Add position-independent diagnostic syntax to syntax tests

[tilladam]

Summary

• Add //-error{message} and //-warning{message} syntax for matching diagnostics by message and level only, ignoring source position
• Useful for layout-related binding loop warnings where the diagnostic position varies across platforms due to analysis traversal order differences
• Includes self-tests verifying the new syntax works (positive match, wrong message, wrong level)

This is needed by #10552 (binding loop detection for callbacks), where several syntax tests fail on CI because layout binding loop warnings are emitted at platform-dependent source locations.

Test plan

• self_test passes — 3 new assertions (positive, wrong message, wrong level)
• syntax_tests passes — no regressions in existing tests
• CI passes

[ogoffart]

I'm curious what makes the position change across platform.
Is there something not deterministic in the diagnostics?
(The #10552 PR fails currently because some loop message have different order.)

What i have noticed is that sometimes diagnostics are not attached to a location and a file and this feature could help, although i don't know if it would work as is because there is also not a file.

[tilladam]

The non-determinism comes from visited: HashSet<PropertyPath> in AnalysisContext (binding_analysis.rs:161). It uses Rust's default RandomState hasher, so the hash seed differs between runs/platforms.

While visited is never iterated, it affects control flow at line 366: if !context.visited.insert(current.clone()) { return }. This early-return determines which traversal path through the binding graph discovers the cycle first. For layout-related cycles (which involve implicit properties like layoutinfo-h, width, layout-cache), the cycle can be entered from multiple directions. Different hash seeds → different traversal order → cycle detected from a different entry point → the currently_analyzing stack has the same properties but in a different order → diagnostics are emitted at different source positions.

Additionally, at line 346, is_in_binding_loop.replace(true) causes the diagnostic emission loop to break early if a property was already marked from a previous traversal. Different traversal orders can cause different properties to be pre-marked, which changes how many diagnostics are emitted and where.

For the case of diagnostics without a file — the current implementation requires matching against a specific file's source, so position-independent matching within a file wouldn't help with that case. That would need a separate mechanism.

[LeonMatthes]

@tilladam Then IMO the better solution is to not use HashSet and use a BTreeSet instead, which is deterministic.
A non-deterministic compiler is a bad idea anyhow, and is causing us issues, like unnecessarily re-writing files that haven't really changed.

[tilladam]

You are right, droppign this one in favor of #11219

[tilladam]

After investigating @ogoffart's point further, I agree that the visited HashSet explanation doesn't hold up — since visited is only used for membership checks (insert returns a bool), its internal hash ordering doesn't affect control flow.

I did find a real source of non-determinism: Element::property_analysis is a HashMap that gets iterated in several passes (inlining, move_declarations, etc.) before binding analysis runs. #11219 fixes this by switching it to BTreeMap.

However, I can't confirm locally that this is the specific cause of the cross-platform diagnostic position differences in #10552 — that would need CI validation. So the position-independent diagnostic syntax proposed here may still be needed if other sources of non-determinism remain.