test(db): reapply DSN pragmas after OpenMemory deserializes its template

[francisrath]

Summary

Follow-up to #2983, from @vavallee's review: synchronous and cache_size are per-schema, so Deserialize reset them on every OpenMemory copy (to 2 and -2000), while the DSN asks for 1 and -16000. OpenMemory now runs the DSN's pragmas again after loading the template, so a copy matches what a plain open gives.

Refs #2982

Implementation notes

• connectionPragmas() reads the pragma list out of connectionPragmaDSN, so the values live in one place and a pragma added to the DSN later is reapplied (and tested) without anyone remembering to.
• The test now checks every DSN pragma on a copy rather than only foreign_keys, and its comment no longer claims they all survive the copy.
• Production Open is unchanged.

How it was verified

The updated test fails on main for the reason in the review:

--- FAIL: TestOpenMemory_CopiesAreIndependent
    openmemory_test.go:43: PRAGMA synchronous = 2, want 1
    openmemory_test.go:43: PRAGMA cache_size = -2000, want -16000

and passes here. The extra pragmas cost nothing measurable: go test -race -run '^TestDownloadClient' ./internal/api is 4.0s.

Checklist

• Every commit carries a Signed-off-by that matches its author (git commit -s)
• Changelog fragment: N/A, test-only change
• Tests added or updated
• docs/DEPLOYMENT.md: N/A
• Wiki pages: N/A
• No new dependency

Test plan

• go test ./internal/db
• make test
• make test-race (all six shards green, no races reported)

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 57.14286% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/db/db.go	57.14%	2 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[vavallee]

Thanks, that's exactly the fix I had in mind, and reading the list out of the DSN means nobody has to remember to keep two lists in sync. OpenMemory's pool is a single connection so the reapplied pragmas land on the only connection there is. Merged current main into it locally (including the sync.Once driver fix from #3037) and go test ./internal/db -count=2 is green. The codecov patch miss is just the two error branches, fine by me. Merging.