fix: correct SQLite file URIs on Windows

[0xMrMasry]

What changed

On native Windows, a fresh codeaf launch disables memory, conversation search, and automations with SQL logic error: out of memory (1). The database openers pass net/url.URL a Windows drive path, producing file://C:%5C.... SQLite rejects that URI before opening a database. With the same driver and machine, :memory:, a plain filename, and file:///C:/... all open successfully.

Add a shared file-URI constructor that uses forward slashes and a rooted URL path, then use it in the graph, automation, collection, and read-only plan-preview database openers. POSIX absolute paths keep their existing URI shape, and query parameters remain with each caller. The regression test exercises persistence at the requested filename with spaces, #, &, and %, connection pragmas, and read-only access.

Related to #1082. This addresses database opening; the separate Windows resume/presence-locking problems remain outside this change.

How it was checked

On Windows amd64 with Go 1.26.5 and the repository's modernc.org/sqlite v1.29.1:

• Before the fix, go test ./internal/automation -run '^TestStoreRefusesAForeignDatabase$' -count=1 fails with the reported SQLite error.
• After the fix, the following eight focused tests pass across all five packages:

go test ./internal/sqliteuri ./internal/store ./internal/automation ./internal/workspace ./internal/plandb -run '^(TestFileOpensTheNamedDatabase|TestOpenMakesItsDirectoryOnAFirstRun|TestOpenIsHappyWhenTheDirectoryIsAlreadyThere|TestStoreRefusesAForeignDatabase|TestCreateWorksOutTheFirstWake|TestTakeCatchesUpOnceAndMarksItLate|TestNestedCollectionsAllowSharedParentsButRejectCycles|TestRootPreviewReadsBesideAWriterAndNeverCreatesAStore)$' -count=1

• git diff --check passes.
• The full automation suite on Windows also encounters existing presence-lock rename and POSIX-path fixture failures; this PR does not claim full Windows support.
• go vet ./internal/sqliteuri ./internal/store ./internal/automation ./internal/workspace ./internal/plandb passes.
• make changelog-check passes, including the entry for fix: correct SQLite file URIs on Windows #1868.
• make pr-ready passes go build ./... and then fails at go vet ./... because unmodified tests use Unix-only syscall.Setpgid, syscall.Kill, and syscall.Flock, and other unmodified test helpers/packages are excluded on Windows. The full local gate therefore does not pass.
• Upstream GitHub Actions requires maintainer approval for the fork PR before CI can run.

Checklist

• Change entry added for fix: correct SQLite file URIs on Windows #1868 and validated, as prescribed in docs/rules/changelog.md.
• No slash command, key, tool, default, limit, or refusal was added or changed, so no manual update is needed.
• No new entry in .github/known-red.txt.
• Only the paths belonging to this fix were staged.

[santoshkumarradha]

Hey, thanks for the PR. Could you sign the CLA when you get a chance so we can review it?

[0xMrMasry]

Hey, thanks for the PR. Could you sign the CLA when you get a chance so we can review it?

Happy to sign! Could you share the CLA signing link? The license/cla check currently shows "Expected" without a link.

[0xMrMasry]

Thanks! I've now signed the CLA as 0xMrMasry, and CLA Assistant showed the confirmation: "Thank you 0xMrMasry for signing the CLA!"

The PR still appears to be waiting for license/cla, with no status reported yet. Could you recheck the CLA integration, or let me know if there's another step I need to complete?

@santoshkumarradha

[AbirAbbas]

Verified, including natively on Windows: cross-compiled test binaries run on Windows pass at this head (drive, forward-slash, UNC and special-character paths, pragmas kept, read-only doesn't create a missing DB), and the same automation test built from dev fails with the original SQL logic error: out of memory (1). On Linux the URIs are byte-identical to what dev produced for spaces, #, &, %, ?, unicode and backslashes, so POSIX users see no change. I added those cases as tests and merged current dev. A few other openers still build URIs by hand; that's pre-existing and filed separately. Thanks!