fix(retention): dedupe contributor-decision-pack to latest-only, not a preserved series

contributor-decision-pack was excluded from LATEST_ONLY_SIGNAL_SNAPSHOT_TYPES on the
theory that it's a bounded trend/change series read by history, with volume "a fraction
of" its three contributor-* neighbors that already dedupe. Both claims were wrong:

- src/services/decision-pack.ts's only reader calls listSignalSnapshots(...)[0] -- the
  exact same latest-only contract as contributor-evidence-graph/outcome-history/strategy,
  which already dedupe correctly.
- Measured against a live export of the hosted D1: 18,549 rows, 6.3 GB, ~350 KB/row --
  71% of the database's entire 8.96 GB file size, and roughly 750x the combined size of
  the three types it was assumed to dwarf.

Every row is still inside the 90-day RETENTION_POLICY age window (the database itself is
only ~65 days old), so age-based pruning had not removed a single one, and the type has
accumulated ~700-1,300 rows/day since 2026-07-06 -- large enough to refill the 10 GB cap
from empty in 3-4 weeks on its own, even with #9415's five-table fix applied (that PR
addressed other tables totaling well under 1 GB combined).

queue-health is the one signal type genuinely read as a historical series
(maintainer-slop-duplicate-trend.ts shapes multiple weeks into a trend card) and stays
excluded, unchanged.

Refs #9435

Author: JSONbored

src/db/retention.ts

@@ -192,9 +192,9 @@ export type SignalSnapshotDedupeResult = { signalType: string; deleted: number }
 
 /** Exported so the D1 size/row-count observability probe (#3810, src/selfhost/d1-size-probe.ts) can scope its
  *  signal_snapshots "rows per dedup key" ratio to exactly the population this dedup job converges to ~1 row
- *  per key -- NOT the whole table, which intentionally keeps bounded multi-row history for other signal
- *  types (queue-health, contributor-decision-pack, ...). Single source of truth: if this list changes, the
- *  probe's ratio scope changes with it automatically. */
+ *  per key -- NOT the whole table, which intentionally keeps bounded multi-row history for the one signal
+ *  type genuinely read as a trend/change series (queue-health). Single source of truth: if this list
+ *  changes, the probe's ratio scope changes with it automatically. */
 export const LATEST_ONLY_SIGNAL_SNAPSHOT_TYPES = [
   "repo-culture-profile",
   "repo-doc-refresh-attempt",
@@ -205,12 +205,25 @@ export const LATEST_ONLY_SIGNAL_SNAPSHOT_TYPES = [
   // weeks at current review volume, refilling D1's 10GB cap before the 90-day age window could ever
   // engage. No reader consumes them as a series (the canonical latest lives in the dedicated
   // contributor_evidence / contributor_scoring_profiles upsert tables; nothing calls
-  // listSignalSnapshots for contributor-* types), so latest-only is lossless for every actual consumer.
-  // contributor-decision-pack stays EXCLUDED: the retention doc above records it as a bounded
-  // trend/change series by design, and its volume is a fraction of these three.
+  // listSignalSnapshots for contributor-* types other than contributor-decision-pack, which itself only
+  // ever reads index [0] — see below), so latest-only is lossless for every actual consumer.
   "contributor-evidence-graph",
   "contributor-outcome-history",
   "contributor-strategy",
+  // #9435/#9459: 2026-07-27 recurrence, and the single largest signal_snapshots offender measured to
+  // date -- 6.3 GB across 18,549 rows (~350 KB/row; the profile/outcome-history/registry-activity payload
+  // each build embeds), 71% of the entire hosted D1's file size, still ENTIRELY inside the 90-day age
+  // window (the database itself is only ~65 days old, so age-based pruning had not touched a single one of
+  // these rows) and accumulating ~700-1,300 rows/day since 2026-07-06 -- refilling the 10GB cap from empty
+  // in roughly 3-4 weeks even with the #9415 fixes applied, since those addressed five OTHER tables
+  // totaling well under 1 GB combined. This entry was previously excluded by a doc comment claiming
+  // decision-pack is "a bounded trend/change series by design" with volume "a fraction of" the three
+  // contributor-* types above -- both claims were wrong: `src/services/decision-pack.ts`'s only reader
+  // (`buildContributorDecisionPack`) calls `listSignalSnapshots(...)[0]`, exactly the same latest-only
+  // contract as its neighbors, and its measured volume is ~750x theirs, not a fraction. queue-health is
+  // the one signal type that genuinely IS read as a series (src/services/maintainer-slop-duplicate-trend.ts
+  // shapes multiple weeks of queue-health snapshots into a trend card) and correctly stays excluded here.
+  "contributor-decision-pack",
   // #8900: these eight writers also INSERT a fresh row every run (persistSignalSnapshot is not an
   // upsert) while every consumer reads only index [0] / the latest row — same latest-only contract as
   // the repo-* and contributor-intelligence types above. queue-health stays EXCLUDED (feeds
@@ -230,8 +243,8 @@ export const LATEST_ONLY_SIGNAL_SNAPSHOT_TYPES = [
  * target_key) on every run rather than replacing the prior one, so within RETENTION_POLICY's 90-day
  * age window a key can accumulate hundreds of superseded snapshots (#3810 -- 342,243 rows for 2,183
  * distinct keys contributed to hitting D1's size cap). Only latest-only cache signal types are
- * deduped; historical series such as queue-health and contributor-decision-pack keep their bounded
- * RETENTION_POLICY history for trend/change readers. This keeps only the latest row per
+ * deduped; the one genuine historical series (queue-health) keeps its bounded RETENTION_POLICY history
+ * for its trend/change reader. This keeps only the latest row per
  * (signal_type, target_key), batched PER signal_type (not one table-wide window-function delete) so
  * each statement stays within D1's per-statement CPU budget -- the same batching split used during
  * the incident's manual remediation. "Latest" is the highest rowid per key: signal_snapshots is

test/unit/retention.test.ts

@@ -438,7 +438,9 @@ describe("dedupeSignalSnapshots", () => {
       await insertSignalSnapshot(env, `${signalType}-new`, signalType, "octocat", "2026-07-02T00:00:00.000Z");
       await insertSignalSnapshot(env, `${signalType}-other`, signalType, "hubot", "2026-07-02T00:00:00.000Z");
     }
-    // decision-pack is a bounded trend series by design — must remain untouched.
+    // #9435/#9459: contributor-decision-pack is NOT a bounded trend series (that earlier assumption was wrong
+    // and was itself the single largest contributor to the D1 filling up) — it dedupes to latest-per-contributor
+    // exactly like its three neighbors above; see the dedicated test below for its own coverage.
     await insertSignalSnapshot(env, "pack-1", "contributor-decision-pack", "octocat", "2026-07-01T00:00:00.000Z");
     await insertSignalSnapshot(env, "pack-2", "contributor-decision-pack", "octocat", "2026-07-02T00:00:00.000Z");
 
@@ -447,13 +449,14 @@ describe("dedupeSignalSnapshots", () => {
     expect(byType["contributor-evidence-graph"]).toBe(1);
     expect(byType["contributor-outcome-history"]).toBe(1);
     expect(byType["contributor-strategy"]).toBe(1);
+    expect(byType["contributor-decision-pack"]).toBe(1);
 
     const remaining = await env.DB.prepare("SELECT id FROM signal_snapshots ORDER BY id").all<{ id: string }>();
     const ids = (remaining.results ?? []).map((row) => row.id);
     expect(ids).toContain("contributor-strategy-new");
     expect(ids).toContain("contributor-strategy-other");
     expect(ids).not.toContain("contributor-strategy-old");
-    expect(ids).toContain("pack-1"); // series preserved
+    expect(ids).not.toContain("pack-1"); // now deduped away
     expect(ids).toContain("pack-2");
   });
 
@@ -509,19 +512,37 @@ describe("dedupeSignalSnapshots", () => {
     expect(rows.results.map((row) => row.id)).toEqual(["private-latest", "public-latest"]);
   });
 
-  it("preserves bounded history for signal types read as historical series", async () => {
+  it("preserves bounded history for queue-health, the one signal type genuinely read as a historical series", async () => {
     const env = createTestEnv();
-    await insertSignalSnapshot(env, "decision-prev", "contributor-decision-pack", "alice", "2026-06-01T00:00:00.000Z");
-    await insertSignalSnapshot(env, "decision-current", "contributor-decision-pack", "alice", "2026-06-02T00:00:00.000Z");
     await insertSignalSnapshot(env, "queue-old", "queue-health", "JSONbored/loopover", "2026-06-01T00:00:00.000Z");
     await insertSignalSnapshot(env, "queue-current", "queue-health", "JSONbored/loopover", "2026-06-02T00:00:00.000Z");
 
     expect(await dedupeSignalSnapshots(env)).toEqual([]);
 
-    expect(await countSignalSnapshots(env, "contributor-decision-pack")).toBe(2);
     expect(await countSignalSnapshots(env, "queue-health")).toBe(2);
   });
 
+  // #9435/#9459: contributor-decision-pack was previously excluded from LATEST_ONLY_SIGNAL_SNAPSHOT_TYPES on the
+  // theory that it's a bounded trend series -- it isn't (src/services/decision-pack.ts reads only index [0],
+  // the same latest-only contract as its contributor-* neighbors below) and it was, measured live, the single
+  // largest contributor to the hosted D1 filling up: 6.3 GB across 18,549 rows, 71% of the whole database, still
+  // entirely inside the 90-day age window and accumulating ~1,000 rows/day.
+  it("dedupes contributor-decision-pack alongside its contributor-* neighbors (#9435/#9459)", async () => {
+    const env = createTestEnv();
+    await insertSignalSnapshot(env, "decision-prev", "contributor-decision-pack", "alice", "2026-06-01T00:00:00.000Z");
+    await insertSignalSnapshot(env, "decision-current", "contributor-decision-pack", "alice", "2026-06-02T00:00:00.000Z");
+    await insertSignalSnapshot(env, "decision-other-prev", "contributor-decision-pack", "bob", "2026-06-01T00:00:00.000Z");
+    await insertSignalSnapshot(env, "decision-other-current", "contributor-decision-pack", "bob", "2026-06-02T00:00:00.000Z");
+
+    const results = await dedupeSignalSnapshots(env);
+    const byType = Object.fromEntries(results.map((r) => [r.signalType, r.deleted]));
+    expect(byType["contributor-decision-pack"]).toBe(2);
+
+    expect(await countSignalSnapshots(env, "contributor-decision-pack")).toBe(2);
+    const rows = await env.DB.prepare("SELECT id FROM signal_snapshots WHERE signal_type = 'contributor-decision-pack' ORDER BY id").all<{ id: string }>();
+    expect(rows.results.map((row) => row.id)).toEqual(["decision-current", "decision-other-current"]);
+  });
+
   it("deletes across multiple batches per signal_type and stops at the per-type cap", async () => {
     const env = createTestEnv();
     for (let i = 0; i < 6; i++) {