[Epic] Record encoding & serialization correctness

Correctness of how Harper encodes and decodes records — the msgpackr/cbor-x/structon layer and the call sites that reach it.

## Scope

**In:** record encoding and decoding, shared-structure dictionaries, content-negotiated serialization of responses, and the guards that decide whether a byte sequence is a valid record.

**Out:** transaction-log framing and audit-entry corruption (#1651) — those are storage-layer concerns even when they surface as decode failures. Replication's handling of records it cannot decode (harper-pro#537) — that's a containment question in the replication path, not an encoding one.

## Why

These defects share a failure mode that makes them worth grouping: an encoding ambiguity produces *plausible* output rather than an error, so the damage is silent and shows up far from its cause. #1762's own title notes the ambiguity "keeps resurfacing", which is the signature of a class rather than a bug.

_Created during backlog triage. Child issues keep their own priorities and milestones._

No source files or tests are named in the issue. Start by reviewing the child issues and referenced #1762, then trace the msgpackr/cbor-x/structon call sites and content-negotiated response paths. Done means the scoped encoding ambiguities and invalid-record guards are addressed without taking on the excluded transaction-log or replication work.

Written by the indexing model from the issue text.