feat(memory): define capacity contract for long-lived Memories

[AlexStocks]

Context

#1709 removed the projection write amplification reported in #1321 by updating active-head/search projection rows incrementally instead of rewriting every projection row on each append.

That PR intentionally did not define the remaining long-lived Memory capacity contract. #1321 was closed by the merge, but several parts of the original expected behavior still need a separate design and implementation track.

Pagination for list_memory_entries is already tracked separately in #1656. This issue should coordinate with that work but should not duplicate it.

Remaining problem

A long-lived Memory can still grow without an explicit, observable capacity boundary:

• each Artifact Revision still stores a complete manifest directory;
• inactive entries remain in the manifest as tombstones;
• revision/history growth has no retention or compaction contract;
• there is no entry-count or manifest-byte budget;
• there is no defined behavior for routing/splitting when a Memory reaches a practical limit;
• scale regression coverage is still focused on projection writes, not the full capacity envelope.

Scope

Define and implement the remaining Memory capacity contract after #1709.

At minimum, cover:

• Current-manifest entry-count budget.
• Current-manifest byte budget.
• Stable over-limit behavior: explicit error, automatic routing to another Memory, or split.
• Split/routing semantics, including artifact identity, citation behavior, and search across split Memories.
• Inactive/tombstone compaction semantics.
• Revision/history retention or bounding semantics.
• Interaction with MemoryIndex.delete() / upsert() and projection rebuilds.
• SQLite and OceanBase behavior documented separately.
• Scale regression benchmarks beyond projection statement count.

Acceptance criteria

• The Memory layer has a documented capacity contract for long-lived Memories.
• Over-limit behavior is deterministic and observable through public APIs.
• Split/routing or stable failure behavior is covered by tests.
• Tombstone compaction defines what happens to old citations and stale references.
• Scale tests record at least:
  • entry count,
  • manifest bytes,
  • database bytes,
  • append latency,
  • final-window append latency,
  • projection row writes,
  • search behavior after split/compaction if applicable.
• SQLite and OceanBase validation results are reported separately.
• The implementation does not regress the incremental projection write guarantees added by perf(memory): update search projections incrementally per revision #1709.

Related work

• Follow-up to bug: Memory append storage and latency grow superlinearly with entry history #1321.
• perf(memory): update search projections incrementally per revision #1709 fixed projection write amplification but not the complete capacity contract.
• feat(memory): add bounded cursor pagination for Memory entries #1656 tracks bounded cursor pagination for Memory entries.