testit v0.3.0 - per-backend submodules, and a working redpanda resource

The testit monolith is now a core module plus four per-backend submodules, mirroring the db/ v0.2.0 split: importing one backend no longer drags every other backend's driver.

• github.com/sonirico/vago/testit v0.3.0 - the backend-agnostic docker resource pool (Resource, Pool, DockerResourcesPool, RunSafe). Depends only on dockertest and go-snaps; lol is replaced by the minimal testit.Logger (lol.Logger satisfies it structurally). Breaking: retryFunc is now the exported RetryOp, and backend constructors moved out of this module.
• github.com/sonirico/vago/testit/postgres v0.1.0 - resource, suite and SQL fixtures.
• github.com/sonirico/vago/testit/clickhouse v0.1.0 - resource, suite, fixtures and the config.d asset.
• github.com/sonirico/vago/testit/redis v0.1.0 - resource and suite.
• github.com/sonirico/vago/testit/redpanda v0.1.0 - rewritten and actually usable: correct --kafka-addr flags (the old --Redpanda-addr spelling never started a broker), injectable image via WithImage, WithClusterConfig/SetClusterConfig applying rpk cluster config through resource.Exec, port reservation via net.Listen instead of deprecated rand.Seed, and a real, un-skipped produce/consume smoke test.

PR: #3

db v0.2.0 — per-backend submodules, and four correctness fixes

db compiled ClickHouse, MongoDB, Redis and the Elastic APM stack into every
consumer that only wanted Postgres. It no longer does.

Verified from a module outside this workspace: importing db + db/postgres
now pulls zero mongo, clickhouse or redis packages.

New modules

db/postgres · db/clickhouse · db/mongo · db/redis — all at v0.1.0.
Core db keeps the backend-agnostic contract (Querier, Handler, Tx,
the Executor family, Context, the bulk SQL builders, In, OrderBy,
NullJSON*) and has no driver dependency at all.

Correctness fixes

Each was verified red against the original code before the fix, not merely
green after it.

• QueryTx never rolled back. The executor callback always returned
  nil, so a failing fn returned its error to the caller and committed
  anyway. Anyone relying on it for atomicity did not have it. The same bug
  ran Do's after-commit hooks when fn had failed.
• ExecutorRWO.RW() returned an executor built on the read-only handler.
• Migrations could apply nothing and report success. LaunchPostgresql
  and LaunchClickhouse matched action against a switch with no
  default, leaving err nil from the preceding call. Any action that was
  not exactly "up" or "down" — a typo, a capital, an empty string —
  applied zero migrations and returned no error, so a deployment would run
  against a schema that was never touched.
• NewNoopRepoContext panicked on first use, handing back a querier
  built on a zero-value *sql.DB whose connector is nil.
• BulkUpdate/BulkUpdateSQL took an unexported parameter type, so
  these exported functions could not be called from outside the package.
  They now take BulkUpdatableRanger, mirroring BulkableRanger.
• ClickHouse HTTP responses leaked their body on every non-expected status.

Breaking changes

• Import paths. OpenPgx*, OpenPg*, OpenPostgres, NewExecutorPgx,
  NewRWOExecutorPgx, LaunchPostgresql → db/postgres. OpenCH,
  OpenClickhouse, NewClickHouseHttp, LaunchClickhouse →
  db/clickhouse. OpenMongo*, MongoParse* → db/mongo. OpenRedis,
  NewApmRedisHook, RedisConfig, RedisTLSConfig → db/redis.
  Symbol names are unchanged.
• db.ErrIsNoRows now checks only sql.ErrNoRows. The pgx-aware
  version is postgres.ErrIsNoRows. Keeping the old one in core would have
  kept pgx in core, which was the whole point.
• db.Logger replaces lol.Logger in NewExecutor,
  NewDatabaseSqlExecutor, NewRWOExecutor and the Bulk* functions. It
  is source-compatible: lol.Logger satisfies it structurally with no
  adapter, asserted at compile time rather than in prose.

One disclosed regression

db.Logger deliberately omits WithField/WithTrace. Go requires exact
signature identity for interface satisfaction, so a self-referential
WithField(...) db.Logger could never be satisfied by lol.Logger's
WithField(...) lol.Logger — verified empirically, not assumed. Keeping
them would have required an adapter importing both, which is the coupling
this release exists to remove.

The cost is two log lines losing a tag: the rollback error no longer carries
WithTrace(ctx), and the read/write executors no longer tag themselves with
mode. Callers wanting the latter can pass an already-tagged logger; the
former genuinely needs the per-call context and would need an adapter
package to restore.

v0.12.0

What's Changed

• BREAKING: streams: make streams context-aware by @sonirico in #2

Full Changelog: v0.11.4...v0.12.0

v0.11.4

Full Changelog: v0.11.3...v0.11.4

v0.11.3

Full Changelog: v0.11.2...v0.11.3

v0.11.2

Full Changelog: v0.11.1...v0.11.2

v0.11.1

Full Changelog: v0.11.0...v0.11.1

v0.11.0

Full Changelog: v0.10.0...v0.11.0

v0.10.0

Full Changelog: v0.9.0...v0.10.0

v0.9.0

What's Changed

• feat(slices): uniq family fns by @sonirico in #1

New Contributors

• @sonirico made their first contribution in #1

Full Changelog: v0.8.2...v0.9.0