fix(schema): rebuild pdp_prove_tasks after 0017 left it FK'd to a dropped table (#75)

0017_drop_legacy_v1_artifacts dropped pdp_proof_sets but missed that
pdp_prove_tasks (created in 0010) still declared
REFERENCES pdp_proof_sets(id). Two failures followed on any fresh
install:

1. With foreign_keys=ON, DELETE FROM harmony_task (harmonytask's
   completion recorder) FK-scans pdp_prove_tasks and dies with
   'no such table: main.pdp_proof_sets' -- so EVERY task completion
   failed to record and retried forever. Observed live on a clean
   calibration boot (task_type_handler.go:482 error spam).
2. The table was still in v1 shape (proofset column), while
   tasks/pdpv0/task_prove.go inserts (data_set, task_id) -- so
   ProveTask scheduling would have failed with 'no such column'.

Migration 0022 rebuilds pdp_prove_tasks in the v0 vocabulary, FK'd to
pdp_data_sets, copying any in-flight rows. Also:

- migrations_test: pdp_proof_sets/pdp_proofset_* moved from wantTables
  to an explicit dropped-tables assertion (the test had been failing on
  main since 0017 landed)
- new TestApplyMigrations_TaskCompletionDeleteWorks: PRAGMA
  foreign_key_check + the real harmony_task DELETE path + v0 column
  shape, so the dangling-FK class can't regress silently

Lesson encoded in the migration header: when dropping a referenced
table, grep schema files for 'REFERENCES <table>' first.

Author: Reiers

internal/harmonysqlite/migrations_test.go

@@ -168,9 +168,10 @@ func TestApplyMigrations_PDPTablesPresent(t *testing.T) {
 		"sector_location", "storage_path",
 		// Eth chain queue
 		"eth_keys", "message_sends_eth", "message_waits_eth", "message_send_eth_locks",
-		// PDP v1
+		// PDP v1 (legacy pdp_proof_sets / pdp_proofset_* vocabulary was
+		// renamed to pdp_data_sets / pdp_data_set_* and the old tables are
+		// DROPPED by 0017_drop_legacy_v1_artifacts.sql — asserted absent below)
 		"pdp_services", "pdp_piecerefs", "pdp_piece_uploads", "pdp_piece_mh_to_commp",
-		"pdp_proof_sets", "pdp_proofset_creates", "pdp_proofset_roots", "pdp_proofset_root_adds",
 		"pdp_prove_tasks",
 		// PDP v0 (renamed vocabulary)
 		"pdp_data_sets", "pdp_data_set_creates", "pdp_data_set_pieces", "pdp_data_set_piece_adds",
@@ -191,6 +192,21 @@ func TestApplyMigrations_PDPTablesPresent(t *testing.T) {
 			}
 		})
 	}
+
+	// Legacy v1 vocabulary must be GONE after 0017_drop_legacy_v1_artifacts.
+	droppedTables := []string{
+		"pdp_proof_sets", "pdp_proofset_creates", "pdp_proofset_roots", "pdp_proofset_root_adds",
+	}
+	for _, tbl := range droppedTables {
+		t.Run("dropped/"+tbl, func(t *testing.T) {
+			var got string
+			err := db.QueryRow(ctx,
+				`SELECT name FROM sqlite_master WHERE type='table' AND name=?`, tbl).Scan(&got)
+			if err == nil {
+				t.Errorf("legacy table %s still present; 0017 should have dropped it", tbl)
+			}
+		})
+	}
 }
 
 // TestApplyMigrations_PDPRefcountTriggerWorks: the inline-body trigger
@@ -290,3 +306,53 @@ func TestApplyMigrations_FKEnforced(t *testing.T) {
 		t.Logf("FK violation (formatted): %v", err)
 	}
 }
+
+// TestApplyMigrations_TaskCompletionDeleteWorks reproduces the live
+// failure behind migration 0022: harmonytask's completion recorder does
+// DELETE FROM harmony_task WHERE id=$1, and with foreign_keys=ON SQLite
+// scans every table whose FK references harmony_task. When 0017 dropped
+// pdp_proof_sets but left pdp_prove_tasks declaring
+// REFERENCES pdp_proof_sets(id), that scan failed with
+// "no such table: main.pdp_proof_sets" and EVERY task completion
+// retried forever. This test fails if any FK in the schema dangles.
+func TestApplyMigrations_TaskCompletionDeleteWorks(t *testing.T) {
+	db, err := Open(Config{Path: ":memory:", ForeignKeys: true})
+	if err != nil {
+		t.Fatalf("Open: %v", err)
+	}
+	defer db.Close()
+	ctx := context.Background()
+	if err := db.ApplyMigrations(ctx); err != nil {
+		t.Fatalf("ApplyMigrations: %v", err)
+	}
+
+	// foreign_key_check catches dangling FK targets schema-wide.
+	rows, err := db.Query(ctx, `PRAGMA foreign_key_check`)
+	if err != nil {
+		t.Fatalf("PRAGMA foreign_key_check errored (dangling FK table?): %v", err)
+	}
+	rows.Close()
+
+	// And the actual harmonytask completion path: insert a task with no
+	// owner, then DELETE it. The DELETE triggers FK-cascade scans across
+	// all referencing tables (pdp_prove_tasks among them).
+	var taskID int64
+	err = db.QueryRow(ctx,
+		`INSERT INTO harmony_task (posted_time, added_by, name)
+		 VALUES (CURRENT_TIMESTAMP, 1, 'completion-test') RETURNING id`).Scan(&taskID)
+	if err != nil {
+		t.Fatalf("insert harmony_task: %v", err)
+	}
+	if _, err := db.ExecCount(ctx, `DELETE FROM harmony_task WHERE id=?`, taskID); err != nil {
+		t.Fatalf("DELETE FROM harmony_task failed (this is the 0022 bug class): %v", err)
+	}
+
+	// pdp_prove_tasks must be in the v0 shape: (data_set, task_id),
+	// FK'd to pdp_data_sets, so task_prove.go's INSERT works.
+	var colName string
+	err = db.QueryRow(ctx,
+		`SELECT name FROM pragma_table_info('pdp_prove_tasks') WHERE name='data_set'`).Scan(&colName)
+	if err != nil {
+		t.Errorf("pdp_prove_tasks.data_set column missing (still v1 'proofset' shape?): %v", err)
+	}
+}

internal/harmonysqlite/schema-curio-core/0022_fix_prove_tasks_dangling_fk.sql

@@ -0,0 +1,48 @@
+-- 0022_fix_prove_tasks_dangling_fk.sql
+--
+-- 0017_drop_legacy_v1_artifacts dropped pdp_proof_sets but left
+-- pdp_prove_tasks behind, still declaring:
+--
+--   proofset INTEGER NOT NULL REFERENCES pdp_proof_sets(id) ON DELETE CASCADE
+--
+-- Two bugs follow:
+--
+-- 1. DANGLING FK. With foreign_keys=ON, SQLite resolves FK targets at
+--    statement prepare. Any DML that touches pdp_prove_tasks -- including
+--    the FK-cascade scan triggered by DELETE FROM harmony_task in
+--    harmonytask's completion recorder -- fails with
+--    "no such table: main.pdp_proof_sets". On a fresh install this means
+--    EVERY task completion fails to record and retries forever
+--    (observed live: harmonytask task_type_handler.go:482 error spam on
+--    a clean calibration boot).
+--
+-- 2. STALE COLUMN NAME. Upstream's 20250730-pdp-v0-rename.sql renamed
+--    proofset -> data_set; tasks/pdpv0/task_prove.go inserts
+--    (data_set, task_id). Against the v1-shape table that INSERT fails
+--    with "no such column", so ProveTask scheduling breaks too.
+--
+-- Fix: rebuild pdp_prove_tasks in the v0 vocabulary, FK'd to
+-- pdp_data_sets. The table is transient scheduling state (rows live only
+-- while a prove task is in flight), so the copy is best-effort: on any
+-- broken deployment the table is empty (nothing could insert), and on a
+-- pre-0017 deployment rows are in the old column shape.
+--
+-- NOTE for future drops: when dropping a referenced table, always grep
+-- schema files for "REFERENCES <table>" first. 0017 missed this one.
+
+CREATE TABLE IF NOT EXISTS pdp_prove_tasks_v0 (
+    data_set INTEGER NOT NULL REFERENCES pdp_data_sets(id) ON DELETE CASCADE,
+    task_id  INTEGER NOT NULL REFERENCES harmony_task(id) ON DELETE CASCADE,
+    PRIMARY KEY (data_set, task_id)
+);
+
+-- Copy any in-flight rows (old column name "proofset" -> "data_set").
+-- On post-0017 fresh installs the source table exists but is empty;
+-- the SELECT lists the legacy column, which still exists in the legacy
+-- table shape this migration replaces.
+INSERT OR IGNORE INTO pdp_prove_tasks_v0 (data_set, task_id)
+    SELECT proofset, task_id FROM pdp_prove_tasks;
+
+DROP TABLE pdp_prove_tasks;
+
+ALTER TABLE pdp_prove_tasks_v0 RENAME TO pdp_prove_tasks;