum supported version. The cast also rounds to precision where the store uses serialize_precision. Routing both through one function makes them agree whatever those settings are. Integers were never affected, as only floats consult the locale.

Neither is a regression from the fix above — the cast is the same conversion set_attribute() was already applying implicitly, one frame later — but both are worth closing while the line is being touched.

INF and NAN

These are scalars, so they passed the check above, but JSON can represent neither. wp_json_encode() on the store returns false, and WP_Script_Modules::print_script_module_data() casts that failure to a string, so the client is served an empty <script type="application/json"> in place of all of its state: every binding on the page loses its value, not only this one. The attribute meanwhile received INF or NAN, which means nothing as an attribute value and does not match the client either, where the same numbers render as Infinity and NaN.

They are now rejected too, with their own message, since describing them as non-scalar would be wrong. Rejecting the binding only draws attention to the problem; removing the value from the state is what resolves it.

Static analysis

The rest is PHPStan level 10 work on the same code path, which is how the missing type checks came to light:

• WP_Interactivity_API::$directive_processors gains an array shape whose values are the literal method names. Without it, the dynamic dispatch through call_user_func_array() is unresolvable, so all eight data_wp_*_processor() methods were reported unused and the callable array was reported as not callable. Ten errors, no ignores.
• get_directive_entries() declares its return type and array shape, and now skips a directive name it cannot parse instead of destructuring null. The get_attribute() null check is defensive: the attribute names come from get_attribute_names_with_prefix(), so it should not be reachable.
• parse_directive_name() documents its shape and guarantees a non-empty prefix. The character check became an anchored whitelist, and a name beginning with -- no longer yields an empty prefix. Neither changes the outcome for a directive that was previously matched — an empty prefix matched no known directive either way — but the return type is now honest.
• extract_directive_value() documents its tuple shape.
• WP_HTML_Tag_Processor::$attributes and get_attribute_names_with_prefix() gain @phpstan- annotations only. No functional change.
• data_wp_bind_processor() gains a void return type.

Testing instructions

npm run test:php -- --group interactivity-api

New coverage in tests/phpunit/tests/interactivity-api/wpInteractivityAPI-wp-bind.php. Non-scalar values are exercised as a cross product with the attribute bound, so the strtr() and esc_url() escaping paths each fail on their own data set rather than one masking the other, and each asserts the exact _doing_it_wrong() message rather than only that it was triggered. There is also a test asserting that the rendered attribute value appears verbatim in the JSON sent to the client, which is the invariant the object handling rests on.

The locale test skips itself where no comma-decimal locale is installed, so it will not fail a runner without one.

To confirm the tests would catch a regression, remove the is_scalar() block from data_wp_bind_processor() and re-run: 32 errors and 8 failures across the 40 data sets.

Verified against PHP 7.4 as well as 8.3, since the float formatting behaves differently between them.

Follow-ups, not addressed here

• state() is the better place to catch a non-finite float. Binding is only where you happen to notice a store which is already broken, and one such value anywhere in the state discards the whole of it.
• _process_directives() restores $context_stack and $namespace_stack only on the unbalanced-tags path, not when an exception is thrown mid-traversal. A caught fatal therefore corrupts every later process_directives() call on the same instance. Moot for this bug now that the throw is prevented, but it is a real fragility.
• The is_array() branch in the directive dispatch in _process_directives() is unreachable: nothing writes to $directive_processors, so no value is ever an array callable.

Trac ticket: https://core.trac.wordpress.org/ticket/65740
Trac ticket: https://core.trac.wordpress.org/ticket/64898

Use of AI Tools

AI assistance: Yes
Tool(s): Claude Code
Model(s): Opus 5
Used for:

• Troubleshooting PHPStan complaining that the data_wp_bind_processor method was unused. See 1638ca3.
• Writing tests.
• Resolving a bound object to whatever it serializes to, and the wording of the _doing_it_wrong() messages.
• Implementing and verifying the number formatting. Using the JSON encoder for it was my own approach, prompted by @dmsnell raising the locale question in review.
• Rejecting INF and NAN, after I asked whether they belonged with the other values which cannot reach the client.
• Drafting this description. The behavior tables and the reproduction snippets above were verified against a local install rather than asserted.

This Pull Request is for code review only. Please keep all other discussion in the Trac ticket. Do not merge this Pull Request. See GitHub Pull Requests for Code Review in the Core Handbook for more details.