Correctness hardening: explicit write-failure surfacing + atomic log cap under contention

[mark-e-deyoung]

Summary

Track deferred correctness work for Group B.

Problem

Several file/log write paths intentionally swallow exceptions, which can mask data-loss conditions. Log-size capping is checked before append, but not atomically under lock, so concurrent writers can overshoot configured caps and fail silently.

Scope

• Replace silent exception swallowing in critical write paths with structured error surfacing (API error, lifecycle event, or error telemetry), while retaining best-effort semantics only for non-critical telemetry.
• Make append_trace_event cap enforcement atomic under file lock.
• Add dropped-event accounting (counter and/or metric) when cap reached.
• Add concurrency tests for append/cap correctness under parallel writers.

Acceptance criteria

• No critical write path silently discards failures without a surfaced signal.
• Log caps are enforced deterministically under concurrent writes.
• Tests cover contention path and dropped-event reporting.

Notes

Deferred intentionally to keep current change focused on Group A race/idempotency fixes.