refactor(core): centralize session message rows

[kitlangton]

What

Centralizes the persisted SessionMessage representation boundary in SessionMessageRow instead of repeating { id, type, data } assembly and splitting throughout session persistence.

How

• Adds packages/core/src/session/message-row.ts with effectful decode, synchronous decode, and synchronous encode operations.
• Migrates message persistence in the store, history, pending promotion, revert planning, projector, and transfer import paths.
• Keeps transaction handling, sequence numbers, timestamps, session ownership, and decode error policy at each call site.
• Keeps session_pending row codecs and fork row copying separate because those operations do not encode or decode the canonical message-row representation.
• Adds focused tests for representation round trips and canonical column precedence over stale JSON keys.

Scope

This is representation-only refactoring. It does not change the session_message schema, durable event behavior, pending input encoding, fork batching, or caller-specific error handling.

Testing

• bun run test test/session-message-row.test.ts test/session-projector.test.ts test/session-create.test.ts test/session-prompt.test.ts from packages/core: 86 passed
• bun typecheck from packages/core
• Push hook: typechecked all 39 workspace packages
• Prettier check on all changed files
• git diff --check

[Enough1122]

AI code review — automated review for reference; please use your judgment.

• packages/core/src/session/message-row.ts — clean extraction; the module follows the repo's self-export namespace pattern, and decode/decodeSync/encode cover all three call-site shapes (effectful, sync, encode). Canonical-column precedence ({ ...row.data, id: row.type && row.id } spread order) is preserved exactly.
• message-row.ts:17 — encode re-wraps with SessionMessage.ID.make(id) after encodeMessage already produced a branded id; harmless but slightly redundant — worth a one-line comment if intentional.
• Migration across history/pending/projector/revert/store/transfer is mechanical and behavior-preserving; transaction handling, seq, and timestamps stay at call sites as described. Fork/pending-table codecs correctly left alone since they don't use the canonical representation.
• Tests pin both round-trip and stale-JSON-key precedence — exactly the invariants this boundary owns. Representation-only claim holds up against the diff.
• Heads-up: PR currently conflicts with dev (mergeable_state: dirty) — rebase needed. No blocking issues found.