fix(cli): honour deletion of extension-managed objects in declarative sync

[avallete]

Removing a pg_cron job or pgmq queue declaration from a manifest-less declarative tree was refused as a legacy pg-delta export, and the refusal then blocked every later sync of that tree, unrelated schema work included. The engine already plans the cron.unschedule / pgmq.drop_queue calls; the refusal was CLI-side classification.

The legacy-export gate now only fires when the plan drops an extension. Extension-managed object removals (cron jobs, pgmq queues) no longer count toward it, so deleting or renaming one on a converged tree generates the unschedule / drop_queue statements like any other change. The gate still enumerates those objects as evidence when their owning extension itself is being dropped, which is the legacy-export case it was built for.

No new flags, prompts, or warning output. The change is confined to legacyClassifyDeclarativeCompatibilityGap in declarative.flow.ts, plus an integration scenario and an e2e scenario that adds a job and a queue to the tree, then renames the job and drops the queue.

Resolves CLI-2282.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Wh8sCVwt29sYWSYdcDUbWN

[github-actions]

Supabase CLI preview

npx --yes https://pkg.pr.new/supabase/cli/supabase@cc83f6009d2386484590c4acfa0cffea00f362a3

Preview package for commit cc83f60.

[github-actions]

🤖 AI Review

Both independent reviews were available. The overlapping data-loss findings are confirmed as one issue: managed-object removals now bypass the legacy gate but are not added to destructive-change warnings. Four additional Claude findings are confirmed, covering missing advertised affordances, two test-quality gaps, and prose wrapping.

Findings

Severity	Location	Category	Sources	Claim
🟠 MAJOR	apps/cli/src/legacy/commands/db/schema/declarative/declarative.flow.ts:95	data-loss	claude+codex	Extension-managed object removals whose owning extension remains declared—including removals encountered after repairing an unrelated implicit extension—proceed without the destructive-change warning promised by the PR, allowing --yes to apply cron.unschedule or pgmq.drop_queue silently.
🟡 MINOR	apps/cli/src/legacy/commands/db/schema/declarative/sync/sync.e2e.test.ts:200	test-quality	claude	The new e2e helper reads only the lexicographically last migration, making the assertions brittle when pg-delta splits a valid plan across multiple migration files.
🟡 MINOR	apps/cli/src/legacy/commands/db/schema/declarative/sync/sync.handler.ts:553	documentation	claude	The PR description advertises an --allow-removals flag and a staged-export Continue with removals choice, but neither exists in the checked-out implementation.
🟡 MINOR	apps/cli/src/legacy/commands/db/shared/legacy-pgdelta-next-adapter.layer.ts:154	test-coverage	claude	The dropped-extension evidence path is not covered with a real pg-delta plan, leaving unverified whether managed-object intents remain available after the summarizer discards child removal facts.
⚪ NIT	apps/cli/src/legacy/commands/db/schema/declarative/sync/SIDE_EFFECTS.md:97	style	claude	The edited paragraph leaves line 97 at 85 columns, disrupting its approximately 80-column prose wrapping.

Findings outside the diff

• 🟡 MINOR apps/cli/src/legacy/commands/db/schema/declarative/sync/sync.handler.ts:553 — The PR description advertises an --allow-removals flag and a staged-export Continue with removals choice, but neither exists in the checked-out implementation.
• 🟡 MINOR apps/cli/src/legacy/commands/db/shared/legacy-pgdelta-next-adapter.layer.ts:154 — The dropped-extension evidence path is not covered with a real pg-delta plan, leaving unverified whether managed-object intents remain available after the summarizer discards child removal facts.

Stats

Claude findings: 5 · Codex findings: 1 · Confirmed: 5 · Refuted: 0 · Uncertain: 0

---

Models: claude-opus-5 + gpt-5.6-sol · Trigger: auto · Workflow run

This review runs once per PR. A maintainer can request another with a /ai-review comment.

[avallete]

On the two review findings without an inline thread:

• Description advertises --allow-removals / "Continue with removals": the description was rewritten after that review ran; neither exists on this branch and the description no longer mentions them.
• Dropped-extension evidence not covered with a real pg-delta plan: the concern is that legacySummarizePgDeltaNextRemovals skips child facts, so intents could vanish when their extension is dropped. In pg-delta's pg_cron and pgmq policies the extensionIntent facts are constructed as root facts with no parent, so the summarizer keeps them. That path is also unchanged by this PR, which only narrows what trips the gate.

---

Generated by Claude Code