fix: Omit non-symbol context attributes and protect against cyclic values

[kinyoklion]

Summary

Attribute references store every path component as a symbol, so a context attribute with a non-symbol name (for example "street" => ... instead of street:) can never be addressed: it is invisible to evaluation and bucketing, and a private-attribute reference can never redact it. Today the event filter still copies such attributes into analytics event payloads, with one extra defect: a top-level string-keyed attribute is emitted with a null value, because the attribute name comes from raw storage but the value is fetched through get_value, which symbolizes the name and misses the entry.

This PR makes ContextFilter:

• Omit attributes with non-symbol names, at the top level and at any nesting depth (including the whole subtree under a non-symbol key). Omitted attributes are not reported in _meta.redactedAttributes, since no reference removed them. This also removes the top-level null emission described above.
• Log an error the first time a non-symbol name is dropped (once per filter instance, guarded by an AtomicBoolean since the formatter runs on flush workers), so the condition is discoverable without log spam.
• Stop recursing into cyclic values. A self-referential hash in a context attribute previously crashed event serialization with SystemStackError. The filter now tracks the chain of hashes from the attribute root down each branch and omits a value that points back into its own chain. The same object appearing in different branches (a DAG) is still copied everywhere it appears. One documented nuance: context creation copies each top-level attribute value, so a cycle through that value appears once in the output before it is cut, and private attribute references match the paths that appear in the output (covered by a pinned spec).

Behavior change note: string-keyed attribute data that was previously sent in event payloads (intact when nested, null at top level) is now omitted.

Performance: same-process A/B against main on representative context shapes shows 0.99-1.05x on flat and redact-all paths and ~0.90x on deeply nested contexts (the cycle-protection chain costs one small array per nested hash).

Regression tests cover: top-level and nested omission, subtree omission, no redactedAttributes reporting for omitted names, composition with private-attribute redaction, log-once behavior, the previous SystemStackError on cycles, and DAG preservation. All new tests fail against the previous implementation.

---

Note

Overview
Context filtering for analytics events now drops context attributes whose names are not symbols (top-level, nested, or entire subtrees under a string key), fixing the prior bug where top-level string keys could appear as null in payloads. Those keys still cannot be targeted by attribute references or private-attribute redaction, so they are no longer listed in _meta.redactedAttributes.

When a logger is supplied (from EventOutputFormatter via config.logger), the filter logs a single error the first time a non-symbol name is skipped.

Nested hash redaction also tracks object identity along each branch and skips edges that cycle back into the same chain, avoiding SystemStackError during serialization while still emitting shared objects (DAG) under separate branches. Private redaction continues to apply to paths that remain in the output, including partial cyclic structures.

^{Reviewed by Cursor Bugbot for commit 4639d68. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit c6de577. Configure here.}