[eli-r-ph]

💡 Motivation and Context

v2 drops an event that holds a NaN or ±Inf property and reports it to Callback.Failure. posthog-go 1.36.1+ (#382) and the other SDKs send these values as null and keep the event, as sdk-specs#140 proposes. The fix in #382 normalizes and re-validates every event, which slows normal events, so v2 left it out.

This PR ports #382's normalizer and its tests (by Manoel), but runs the normalizer only after go-json fails on a non-finite float64.

• prepareForSend encodes each event once, as before. On a non-finite number error only, it copies the properties with those numbers set to null and encodes again.
• Any other encoding error still drops the event and reports it to Callback.Failure.
• event_properties_json.go is fix: preserve events with non-finite numeric properties #382's file plus marshalEventPayload, the one-pass entry point.
• The tests port fix: preserve events with non-finite numeric properties #382's tests to v2. New tests check that a finite event is encoded once, that the slower path changes only the non-finite values, and that a failing event is dropped without affecting its neighbor.
• The migration guide and a changeset describe the behavior and its known gap.

Known gap, unchanged from the v1 branch: go-json writes a non-finite float32 as a bare NaN/+Inf token without an error. It also passes some invalid custom MarshalJSON output, such as {"value":01}. If no non-finite float64 sends the event down the slower path, the request is invalid JSON and capture rejects it. #382 fixes this as a side effect of re-validating every event.

I tried a cheaper check: a byte scan of the encoded event for a bare NaN/Inf value token. On a 1 KB event it cost about 0.7% of encode time. On a 450 KB AI event it cost 26–78% of encode time, depending on the text, so I left it out. Even a single IndexByte pass costs about 4%.

💚 How did you test it?

• go build ./..., go vet ./..., gofmt -l . and go test -p 1 ./... pass. The new tests also pass with -race and with -count=100.
• With prepareForSend changed back to a plain json.Marshal, 10 of the new and ported tests fail.
• Benchmark of BenchmarkPrepareForSendFinite against the v1 base on an Apple M4 Pro (-count 10, benchstat):

Event	Base sec/op	This PR sec/op	Base B/op	This PR B/op	allocs/op
typical, 871 B	5.757µs ± 2%	5.848µs ± 4% (~, p=0.218)	7.350Ki	7.348Ki	30 → 30
AI, ~450 KB	254.8µs ± 2%	254.0µs ± 3% (~, p=0.481)	504.3Ki	511.5Ki (~, p=0.075)	34 → 34

The allocation count of the typical event is 29 or 30 per process on the base too.

An event with a non-finite value costs about 2.2x (typical) and 5.3x (AI) to encode. Only those events pay this.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed.
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran pnpm changeset to generate a changeset file

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

Written with Claude Code (Claude Opus 5.5). The design comes from the person driving the work: encode strictly first, as posthog-python does, and fall back to a sanitized copy only on failure.

The non-finite section of the migration guide is also added by #390, which says that v2 drops these events. Whichever PR merges second must keep this PR's wording.

🤖 Generated with Claude Code