test(db): migrate OpenMemory's schema once per process and copy it

[francisrath]

Summary

db.OpenMemory() ran all 95 migrations for every test. Under -race, modernc's transpiled SQLite made that ~2.5s of setup per test, which is why the race shards were at 840–847s of a 900s budget on CI and timed out outright on Apple Silicon. OpenMemory now migrates a template once per process, serializes it, and deserializes a copy into each new single-connection :memory: database. Production Open is unchanged.

Closes #2982

Implementation notes

• memoryTemplate() runs inside a sync.Once: open, migrate, Serialize(). Any error is cached and returned to every caller, the same as a migration failure today.
• openMemoryConn() is the old open path (DSN pragmas, SetMaxOpenConns(1), setPragmas) pulled out so the template and the copies are opened the same way.
• Serialize/Deserialize are reached via sql.Conn.Raw through a two-method interface. No new dependency, and go.mod is untouched.
• Connection-level pragmas (foreign_keys, busy_timeout, …) belong to the connection, not the database image, so they survive the deserialize. The new test checks foreign_keys.

Scope decisions

• Isolation is the same as before: each call returns its own database, pinned to one connection, with exactly the state migrate produces (schema, seed rows, schema_migrations).
• Migration tests didn't need a separate path. They delete a schema_migrations row and call migrate() on the OpenMemory database, which still runs the real code. No test swaps migrationsFS or any migration input.
• No changelog fragment: nothing user-visible changed.

How it was verified

The change is about speed, so the evidence is timings rather than a test that fails on main. All runs use GOTOOLCHAIN=go1.26.8 on an Apple Silicon Mac:

Run	main	This branch
go test -race -run '^TestDownloadClient' ./internal/api	157s	6.6s
make test-race, internal/api ^Test[A-B]	>15m (timeout)	61s
make test-race, internal/api ^Test[C-K]	>15m	57s
make test-race, internal/api ^Test[L-Q]	>15m	48s
make test-race, internal/api rest	>15m	30s
make test-race, internal/db	>15m	78s

On CI, the validate (Go race) shards of this PR's run compared with a recent successful CI run on another PR (37187253010):

Shard	Recent PR run	This PR
api-a-b	934s	116s
api-c-k	857s	135s
api-l-q	1305s	103s
api-rest	660s	105s
db	1403s	95s
non-api	1757s	169s

The non-api shard drops too, because other packages also call OpenMemory. codecov/patch reports 54%: the uncovered lines are the error branches (template migration, serialize/deserialize, and a driver conn without Serialize), which would need a fake driver to reach.

TestOpenMemory_CopiesAreIndependent checks the copy semantics: a row written to one OpenMemory database doesn't appear in the next, foreign_keys is still on, and every migration version is recorded.

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

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

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 78.37838% with 8 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/db/db.go	78.37%	4 Missing and 4 partials ⚠️

📢 Thoughts on this report? Let us know!

[vavallee]

Thanks Francis, nice find. The race shards timing out had been a pain for a while and this is a small, tidy fix.

I merged current main into it (migrations 099 and 100 included) and checked a few things. A freshly migrated database and an OpenMemory copy have identical schema and rows, 32 parallel copies under race stay isolated, and 20 opens under race went from 259s to about 16s on my loaded box. Full suite passes.

One small nit, not blocking: synchronous and cache_size are per schema, so they reset on deserialize (they come back as 2 and -2000). Harmless for in memory tests, but the doc comment says the pragmas carry over. Happy to take a follow up that reruns those two after Deserialize or softens the comment.

Merging this one now, thanks again!