fix(telemetry): binary-search Truncate drop count to avoid O(n) re-renders (#153) - #154

Fixes #153.

Problem

telemetry.Truncate drops operation Result fields (step 1) and whole operations (step 2) one at a time, re-rendering the entire record to JSON after every single drop (fits → render). For an invocation with tens of thousands of operations that is O(n) full renders of a multi-megabyte record, inside OnInvocationEnd:

• A 29,999-operation invocation finished its work in ~22 s, then stalled in truncation to the 900 s Lambda timeout and failed with ExecutionLimitExceeded (the same handler without the plugin succeeds in ~25 s). A CPU profile of a 5,000-operation run showed ~79 % of time in Truncate → fits → json.Marshal.
• A related consequence: Truncated/DroppedOperations were set after the fits check, so the final record could land a few bytes over the cap — the reported 262,118-byte event exceeded DefaultMaxRecordSizeBytesLog and was cut by CloudWatch Logs so it no longer parsed.

Fix

Dropping more content only shrinks the rendered record, so "fits after dropping the oldest k" is monotonic. Replace both one-at-a-time loops with sort.Search, cutting thousands of renders to O(log n) (~15). Each candidate is sized with the final Truncated/DroppedOperations metadata already applied, so the returned record accounts for those bytes and cannot creep back over the cap.

Drop order (results oldest-first, then whole operations oldest-first, then Input, then Output), the set of retained operations, and all truncation flags are unchanged — this is purely an algorithmic change, no exported API change.

Validation

• Added telemetry/truncate_dropops_test.go: asserts a 20,000-operation oversized record truncates to fit and the render count stays O(log n) (≤100). On the old code this test does not complete — the one-at-a-time loop renders many thousands of times (locally it exceeds a 2-minute timeout, mirroring the production stall); with the fix it finishes in well under a second.
• Full go test ./telemetry/... passes; go vet and gofmt clean.

Developed with AI assistance (Claude Code): the root cause from the issue was confirmed, the binary-search fix implemented, and the regression test written and run locally. I reviewed and verified the change and take responsibility for it.