Documentation, one new validation rule, and test coverage on top of wsman's
GitHub PR #6. No change to what _defensive itself does.

conn.go: the comment above the db_config calls had been generalized to
"connection-level sqlite3_db_config options are applied before
applyQueryParams because the SQLite contract requires
sqlite3_db_config(SQLITE_DBCONFIG_DQS_*) to be set before any statement is
prepared". That contract is DQS-specific and does not extend to DEFENSIVE,
which may be toggled at any point in a connection's life; defensive goes
first because the PRAGMAs that follow must run under the restriction, which
is a choice, not an API requirement. Restore the DQS comment verbatim on the
DQS call and give _defensive its own rationale.

sqlite.go: reject _defensive=1 together with _journal_mode=OFF (or
_journal=OFF). SQLite turns PRAGMA journal_mode=OFF into a no-op that still
reports success under defensive mode, so the combination previously opened a
connection in which neither parameter had been honoured and nothing was
reported. The check sits in the validation phase from v1.55.0, before any
statement executes, so a rejected DSN still cannot leave the database
half-configured. _pragma stays the documented exception: _pragma
=journal_mode(OFF) alongside _defensive=1 runs and is silently ignored by
SQLite. Only DSNs using _defensive can be affected and that parameter is new,
so no DSN that opened before changes behavior.

driver.go: expand the _defensive documentation to the depth of its _dqs and
_error_rc neighbours -- what the mode observably does (writable_schema=ON,
journal_mode=OFF and schema_version=N become silent no-ops; shadow-table and
sqlite_dbpage writes error; reads, ordinary virtual table use and VACUUM are
unaffected), the new _journal_mode conflict, and the two limits the name
invites callers to overlook: it is a hardening measure rather than a sandbox
for hostile files (this build has neither SQLITE_TRUSTED_SCHEMA=0 nor
SQLITE_DQS=0 and there is no authorizer), and it is a property of the
connection, not of the database file.

defensive_test.go: rename TestUnmodifiedBehaviorIsNotAnADEDefensiveControl,
whose "ADE" is an acronym from the contributor's downstream and means nothing
here, to TestDefensiveAbsentOrFalseIsBaseline, and TestUnprotectedPoolIsA
NegativeControl to TestDefensiveOffPoolIsANegativeControl so every test in
the file groups under the feature. Add TestDefensiveRejectsJournalModeOff for
the new rule, TestDefensiveDSNForms for the file: URI shape -- which keeps its
query string all the way into sqlite3_open_v2, so SQLite parses _defensive too
and must ignore it -- plus :memory: and shared-cache memory URIs,
TestDefensiveLeavesOrdinaryUseIntact for the other half of the contract
(fts5 create/insert/MATCH, foreign keys, VACUUM, integrity_check and every
other DSN parameter that could collide, with direct fts5 shadow-table and
sqlite_dbpage writes refused), and TestDefensiveIsPerConnection to pin the
documented scope. Also adopt the file's `package sqlite // import ...` form.

CHANGELOG.md: document both the parameter and the new conflict rule.

TestConcurrentGoroutines re-invokes itself under -race and treats anything
the recursive run prints other than two known "cannot run here" messages as a
failure. With CGO_ENABLED=0 the go tool refuses with "-race requires cgo;
enable cgo by setting CGO_ENABLED=1", which matched neither, so the whole
suite failed on an environment where the check simply cannot run -- and a
CGo-free driver is a natural thing to build and test with cgo disabled.
Accept that message alongside the existing two and skip, as the test already
does for a toolchain without race support and for an unsupported VMA range.
Nothing changes when cgo is available: the recursive -race run still executes
and still has to pass.

Found while reviewing GitHub PR #6, whose author hit it in a CGO_ENABLED=0
lane; unrelated to that change.

Deln0r has had Maintainer rights on the GitLab project since 2026-06-09.
The file has named only me since it was added in February and predates
that grant.

Split out of !135, where the hunk was proposed alongside a code change.

vec/ has carried the transpiled sqlite-vec sources since v1.47.0, but the
module shipped only its own BSD-3-Clause LICENSE and the public-domain SQLite
notice. sqlite-vec is Copyright (c) 2024 Alex Garcia and dual-licensed
Apache-2.0 OR MIT; modernc.org/libsqlite_vec's generator elects MIT, whose
terms require the copyright and permission notice to accompany substantial
portions of the software. 2.8 MB of transpiled vec/ is a substantial portion.
Attribution was never absent -- vec's package documentation names the
extension, pins v0.1.9 and links upstream -- but the license text was.

LICENSE-SQLITE_VEC: the notice, byte-identical to LICENSE-MIT in the upstream
v0.1.9 archive and to the file libsqlite_vec extracts it into.

vendor_libs/main.go: the omission was mechanical -- the tool copied the
per-target transpiles and nothing else, so a plain cp of the notice would have
survived only until the next `make vendor`. Copy it alongside the sources it
belongs to, and treat a missing source as fatal: shipping the code without the
notice is worse than not vendoring at all.

SQLITE-LICENSE -> LICENSE-SQLITE, contents unchanged. This matches the new file
beside it and the LICENSE-<upstream> convention the rest of the modernc.org
repositories follow, but it is not only cosmetic. `go mod vendor` picks the
metadata files it copies into a downstream vendor/ tree by matching each name
against a fixed prefix list (cmd/go/internal/modcmd/vendor.go, metaPrefixes)
that includes LICENSE, so a name merely ending in LICENSE was never propagated.
Both notices now reach vendored builds, which is where the MIT terms on vec/
keep applying. Direct links to the old path will break.

vec/patches.go: a License section on the package documentation, so an importer
of vec sees on pkg.go.dev that this package is under a different license from
the rest of the module.

Found by an SBOM audit of the published module.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The vendoring reads one full per-target file at a time from ../libsqlite3 and
../libsqlite_vec, which assumes those checkouts ship expanded. They do today,
but either may adopt the deduplicated layout that modernc.org/wa2c already ships
(modernc.org/builder's NW autogen; see NW_GENERALIZATION_HANDOFF.md there) — and
then those per-target files hold only each target's residue, so vendoring them
would silently drop most of the package.

Detect it instead of remembering it later. srcDir applies undup's own rule
(base.go or base_g_*.go present means folded) and:

  - returns an expanded checkout as is: no copy, no subprocess, exactly the path
    this tool has always taken;
  - copies a deduplicated one to a temp directory and expands it THERE, never in
    place — expanding the sibling would leave that checkout dirty and tempt a
    "restore" that discards whatever else is uncommitted in it. go.mod and go.sum
    travel with the copy because undup resolves each unaliased import's package
    name by running "go list" inside the directory it expands.

The undup pin is passed in from the Makefile, so the repo keeps one version of
record wherever undup is invoked, and expansion that leaves shared files behind
is a hard error naming the pin — that is what a version too old for the layout
looks like.

Also adds -libsqlite3 / -libsqlite_vec flags so this can be exercised against
scratch copies without touching the real checkouts.

Verified both ways: output byte-identical to the previous binary on today's
expanded checkouts (38/38 files); and from deduplicated copies of both siblings,
a complete vendoring whose lib/ and vec/ build for linux/{amd64,s390x},
darwin/arm64, windows/{amd64,386}, freebsd/386 and netbsd/amd64.

Expect one-time textual churn whenever a sibling does flip: undup reconstructs
declarations in its own order and recomputes imports, dropping ccgo's
`var _ = math.Pi` and friends along with the imports they exist to keep. Same
package, different bytes.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

TestDBPageVtab's comment about -DSQLITE_ENABLE_DBPAGE_VTAB carried a trailing
space, which made all_test.go the one hand-written file in the tree that
gofmt -l reports. Comment text only; nothing else changes.

Noticed by Ian Chechin while preparing GitLab merge request #135, which does
not touch this file.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The test registered its driver under the fixed name "sqlite_conn_hook_test".
sql.Register panics on a name it has already seen and offers no way to undo a
registration, so the second iteration of go test -count=2 took the whole test
binary down with "sql: Register called twice for driver
sqlite_conn_hook_test" -- not a failure of the code under test, and it hid
whatever the remaining iterations would have found.

Derive the name from an atomic counter instead, so each invocation registers
its own driver, and close the sql.DB the test opens while here: it was leaked
once per iteration.

Nothing changes for a single run. Ian Chechin's driver_register_test.go in
GitLab merge request #135 solves the same problem with a uniqueDriverName
helper; the two can be folded together once that lands.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Driver holds four categories of registration state but only connection
hooks could be put on a constructed one. Functions and collations were
reachable through the package-level API alone, and modules through the
package-level driver only, so a constructed Driver was half-built: its
modules field was written and read through the package-level instance,
making it process-global state wearing a per-instance field.

This is the additive half of #254, with one narrow, loud exception
spelled out below:

  - registerFunction and registerCollation become methods on *Driver.
    The package-level RegisterFunction, RegisterScalarFunction,
    RegisterDeterministicScalarFunction and RegisterCollationUtf8 keep
    targeting the package-level driver, spelled explicitly now rather
    than by relying on the receiver named d shadowing the package
    variable of the same name. newDriver is renamed defaultDriver,
    since it read like a constructor but returns the singleton.

  - Driver gains RegisterFunction, RegisterScalarFunction,
    RegisterDeterministicScalarFunction, RegisterCollationUtf8 and
    RegisterModule, plus Must* variants of the first four, each
    registering on that Driver alone. The zero Driver stays usable: the
    maps are created on demand, since a constructed &Driver{} has nil
    maps where the package-level instance is built with make. A mutex
    on Driver makes concurrent registrations safe, so a *Driver can be
    handed out for several packages to fill from their init functions.

  - A connection now receives the union of the modules registered on the
    package-level driver and those registered on the Driver that opened
    it. The package-level set is applied unconditionally, which is what
    every connection has received since module support was added, so the
    isolating change #254 warns about is not made here. A name present
    in both resolves to the package-level implementation.

  - The ID handed to sqlite3_create_module_v2 as pAux is allocated per
    (Driver, name) pair, not per name. It is what every trampoline
    dispatches on, and a per-name ID would make two Drivers registering
    the same module name share one entry, with the last registration
    winning process-wide and retroactively, on connections already open.

  - vtab.RegisterModule honours its db argument: a non-nil db registers
    on the driver backing it when that driver implements the new
    vtab.ModuleRegisterer, and a nil db keeps targeting the driver this
    package registers as "sqlite". For a db opened on "sqlite" the two
    are the same driver, so the outcome is unchanged.

The exception: vtab.RegisterModule(db, ...) where db was opened on a
caller-constructed Driver used to discard the db argument and land on
the "sqlite" driver, reaching every connection in the process. It now
lands on the constructed driver alone, so a sql.Open("sqlite")
connection that used to resolve such a module gets "no such module"
instead.

That same call is also the one route by which existing code can hold a
module name on both a constructed Driver and the package-level one, so
the collision rule above is a second, narrower part of the exception
rather than a continuation. Before, the first of the two registrations
won and the second was refused as already registered; now the
package-level implementation wins on the constructed Driver's
connections whichever order they ran in. No order-independent rule
reproduces the old behavior, and reaching the case at all means the
program ignored the error the older version returned.

Two smaller deviations round it out: Driver.RegisterModule reports no
error for such a collision, and vtab.RegisterModule validates its name
and module arguments before the not-implemented check, so a call with an
empty name that returned ErrNotImplemented outside this driver now
returns "vtab: module name must be non-empty".

Tests cover the new registrations, that they reject duplicates, that two
constructed Drivers stay isolated from each other and from the
package-level driver -- including when both register the same module
name, exercised on a connection pinned before the second driver opens
one -- that concurrent registrations survive the race detector, that
globally registered modules still reach a constructed Driver, and that
the db argument selects the driver. The behavior-preserving properties,
the same-name case and the concurrency claim were each verified by
breaking them deliberately and confirming the test fails.

Updates #254

sqlite: let a caller-constructed Driver register its own functions, collations and modules

See merge request !135

All three shipped as experimental in v1.53.0 and were deliberately left out
of the supported platforms table until they had accumulated some real-world
exposure. That period has elapsed: they have been in the builder matrix and
in build_all_targets since, they pass on this release's commit alongside the
seventeen platforms already listed, and no open issue reports a defect in any
of them. The table therefore listed seventeen entries while the module
shipped, cross-built and tested twenty; it now lists all twenty.

Also set the v1.57.0 CHANGELOG date to the release date, and drop a stale
claim in lib/hooks_linux_arm64.go that this module is "stuck on libc@v1.55.3"
-- go.mod has pinned v1.74.4 since v1.56.0. The comment now says what the
run-time patch is actually for.