chore(release): v0.3.2

Recall query telemetry + hit-rate report tool. Pure instrumentation —
no scoring, schema, or recall behaviour changes. Feeds the v0.4.0 batch
ordering decision after 7 days of observation.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: tznthou

CHANGELOG.md

@@ -11,6 +11,50 @@ more like an iteration counter than a strict SemVer major).
 
 ---
 
+## [0.3.2] — 2026-05-21
+
+### Added
+
+- **`recall_query` telemetry for hit-rate analysis.** Every `GET /memory/query`
+  call now appends one JSONL row to `~/.ccrecall/recall-query.log.jsonl`:
+  truncated query (80 chars), original `queryLen`, `hitCount`, `projectId`,
+  `limit`, `maxTokens`. Append failures are swallowed — telemetry must never
+  affect endpoint response. Two opt-out knobs:
+  `CCRECALL_RECALL_TELEMETRY_OFF=1` disables writes; `CCRECALL_RECALL_TELEMETRY_PATH`
+  redirects the log (used by the test suite to keep host telemetry clean).
+
+- **`scripts/recall-hit-rate-report.ts` analysis tool.** Reads the telemetry log
+  and cross-references the `memories` table to split zero-hit queries into two
+  buckets — *literal mismatch* (keyword appears in some memory body but FTS5
+  missed it) vs *truly absent* (keyword nowhere in any memory). Output (markdown
+  or `--json`): totals, hit rate, zero-hit breakdown, per-project counts,
+  query-length percentiles, and up to 10 samples per category. Designed as the
+  L0 quick-fix that feeds the v0.4.0 batch ordering: 1 day of instrumentation +
+  7 days of observation → evidence-driven decision between `#28` surfacing UX,
+  `#15` tag first-class + Topic CJK, and `#29` scorer epistemic.
+
+### Why
+
+The cold-rate signal (96/116 memories never recalled, 4 batches of major events
+that landed in auto memory instead of ccRecall DB) needs an evidence base before
+the v0.4.0 batch can pick a direction. This release is pure instrumentation —
+no scoring, schema, or recall behaviour changes — so the 7-day observation
+period stays clean.
+
+### Notes
+
+- Privacy: query string is truncated to 80 chars before logging. `queryLen`
+  preserves the original length so we can still characterise long-query
+  distribution. Hashing was considered and rejected because hashing would
+  prevent the substring cross-reference that splits literal-mismatch from
+  truly-absent.
+- Performance: `appendFileSync` runs on the response path; E2E tests assert the
+  total round-trip stays under 100ms even with telemetry on.
+- This release does NOT flip `CCRECALL_SESSION_START_STRATEGY` default — that
+  ships with v0.4.0 once the metric evidence is in.
+
+---
+
 ## [0.3.1] — 2026-05-13
 
 ### Fixed

CHANGELOG_ZH.md

@@ -8,6 +8,44 @@ ccRecall 的重要版本變更記錄在這裡。
 
 ---
 
+## [0.3.2] — 2026-05-21
+
+### 新增
+
+- **`recall_query` telemetry,為 hit-rate 分析鋪基。** 每次 `GET /memory/query`
+  呼叫現在會 append 一筆 JSONL 到 `~/.ccrecall/recall-query.log.jsonl`:
+  truncate 過的 query(80 chars)、原始 `queryLen`、`hitCount`、`projectId`、
+  `limit`、`maxTokens`。Append 失敗會被吞掉 — telemetry 絕不能影響 endpoint
+  回應。兩個 opt-out 開關:`CCRECALL_RECALL_TELEMETRY_OFF=1` 關掉寫入;
+  `CCRECALL_RECALL_TELEMETRY_PATH` 改寫 log 路徑(test 用,避免污染 host 真實 log)。
+
+- **`scripts/recall-hit-rate-report.ts` 分析腳本。** 讀 telemetry log + 對照
+  `memories` 表,把 zero-hit query 拆兩桶 — *literal mismatch*(keyword 出現在
+  某筆 memory body 但 FTS5 沒抓到) vs *truly absent*(keyword 不存在任一
+  memory)。輸出(markdown 或 `--json`):總數、hit rate、zero-hit 拆分、
+  per-project 計數、query length 分布、每類最多 10 筆 sample。設計為 L0
+  quick-fix 餵 v0.4.0 batch 排序:1 天 instrumentation + 7 天觀察 → 用實證
+  決定 `#28` surfacing UX、`#15` tag first-class + Topic CJK、`#29` scorer
+  epistemic 之間的優先序。
+
+### 為什麼
+
+Cold rate 訊號(116 筆 memories 96 筆從未召回,4 批重大事件落到 auto memory
+而非 ccRecall DB)需要 evidence 基底才能決定 v0.4.0 batch 方向。本版純
+instrumentation — 不動 scoring、schema、recall 行為 — 讓 7 天觀察期保持乾淨。
+
+### 備註
+
+- Privacy:query 字串 log 前 truncate 到 80 chars。`queryLen` 保留原長度可看
+  長 query 分布。考慮過 hashing 但拒絕,因為 hashing 會阻擋拆分 literal
+  mismatch / truly absent 所需的 substring 對照。
+- 效能:`appendFileSync` 跑在 response path 上;E2E test 驗證 telemetry 開
+  狀態下整段 round-trip 仍 < 100ms。
+- 本版**不**翻 `CCRECALL_SESSION_START_STRATEGY` 預設值 — 等 metric 證據到位後
+  跟 v0.4.0 一起出。
+
+---
+
 ## [0.3.1] — 2026-05-13
 
 ### 修正

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tznthou/ccrecall",
-  "version": "0.3.1",
+  "version": "0.3.2",
   "description": "AI memory service for Claude Code — on-demand context injection with metacognition",
   "type": "module",
   "main": "dist/index.js",