perf(storage): persist artifact metadata deltas

[likun666661]

Summary

• replace the snapshot-shaped Artifact metadata write contract with explicit upsert/delete changes
• keep each mutation atomic while hashing, serializing, and writing only changed records
• preserve the existing cross-process reload and recovery behavior
• add a trigger-backed regression test proving an unchanged row receives no SQL write

This removes the O(N) DELETE plus re-insert cycle from every Artifact mutation and bounds metadata WAL traffic to the changed rows.

Fixes#4037
Refs #4704

Verification

• npm run clean
• npm --workspace @maka/core run build
• npm --workspace @maka/storage run build
• NODE_NO_WARNINGS=1 npm --workspace @maka/storage run test:dist — 1107 tests, 1099 passed, 8 skipped, 0 failed
• npx biome check on all five changed files
• git diff --check

NODE_NO_WARNINGS=1 suppresses the Node 22 experimental SQLite warning; without it, the strict child-process authority test treats that warning as unexpected stderr.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: OpenAI Codex audited the write path, implemented the delta protocol and regression test, and ran verification. The commit includes the required Generated-by: OpenAI Codex trailer.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — Artifact metadata mutations now persist only changed rows while retaining atomic commit semantics.
• No

[me2seeks]

Verified the delta protocol end to end, not just the diff:

• All six metadata mutation sites (create, revive, both delete paths, purge, orphan adoption) now pass exact deltas; no replaceAll callers remain, and the interface has a single production implementation.
• storage_key is the PRIMARY KEY and every compared column is NOT NULL, so the ON CONFLICT ... WHERE ... IS NOT excluded.* guard is valid and NULL-safe.
• Independently re-ran the upsert semantics against node:sqlite: identical re-run produces zero SQL writes, a changed row updates exactly once — matching the trigger-backed regression test's claim.
• The delete no-op early-return and atomic transaction wrapping are preserved.

One deliberate trade-off worth naming (non-blocking): replaceAll used to incidentally repair any DB↔memory drift on every mutation; delta writes only heal the rows they touch. Given the writer lock plus readAll on open/reload, the drift window is crash-only and the upsert path (insert-if-missing) still heals the touched row. Acceptable.

This removes the O(N) metadata rewrite that dominated the WAL amplification measured in #4704.