feat(mcp): bound MemoryDelta results with a limit

[radimsem]

Summary

Bounds MemoryDelta so a large vault can't blow the LLM context window or overrun the MCP transport in one call — it was the only read tool without a result cap (Search=50, FetchBatch=256, Related=100).

Closes #221.

What changed

• limit param on DeltaInput (default 500, ceiling 5000, over-asks clamp rather than error). Plumbed HandleDelta → Engine.Delta → GetDiffsSince → qSelectDiffsSince (LIMIT ?).
• Snapshot-granular cursor preserved (since_snapshot stays a snapshot id — no client contract change). Engine.Delta peeks limit+1 and trims the trailing partial snapshot, so resuming from the hinted since_snapshot is lossless.
• Tiered note: continuation hint (returned as text, since results are text-only):
  • boundary truncation → Re-call MemoryDelta(since_snapshot=<last complete id>)
  • a single snapshot below the ceiling → raise limit (max 5000)
  • a single snapshot above the ceiling → exact MemoryDiff(from_snapshot_id=…, to_snapshot_id=…) escape hatch (that snapshot is provably the immediate-next one, so the diff is scoped to it alone — no inaccessible tail).
• Non-positive limit guarded (exported Engine.Delta returns empty instead of panicking).
• §10 contract docs updated: skills/remind/references/snapshots-diffs.md, docs/versioning.md.

Tests

TestDelta_Limit (fit / trim / boundary / overflow / non-positive guard), TestDeltaTruncationNote (all three note branches), TestHandleDelta_LimitTruncationNote. Full suite, make check-skills, gofmt, and go vet all green.

Review

Project reviewers (go-style-reviewer, mcp-surface-reviewer) clean; Codex review converged over two rounds (caught the >5000-diff inaccessible-tail regression and a non-positive-limit panic, both fixed).

🤖 Generated with Claude Code