fix(db): drop the retired @hasna/cloud adapter, restore bun:sqlite

[andrei-hasna]

Why

@hasna/cloud is retired. Its npm deprecation notice reads:

@hasna/cloud is retired and no longer supported by Hasna. The source repo has been deleted. Do not add new dependencies on it; services now own their storage (local SQLite / self-hosted API).

PR #16 (1c37dae, merged 2026-07-30T09:05Z) added "@hasna/cloud": "^0.1.41" plus imports across 11 src/db/* files — six days after that deprecation. Its GitHub repo is confirmed gone (246 repos enumerated across hasna + hasnaxyz with a repo-scoped token, zero matches; 8 exact-name probes 404), so a defect in that adapter can never be fixed upstream.

This repo had a working, self-owned local SQLite store before #16. The replacement home named by the deprecation notice is therefore the one it already had — so this restores it rather than inventing a third path.

What

A clean inverse of 1c37dae, which was the tip of main and a purely mechanical adapter swap:

before #16	#16	this PR
Database from bun:sqlite	SqliteAdapter from @hasna/cloud	Database
type Database	type DbAdapter	type Database
db.run(sql, [a, b])	db.run(sql, a, b)	db.run(sql, [a, b])
db.query(sql).get()/.all()	db.get(sql)/.all(sql)	db.query(sql).get()/.all()
db.transaction(fn)(rows)	db.transaction(fn)	db.transaction(fn)(rows)

No behaviour beyond the adapter swap is changed.

Regression test

Adds src/db/storage-ownership.test.ts. It was written to fail while the dependency is present and pass once it is gone — not merely to show the code still works.

Every assertion carries a positive control asserting the same reader/scanner finds something genuinely present, so a broken reader returning {} or an empty file list cannot make it pass vacuously. Each assertion was then verified to die under targeted mutation:

mutation	result
re-add the dep to package.json only	assertion 1 fails, other 2 pass
re-add the import to one source file only	assertion 2 fails and names that file, other 2 pass
wrap the handle in a pass-through adapter	assertion 3 fails, other 2 pass
restored tree	3 pass, 0 fail, 6 expect() calls

The mutation-3 wrapper passed exec/run/query straight through and still failed — a working wrapper is still a wrapper.

The guard needs no node_modules, so it runs anywhere.

Verification

Worktree 3281f770-decloud, 177 packages installed, @hasna/cloud absent from node_modules:

• tsc --noEmit — rc=0
• bun test — 1055 pass, 1 fail, 3969 expect() calls, 1056 tests across 18 files
• /usr/bin/grep -rn "@hasna/cloud" src package.json — no match, with a positive control confirming the same command finds @hasna/events
• staged secrets scan — clean, with a positive control confirming the pattern fires on a synthetic credential

The single failure is pre-existing and unrelated. hooks/codewith-native-common.test.ts — a destructive-shell-guard test — overshoots its 5000ms budget. It references no db code (grep for src/db/getDb/createTestDb returns nothing) and is not in this diff. Running that same file against unmodified origin/main gives an identical 123 pass / 1 fail, so this PR introduces no regression.

Not touched, deliberately

The wider @hasna/cloud removal programme — 16 published @hasna packages whose latest still declares it, 15 as production dependencies — is a separate republish step and is out of scope here. This PR lands the source change only and publishes nothing.

Refs: todos 3281f770, knowledge k_ms8ngox0_eb1tet

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[andrei-hasna]

[REVIEW] MERGE — #21 @ 3ae872d — lens: revert-completeness+storage-correctness, reviewer Cato (1 of 1)

Independent adversarial review, no author context inherited. All figures measured first-hand in a dedicated worktree at the PR head, against a like-for-like origin/main baseline installed from its own manifest. Canonical checkout open-hooks used throughout — the bare-name repos repo hooks lookup resolves to the stale _factory_src/hooks mirror and was discarded.

1. Clean inverse — yes, provably. origin/main is the adding commit 1c37dae, which permits a decisive test: git diff 1c37dae^ 3ae872d returns one file, src/db/storage-ownership.test.ts, +116 and nothing else. The tree at the PR head is therefore byte-identical to the pre-dependency tree apart from the new guard. No drive-by change is possible. Arithmetic is consistent: the original was +120/−116, its inverse +116/−120, plus the 116-line guard = the stated +232/−120 across 13 files.

2. Dependency actually gone. At the PR head the only surviving @hasna/cloud strings are inside the guard that asserts its absence (1 file). Positive control @hasna/events returns 3 files, so the scan can find what is there; the same scan on origin/main returns 12 files including package.json:53. Absence is controlled in both directions.

3. Mutation evidence reproduced. Two of the three killed independently, each killing exactly one assertion:

• dep re-added to package.json → 2 pass / 1 fail, only "package.json declares no @hasna/cloud"
• createTestDb returning a delegating pass-through adapter → 2 pass / 1 fail, only "the database handle is a bun:sqlite Database"

The second is the one worth having: it fires on a wrapper whose import string never appears, so the behavioural assertion genuinely backstops the textual scan rather than restating it. Both mutations reverted; worktree ends clean.

4. Controls are load-bearing. I verified the equivalent guard's positive control on the sibling repo by emptying the manifest reader's input: the test fails on the control assertion rather than passing vacuously. The guard cannot pass against a reader that sees nothing.

5. Gates, measured here. tsc --noEmit rc=0, 0 errors — and a deliberately injected type error produced exactly 1, proving tsc actually checks this tree. Secrets scan over the diff: 0 matches, with a synthetic credential confirming the pattern fires and a harmless line confirming it does not over-fire. Author/committer Andrei Hasna <andrei@hasna.com> from /home/hasna/.gitconfig with no repo-local override; exactly one Agent: Vespasian; zero Co-Authored-By — note the reverted 1c37dae carried Co-authored-by: hasna-drain, correctly not propagated. Local ref == origin head.

6. Suite. 1055 pass / 1 fail / 18 files, 1056 tests. The single failure is hooks/codewith-native-common.test.ts › a command too large to tokenize…, which contains no db reference and is not in the diff. Pre-existence confirmed by running that file on unmodified origin/main: 123 pass / 1 fail, same test name. Read it correctly — bun reports "this test timed out after 5000ms", so 5000 ms is the budget and the true runtime is unbounded; the 5893 ms figure is wall time on a kill, not a near-miss.

Non-blocking observations (neither affects this revert):

• bun.lock is gitignored in this repo (.gitignore:9), so there is no reproducible install and the guard correctly carries no lockfile assertion — unlike the sibling repo, which tracks its lock and does. The asymmetry is justified, not an oversight. A gitignored lockfile in a published package is worth its own task.
• The canonical checkout's node_modules and local lock date from Jul 20, ten days before this dependency merged on Jul 30, and contained only @hasna/events — so @hasna/cloud was never resolved there. It is installable (the baseline install pulled it, 179 vs 177 packages); the checkout was simply stale. So the accurate statement is "nobody ever ran this suite with the dependency installed", not "the dependency is unresolvable".
• src/lib/installer.test.ts reads the real ~/.claude/settings.json, and the run reports "killed 1 dangling process". Pre-existing test-isolation leak.
• I measured 3970 expect() calls against the 3969 stated; a one-call drift, immaterial, likely a conditional assertion.

Verdict MERGE. The revert is exactly the dependency removal and nothing else, the restored storage path is bun:sqlite verified behaviourally, and the guard is mutation-proven and non-vacuous. Republishing the sixteen packages that still declare the dependency is a separate programme and explicitly out of scope here.