Cache test database schema in CI to reduce compute time

[dannyroberts]

Product Description

N/A — CI-only change, no user-facing effects.

Technical Summary

Uses GitHub Actions cache to store a pg_dump of the test database schema. On cache hits (~every run without migration changes), restores the dump in seconds instead of running the full ~200s setup_databases() call via Django test framework.

How it works:

1. New actions/cache@v4 step caches artifacts/test_db.dump, keyed on hashFiles('**/migrations/**/*.py')
2. Before tests: if dump file exists (cache hit), start postgres, restore the dump, skip schema creation
3. After tests: if no dump existed (cache miss), dump the DB for next run's cache
4. On restore failure: delete the dump and fall back to normal setup

Scope: 3 non-sharded python test shards (05, 6a, bf). The python-sharded-and-javascript shard is excluded (different DB config, only 1 job).

Expected savings: ~600s total CI compute per run (3 shards × 200s) on cache hits, with no wall-clock impact.

Feature Flag

N/A

Safety Assurance

Safety story

• CI-only change — no production code modified
• Restore failure is handled gracefully: deletes the dump file and falls back to normal 200s schema setup
• Dump failure is silently ignored — next run is just a cache miss
• Cache key invalidates on any migration file change, so schema drift is not possible
• The existing --reusedb=1 fast path in reusedb.py already handles the "DB already exists" case

Automated test coverage

The CI tests themselves serve as validation — if the restored schema is wrong, tests will fail.

QA Plan

1. First CI run: cache miss, all shards do normal setup, dump is cached
2. Push a no-op commit: cache hit, verify tests still pass and --durations output shows no ~200s setup time
3. Verify python-sharded-and-javascript shard is unaffected

Rollback instructions

• This PR can be reverted after deploy with no further considerations

Labels & Review

• Risk label is set correctly
• The set of people pinged as reviewers is appropriate for the level of risk of the change

🤖 Generated with Claude Code

[dannyroberts]

I don't love how much boilerplate we need in order to do this, but first I just want to see if it'll even save any time at all. If it does, then I'll think about how to make this cleaner / more concise