docs(superpowers): v0.20.x logging r2 spec + plan artifacts

logging round 2 (4 enhancement: image_w/h + V008 SQLite mirror + CLI inspect + retention) 의 spec/plan ACCEPT 후 round artifacts.

- spec: 751 line (ACCEPT, 7/7 critic round 1 finding + 7/7 closure r2 traceability)
- plan: 576 line (ACCEPT, 6/6 step + 13/13 AC + G1/G2/G3 plan-level resolve)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-28 08:04:32 +00:00
parent 9a36a06f97
commit 7c24734cc7
2 changed files with 1327 additions and 0 deletions

View File

@@ -0,0 +1,751 @@
---
title: v0.20.x ingest log round 2 — 4 enhancement spec
created: 2026-05-28
status: DRAFT round 0
parent_spec: 2026-04-27-kebab-final-form-design.md
target_version: v0.20.x
branch: feat/ingest-log-round2-enhancements
---
# v0.20.x ingest log round 2
## §1 Motivation
### §1.1 Sweet-spot analysis — progressive dogfood tuning
v0.20.0 sub-item 1 의 round 1 ingest log (PR #189 merged) 는 per-run ndjson 의 file-only logging 으로 배포. 3개월 실사용(dogfood) 중 OCR engine 의 timeout / performance sweet-spot 을 점진적으로 조정할 필요.
- **현재 상태**: 각 ingest run 의 OCR 샘플(ms, success/fail) 이 ndjson file 에만 기록 → historical aggregate query 불가 (per-run 단위).
- **요구**: 누적 데이터베이스에서 p90 / p99 / 극값 을 조회 → timeout default 축소(e.g. 300s → 180s) 결정 가능.
- **해결**: SQLite `pdf_ocr_events` mirror table — v0.20.x round 2 주요 enhancement.
### §1.2 Image dimension 결함 — null emit 문제
`crates/kebab-app/src/pdf_ocr_apply.rs` 의 6개 emit point (line 155, 188, 265 등) 에서 `image_width: None, image_height: None` hardcode.
- **현재**: raster JPEG 가 memory 에 있으나 dimension 측정 미수행.
- **영향**: wire schema 의 optional field 이지만, 실제 use case (e.g. "100MB+ 이미지만 timeout 조정") 를 위해 필수 데이터.
- **fix**: raster decode via `image` crate (transitive dep via `kebab-parse-image`).
### §1.3 CLI inspect subcommands — 운영 visibility
ndjson log 는 human-readable 이지만, 스크립트/automation 용 corpus-wide 통계가 부족.
- `kebab inspect ocr-stats --json` — 전체 OCR 성공률, p50/p90/p99 latency, engine 별 breakdown.
- `kebab inspect ocr-failures --doc-id <id> --json` — 특정 doc 의 failure history.
- `kebab inspect ocr-failures --json` — 최근 failure 나열 (corpus-wide).
### §1.4 Log retention — 무한 증가 방지
ingest 가 반복되면서 `~/.local/state/kebab/logs/ingest-*.ndjson` file 누적.
- **현재**: 수동 정리 필수.
- **요구**: 자동 cleanup — `keep_recent_runs` (e.g. 100) + `retention_days` (e.g. 30).
- **적용**: SQLite `pdf_ocr_events` 도 동일한 retention 정책 적용.
---
## §2 Scope + non-scope
### §2.1 Included
**Enhancement 1: image_width + image_height capture (trivial)**
- raster JPEG dimension decode in `pdf_ocr_apply.rs` 6 emit point.
- image crate import (already transitive).
- Some(u32) 로 fill.
**Enhancement 2: SQLite mirror — pdf_ocr_events table (medium)**
- V008 migration: `pdf_ocr_events` table 신규 (run_id, ts, doc_id, page, image_byte_size, image_width, image_height, ms, chars, success, reason, ocr_engine).
- index: run_id, doc_id, ts.
- insert path: `IngestLogWriter` 에서 file ndjson + SQLite 동시 write (dual-write).
- SqliteStore::record_pdf_ocr_event(…) API.
**Enhancement 3: CLI inspect commands (medium)**
- `kebab inspect ocr-stats --json` — corpus-wide aggregate (total_events, success_count, failure_count, p50/p90/p99_ms, by_engine, top 10 docs by failure).
- `kebab inspect ocr-failures --doc-id <id> --json` — single doc failure list.
- `kebab inspect ocr-failures --json` (no doc-id) — corpus-wide recent failures (--limit configurable).
- wire schemas: `ocr_stats.v1` + `ocr_failures.v1` (additive minor to schema.v1).
**Enhancement 4: log retention + rotation (low)**
- `[logging] keep_recent_runs: u32` (default 100) + `retention_days: u32` (default 30).
- file cleanup: IngestLogWriter::open 시 prune helper 호출.
- SQLite cleanup: SqliteStore::prune_pdf_ocr_events(retention_days).
- backward compat: old config (no `[logging]` fields) parses with default.
### §2.2 Out of scope
- wire schema public API (pdf_ocr_events 는 internal SQLite table, wire expose 안 함).
- `ask` 명령의 한국어 phrasing-sensitive refusal (이번 round 범위 외).
- migration rollback automation (standard CLAUDE.md protocol follow).
- concurrent ingest lock manager (현재 single-process ingest 가정, future spec).
---
## §3 Design decisions
### §3.1 image_width / image_height — raster decode path
**선택**: raster JPEG 를 ImageReader 로 decode → (width, height) 추출.
- **이유**: OCR 호출 시 bytes 가 이미 memory (extract_dctdecode_page_image), decode latency << OCR latency (negligible <1ms).
- **대안 거절**: PDF MediaBox 사용 → actual raster 와 page size 다를 수 있음 (less accurate).
- **구현**: `image` crate 의 `ImageReader::new(Cursor::new(&bytes)).with_guessed_format()?.into_dimensions()?`.
- **error handling**: decode fail → (None, None) fallback. OCR 결과는 여전히 valid.
### §3.2 SQLite mirror — V008 migration + dual-write
**선택**: v0.20.0 round 1 의 file-only ndjson 을 보완하는 SQLite mirror (non-breaking).
- **이유**: historical query 를 위해 structured storage 필수. file 만으로는 corpus-wide aggregate 불가.
- **doc_id wiring**: LogEvent::Ocr 의 `doc_id` field 는 closure scope 에서 미리 capture 되어야 함. apply_ocr_to_pdf_pages 호출 전에 canonical.doc_id 를 local var 로 binding 후, closure 내에서 동일한 doc_id 로 file ndjson + SQLite insert 수행. 이를 통해 dual-write 의 일관성 보장.
- **dual-write 구조**:
1. `IngestLogWriter::write_event(&LogEvent::Ocr)` 시 file ndjson + SQLite insert.
2. insert 는 `Arc<SqliteStore>` clone 을 emit_progress closure 가 직접 호출.
3. transaction safety: file write first (failures → log), then SQLite (non-critical).
- **non-breaking**: old config 가 없어도 logging 정상 작동 (file only). SQLite 는 upgrade 시 자동 생성.
### §3.3 CLI inspect commands — ocr-stats + ocr-failures
**wire schema**: 기존 `schema.v1``wire.schemas` list 에 `ocr_stats.v1` + `ocr_failures.v1` additive minor 추가.
- **이유**: 새 wire shape 은 public API 가 아님 (inspect command 만 emit). wire.v1 의 확장으로 additive.
- **구현**: `kebab-cli/src/main.rs``Subcommand::Inspect``InspectCommand::OcrStats / OcrFailures` arm 추가.
### §3.4 retention — keep_recent_runs + retention_days
**선택**: 두 조건 모두 충족 시만 보존 (OR-on-stale = AND-on-fresh semantics).
- **이유**:
- `keep_recent_runs=100` — deterministic "최근 N 개 run 보존".
- `retention_days=30` — time-based cleanup (dogfood 중단 후 obsolete log 자동 삭제).
- **Delete if** (idx >= keep_recent) **OR** (modified <= cutoff) — 둘 중 하나라도 stale 시 삭제. 동등: **Keep iff** (idx < keep_recent) **AND** (modified > cutoff) — 둘 다 fresh 일 때만.
- **구현**: `IngestLogWriter::open()` 시 cleanup helper, `SqliteStore::prune_pdf_ocr_events(retention_days)` 별도 routine.
---
## §4 Implementation specification
### §4.1 image_width / image_height decode helper
**파일**: `crates/kebab-app/src/pdf_ocr_apply.rs`
**변경**:
1. `crates/kebab-app/Cargo.toml` dependency update:
```diff
-image = { version = "0.25", default-features = false, features = ["png"] }
+image = { version = "0.25", default-features = false, features = ["png", "jpeg"] }
```
2. import `use image::io::Reader as ImageReader;` (transitive via kebab-parse-image).
3. 새 helper function:
```rust
fn extract_image_dimensions(jpeg_bytes: &[u8]) -> Option<(u32, u32)> {
let reader = ImageReader::new(std::io::Cursor::new(jpeg_bytes))
.ok()?
.with_guessed_format()
.ok()?;
reader.into_dimensions().ok()
}
```
4. 6 emit point (line 155, 188, 265 및 PdfOcrProgress::Finished 의 다른 3곳):
```rust
let (w, h) = extract_image_dimensions(&page_image_bytes).map(|(w, h)| (Some(w), Some(h)))
.unwrap_or((None, None));
emit_progress(PdfOcrProgress::Finished {
image_width: w,
image_height: h,
...
});
```
**test**: 기존 `pdf_ocr_roundtrip` + 새 `pdf_ocr_image_dimensions` integration test.
### §4.2 V008 migration SQL + SqliteStore API
**파일**: `migrations/V008__pdf_ocr_events.sql` (신규)
```sql
CREATE TABLE pdf_ocr_events (
id INTEGER PRIMARY KEY,
run_id TEXT NOT NULL,
ts TEXT NOT NULL, -- ISO 8601 UTC (RFC 3339)
doc_id TEXT, -- nullable (file detect skip)
doc_path TEXT NOT NULL,
page INTEGER NOT NULL,
image_byte_size INTEGER,
image_width INTEGER,
image_height INTEGER,
ms INTEGER NOT NULL,
chars INTEGER NOT NULL,
success INTEGER NOT NULL, -- 0 = fail, 1 = success
reason TEXT, -- "timeout" / "ocr_error" / NULL
ocr_engine TEXT NOT NULL
);
CREATE INDEX idx_pdf_ocr_events_doc_id ON pdf_ocr_events(doc_id);
CREATE INDEX idx_pdf_ocr_events_run_id ON pdf_ocr_events(run_id);
CREATE INDEX idx_pdf_ocr_events_ts ON pdf_ocr_events(ts);
```
**파일**: `crates/kebab-store-sqlite/src/lib.rs` (SqliteStore 확장)
**신규 method** (Mutex lock 명시):
```rust
impl SqliteStore {
pub fn record_pdf_ocr_event(
&self,
run_id: &str,
ts: &str,
doc_id: Option<&str>,
doc_path: &str,
page: u32,
image_byte_size: Option<u64>,
image_width: Option<u32>,
image_height: Option<u32>,
ms: u64,
chars: u32,
success: bool,
reason: Option<&str>,
ocr_engine: &str,
) -> anyhow::Result<()> {
let conn = self.conn.lock().expect("sqlite lock poisoned");
conn.execute(
"INSERT INTO pdf_ocr_events
(run_id, ts, doc_id, doc_path, page, image_byte_size, image_width, image_height, ms, chars, success, reason, ocr_engine)
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
rusqlite::params![
run_id, ts, doc_id, doc_path, page,
image_byte_size, image_width, image_height, ms, chars,
if success { 1 } else { 0 }, reason, ocr_engine
]
)?;
Ok(())
}
pub fn prune_pdf_ocr_events(&self, retention_days: u32) -> anyhow::Result<u64> {
let conn = self.conn.lock().expect("sqlite lock poisoned");
let cutoff_ts = time::OffsetDateTime::now_utc()
.checked_sub(time::Duration::days(retention_days as i64))
.map(|dt| dt.format(&time::format_description::well_known::Rfc3339).ok())
.flatten()
.unwrap_or_default();
let n = conn.execute(
"DELETE FROM pdf_ocr_events WHERE ts < ?",
rusqlite::params![cutoff_ts],
)?;
Ok(n as u64)
}
}
```
### §4.3 wire schema — ocr_stats.v1 + ocr_failures.v1
**파일**: `docs/wire-schema/v1/ocr_stats.schema.json` (신규)
```json
{
"$schema": "http://json-schema.org/draft-07/schema#",
"title": "ocr_stats.v1",
"type": "object",
"properties": {
"schema_version": { "const": "ocr_stats.v1" },
"total_events": { "type": "integer" },
"total_runs": { "type": "integer" },
"success_count": { "type": "integer" },
"failure_count": { "type": "integer" },
"success_rate": { "type": "number" },
"p50_ms": { "type": "integer" },
"p90_ms": { "type": "integer" },
"p99_ms": { "type": "integer" },
"max_ms": { "type": "integer" },
"by_engine": { "type": "object", "additionalProperties": { "type": "integer" } },
"by_doc": {
"type": "array",
"items": {
"type": "object",
"properties": {
"doc_id": { "type": "string" },
"failure_count": { "type": "integer" },
"success_count": { "type": "integer" },
"p90_ms": { "type": ["integer", "null"] }
}
}
}
},
"required": ["schema_version", "total_events", "total_runs", "success_count", "failure_count", "success_rate"]
}
```
**파일**: `docs/wire-schema/v1/ocr_failures.schema.json` (신규)
```json
{
"$schema": "http://json-schema.org/draft-07/schema#",
"title": "ocr_failures.v1",
"type": "object",
"properties": {
"schema_version": { "const": "ocr_failures.v1" },
"doc_id": { "type": ["string", "null"] },
"failure_count": { "type": "integer" },
"failures": {
"type": "array",
"items": {
"type": "object",
"properties": {
"ts": { "type": "string" },
"page": { "type": "integer" },
"ms": { "type": "integer" },
"reason": { "type": "string" },
"image_byte_size": { "type": ["integer", "null"] }
}
}
}
},
"required": ["schema_version", "failure_count", "failures"]
}
```
**파일**: `docs/wire-schema/v1/schema.schema.json` 갱신
`schema.v1.wire.schemas` list 에 `ocr_stats.v1` + `ocr_failures.v1` 추가 (additive).
### §4.4 CLI inspect ocr-stats + ocr-failures
**파일**: `crates/kebab-cli/src/main.rs`
**신규 subcommand**:
```rust
#[derive(Subcommand)]
pub enum Subcommand {
Inspect(InspectCommand),
// ...
}
#[derive(clap::Subcommand)]
pub enum InspectCommand {
OcrStats {
#[arg(long, default_value = "false")]
json: bool,
},
OcrFailures {
#[arg(long)]
doc_id: Option<String>,
#[arg(long, default_value = "10")]
limit: usize,
#[arg(long, default_value = "false")]
json: bool,
},
}
```
**파일**: `crates/kebab-app/src/lib.rs` 확장
```rust
impl App {
pub fn inspect_ocr_stats(&self) -> anyhow::Result<OcrStatsV1> {
// SELECT 쿼리: pdf_ocr_events 에서 aggregate.
// 1. total_events, success_count, failure_count, success_rate 계산.
// 2. percentile via in-memory sort: SELECT ms FROM pdf_ocr_events WHERE success=1 ORDER BY ms.
// Vec 로 fetch 후 idx 계산 (p50 = idx 50%, p90 = idx 90%, p99 = idx 99%).
// 3. by_engine groupby (engine 별 success count).
// 4. by_doc top 10 (failure_count DESC).
}
pub fn inspect_ocr_failures(
&self,
doc_id: Option<&str>,
limit: usize,
) -> anyhow::Result<OcrFailuresV1> {
// SELECT failure records WHERE success=0.
// doc_id 있으면 WHERE doc_id=?; 없으면 ORDER BY ts DESC LIMIT limit.
}
}
```
### §4.5 retention cleanup helper — file + SQLite
**파일**: `crates/kebab-app/src/ingest_log.rs` 확장
```rust
impl IngestLogWriter {
pub fn open(cfg: &kebab_config::LoggingCfg) -> anyhow::Result<Option<Self>> {
if !cfg.ingest_log_enabled {
return Ok(None);
}
let run_id = generate_run_id();
let log_dir = expand_log_dir(&cfg.ingest_log_dir);
std::fs::create_dir_all(&log_dir)?;
// Cleanup file logs (before creating new log).
if let Err(e) = Self::cleanup_old_logs(&log_dir, cfg.keep_recent_runs, cfg.retention_days) {
tracing::warn!(target: "kebab-app", "ingest log cleanup failed: {e}");
// non-critical — continue without failing ingest.
}
let path = log_dir.join(format!("ingest-{run_id}.ndjson"));
let file = BufWriter::new(File::create(&path)?);
Ok(Some(Self { file, path, run_id, started_at: SystemTime::now() }))
}
fn cleanup_old_logs(log_dir: &Path, keep_recent: u32, retention_days: u32) -> anyhow::Result<()> {
let mut entries: Vec<_> = std::fs::read_dir(log_dir)?
.filter_map(|e| e.ok())
.filter(|e| e.path().file_name()
.and_then(|n| n.to_str())
.map(|s| s.starts_with("ingest-") && s.ends_with(".ndjson"))
.unwrap_or(false))
.collect();
// Sort by modified time descending (newest first).
entries.sort_by_key(|e| std::cmp::Reverse(e.metadata().ok().and_then(|m| m.modified().ok())));
let cutoff_time = SystemTime::now() - std::time::Duration::from_secs(retention_days as u64 * 86400);
for (idx, entry) in entries.into_iter().enumerate() {
let path = entry.path();
let metadata = entry.metadata()?;
let modified = metadata.modified()?;
// Delete if (idx >= keep_recent) OR (modified <= cutoff).
// Equivalent: keep iff (idx < keep_recent) AND (modified > cutoff) — both fresh.
// Per §3.4 OR-on-stale semantics.
if idx < keep_recent as usize && modified > cutoff_time {
continue;
}
std::fs::remove_file(&path)
.map_err(|e| anyhow::anyhow!("failed to remove {}: {}", path.display(), e))?;
}
Ok(())
}
}
```
### §4.6 Config extension — LoggingCfg
**파일**: `crates/kebab-config/src/lib.rs` (LoggingCfg 확장)
```rust
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct LoggingCfg {
pub ingest_log_enabled: bool,
pub ingest_log_dir: PathBuf,
#[serde(default = "default_keep_recent_runs")]
pub keep_recent_runs: u32,
#[serde(default = "default_retention_days")]
pub retention_days: u32,
}
fn default_keep_recent_runs() -> u32 { 100 }
fn default_retention_days() -> u32 { 30 }
impl Default for LoggingCfg {
fn default() -> Self {
Self {
ingest_log_enabled: true,
ingest_log_dir: PathBuf::from("{state_dir}/logs"),
keep_recent_runs: 100,
retention_days: 30,
}
}
}
```
**파일**: `docs/SMOKE.md` — config example 갱신
```toml
[logging]
ingest_log_enabled = true
ingest_log_dir = "{state_dir}/logs"
keep_recent_runs = 100
retention_days = 30
```
### §4.7 IngestLogWriter dual-write integration — canonical.doc_id closure capture (F1)
**파일**: `crates/kebab-app/src/ingest_log.rs` — write_event 확장
```rust
impl IngestLogWriter {
pub fn write_event_with_db(
&mut self,
event: &LogEvent<'_>,
store: Option<&SqliteStore>,
) -> anyhow::Result<()> {
// Write to file.
serde_json::to_writer(&mut self.file, event)?;
writeln!(self.file)?;
// Write to SQLite if store provided and event is Ocr.
if let (Some(store), LogEvent::Ocr {
ts, doc_id, doc_path, page, image_byte_size, image_width, image_height,
ms, chars, success, reason, ocr_engine,
}) = (store, event) {
let _ = store.record_pdf_ocr_event(
self.run_id(),
ts,
doc_id.as_deref(), // doc_id must be captured in closure scope (see below)
doc_path,
*page,
*image_byte_size,
*image_width,
*image_height,
*ms,
*chars,
*success,
*reason,
ocr_engine,
).map_err(|e| {
// Non-critical — log warning but don't fail ingest.
tracing::warn!(target: "kebab-app", "sqlite ocr event insert failed: {e}");
});
}
Ok(())
}
}
```
**caller 예시** (kebab-app/src/ingest_one_pdf_asset 또는 apply_ocr_to_pdf_pages 내 emit_progress closure):
```rust
// Pre-capture canonical.doc_id before apply_ocr_to_pdf_pages closure:
let doc_id_for_log: String = canonical.doc_id.0.clone();
let doc_path_for_log = asset.workspace_path.0.clone();
let store = Arc::new(SqliteStore::open(&cfg.storage.sqlite)?);
let mut log_writer = IngestLogWriter::open(&cfg.logging)?;
let emit_progress = move |progress: PdfOcrProgress| {
if let Some(writer) = &mut log_writer {
let event = LogEvent::Ocr {
ts: now_ts(),
doc_id: Some(doc_id_for_log.clone()), // ← captured in closure scope
doc_path: doc_path_for_log.clone(),
page: page_n,
// ... other fields ...
};
let _ = writer.write_event_with_db(&event, Some(&store));
}
};
```
---
## §5 Acceptance criteria
**AC-1**: image_width + image_height non-null after PDF OCR.
- Integration test: scanned PDF 로 `ingest_one_pdf_asset` → IngestReport check `pdf_ocr_summary.pages_ocrd > 0`, log file 의 `image_width: Some(_)`, `image_height: Some(_)` verify.
**AC-2**: V008 migration successful + `pdf_ocr_events` table exists.
- test: fresh DB 생성 → migration apply → `SELECT name FROM sqlite_master WHERE type='table' AND name='pdf_ocr_events';` verify.
**AC-3**: ingest 시 SQLite row 가 ndjson file 의 OCR record 와 1:1 일치.
- Integration test: ingest 후 `SELECT COUNT(*) FROM pdf_ocr_events WHERE success=1` = ndjson 의 `success=true` OCR line count.
**AC-4**: `kebab inspect ocr-stats --json` 정상 emit + `ocr_stats.v1` schema_version.
- CLI test: `kebab inspect ocr-stats --json | jq '.schema_version'` = `"ocr_stats.v1"`, `total_events`, `success_rate` present.
**AC-5**: `kebab inspect ocr-failures --doc-id <id> --json` 정상 emit + `ocr_failures.v1`.
- CLI test: failure 가 있는 doc_id 로 조회 → `failures[]` array non-empty, `reason` field present.
**AC-6**: `kebab inspect ocr-failures --json` (no doc-id) corpus-wide.
- CLI test: `--limit 5` 로 최근 5개 failure 반환, `failure_count >= 5`.
**AC-7**: log retention — keep_recent_runs=2 시 3rd ingest 후 oldest file deleted.
- Integration test: temp log dir, 3 ingest run with `keep_recent_runs=2` → oldest 2 file only remain.
**AC-8**: SQLite retention — retention_days=0 시 old row deleted.
- test: insert old row (ts = 90 days ago) → `prune_pdf_ocr_events(0)` → row deleted.
**AC-9**: backward compat — old config (no `[logging] retention_*` field) parses with default.
- test: pre-v0.20.x config (no `[logging]` section) → load → `logging.keep_recent_runs == 100` (default).
**AC-10**: workspace test + clippy green.
- `cargo test --workspace -j 1`, `cargo clippy --all-targets`.
**AC-11**: integration test (`ocr_inspect_smoke` + `pdf_ocr_events_insert_smoke`).
- new test binary: scanned PDF ingest → `kebab inspect ocr-stats / ocr-failures` 검증.
- `crates/kebab-store-sqlite/tests/pdf_ocr_events_insert_smoke.rs`.
**AC-12**: `[logging] retention_*` default emit in `kebab init` config.
- test: `kebab init --config /tmp/test-cfg.toml` → `[logging] keep_recent_runs = 100` + `retention_days = 30` present.
**AC-13**: wire schema additive list sync in `integrations/claude-code/kebab/SKILL.md`.
- test: `grep -c 'ocr_stats\.v1' integrations/claude-code/kebab/SKILL.md` returns ≥1, same for `ocr_failures.v1`.
---
## §6 Risks + open questions
### R-1 dual-write transaction safety (file vs SQLite race)
**Issue**: emit_progress closure 가 file write 후 SQLite insert 실패 시, ndjson 과 DB 불일치.
**Mitigation**:
- file write first (durable, may fail).
- SQLite write second (non-critical, warn on fail, don't propagate error).
- per-run 단위 reconciliation tool (future enhancement, not in scope).
### R-2 V008 migration rollback (F6)
**Issue**: user 가 v0.20.x → older version downgrade 시 V008 rolled back?
**Mitigation**: CLAUDE.md migration policy follow. V008 은 additive table → old version 이 table ignore 하면 작동 OK.
**Manual rollback** (v0.19.x ↔ v0.20.x alternating dogfood):
```sql
DELETE FROM refinery_schema_history WHERE version=8;
DROP TABLE IF EXISTS pdf_ocr_events;
```
**Out-of-scope**: v0.19.x ↔ v0.20.x alternate run 의 자동 rollback path 미제공.
### R-3 prune helper 가 concurrent ingest 시 stale lock
**Issue**: cleanup_old_logs 가 file 삭제 중인데 다른 process 가 write?
**Mitigation**:
- cleanup 은 IngestLogWriter::open 시만 (ingest 시작 전).
- per-process single ingest 가정 (현재 design).
- concurrent ingest support 는 future phase.
### R-4 image decode failure handling (corrupt JPEG fallback)
**Issue**: JPEG 가 corrupt → extract_image_dimensions 실패.
**Mitigation**: helper returns Option<(u32, u32)> → (None, None) fallback. OCR 완료도 유효. warning event push (optional, future enhancement).
### R-5 wire schema additive minor — old consumer 의 schema 미인식
**Issue**: old `kebab` binary (v0.19.x) 가 v0.20.x `kebab inspect ocr-stats` 의 output consume?
**Mitigation**:
- `schema_version` = `ocr_stats.v1` explicit (old consumer 는 schema 미인식 → skip OK).
- wire.v1 의 additive list → backward compat (old consumer 는 list 만 ignores).
- new consumer 만 `ocr_stats.v1` / `ocr_failures.v1` 인식.
### R-6 Concurrent cleanup 에 의한 log file loss
**Issue**: keep_recent_runs / retention_days 정책 으로 파일 삭제 중 user 가 tail 시도?
**Mitigation**: cleanup 은 ingest start 때만. user 가 tail 하는 중인 파일은 일반적으로 recent (top N 내) → 안전.
### R-7 doc_id NULL wiring in LogEvent::Ocr closure (F1)
**Issue**: emit_progress closure 에서 doc_id 가 None 또는 mismatch 시, file ndjson 과 SQLite record 의 doc_id 가 불일치.
**Verification** (spec §6 R-7 add):
```bash
# canonical.doc_id 가 set 되는 시점 확인
grep -n "canonical.doc_id\|\.doc_id\s*=" crates/kebab-app/src/lib.rs | head -10
```
**Mitigation**:
- closure scope 에서 doc_id 를 pre-capture (let doc_id_for_log: String = canonical.doc_id.0.clone()).
- LogEvent::Ocr 생성 시 captured value 사용.
- per-run integration test 에서 file ndjson 의 doc_id 와 SQLite SELECT 의 doc_id match verify.
---
## §7 References
- **Parent spec**: `docs/superpowers/specs/2026-04-27-kebab-final-form-design.md` (design contract).
- **Round 1 spec**: v0.20.0 sub-item 1 ingest log (PR #189).
- **Code ranges**:
- `crates/kebab-app/src/pdf_ocr_apply.rs` lines 155, 188, 265 (6 emit point).
- `crates/kebab-app/src/ingest_log.rs` (LogEvent, IngestLogWriter).
- `crates/kebab-config/src/lib.rs` (Config, LoggingCfg).
- `crates/kebab-store-sqlite/src/lib.rs` (SqliteStore).
- `crates/kebab-cli/src/main.rs` (Subcommand).
- **Dependencies**:
- `image` crate (transitive via kebab-parse-image).
- `time` crate (RFC 3339 timestamp, already in workspace).
- `rusqlite` (already in kebab-store-sqlite).
- **Config sections**:
- `[logging]`: ingest_log_enabled, ingest_log_dir, keep_recent_runs, retention_days.
- `[pdf.ocr]`: (unchanged from v0.20.0).
- **Wire schemas**:
- `docs/wire-schema/v1/ocr_stats.schema.json` (신규).
- `docs/wire-schema/v1/ocr_failures.schema.json` (신규).
- `docs/wire-schema/v1/schema.schema.json` (additive: wire.schemas list 에 두 schema 추가).
- **HOTFIXES contract**: 새로운 deviations 는 `tasks/HOTFIXES.md` 에 dated entry + cross-link to this spec.
- **Version cascade**: image_width/height, SQLite table schema 추가는 index_version cascade 아님 (chunks/embeddings 미영향).
- **Backward compat**: old config parses with `[logging]` defaults, wire schema additive minor.
---
## §8 Dependencies + imports
### Allowed dependencies
- `image` crate (for ImageReader::new, into_dimensions).
- `time` crate (RFC 3339 formatting, already in workspace).
- `rusqlite` (for SQL execute / query, already in kebab-store-sqlite).
- `serde_json` (for wire schema export, already in kebab-app).
### Forbidden dependencies
- **None new introduced.** All uses are transitive or workspace-existing.
---
## §9 Testing strategy
### Unit tests
- `extract_image_dimensions` helper: valid JPEG → Some((w, h)), corrupt JPEG → None.
- `cleanup_old_logs`: keep_recent_runs / retention_days logic, file deletion.
- LoggingCfg defaults: serde round-trip, backward compat.
### Integration tests
**New test files**:
- `crates/kebab-app/tests/ocr_inspect_smoke.rs`: scanned PDF ingest → inspect ocr-stats / ocr-failures validation.
- `crates/kebab-store-sqlite/tests/pdf_ocr_events_insert_smoke.rs`: V008 migration, dual-write, prune logic.
**Existing test updates**:
- `pdf_ocr_roundtrip` → verify image_width/height non-null.
- `ingest_report_snapshot` → verify ocr_stats output shape.
### Smoke test (docs/SMOKE.md)
- `kebab ingest` with scanned PDF → `~/.local/state/kebab/logs/ingest-*.ndjson` + `pdf_ocr_events` table check.
- `kebab inspect ocr-stats --json | jq '.schema_version'` = `"ocr_stats.v1"`.
### Regression tests
- 기존 1370 workspace test suite — regression 0 기대 (cleanup 은 non-critical, file-only logging 은 unchanged).
---
## §10 Rollout + dogfood
### v0.20.x milestone
1. **spec approval** (이번 round 0).
2. **implementation** (A6 round 1, estimate 3-4 days for 4 enhancement).
3. **review + merge** (pull request via gitea-ops).
4. **dogfood** (user runs v0.20.x binary, accumulates OCR stats over 2-4 weeks).
5. **data-driven tuning** (inspect ocr-stats → timeout default adjust, release note v0.20.y).
### Backward compat notes
- Old binary (v0.19.x) + new config (v0.20.x with `[logging] retention_*`): config parses, logging ignores new fields.
- New binary (v0.20.x) + old config (v0.19.x without `[logging]`): defaults apply, logging works.
- wire schema: additive, consumers ignore unknown fields.
### Release notes
- **wire schema additive minor** (`ocr_stats.v1`, `ocr_failures.v1` 추가) → release trigger 아님 (CLAUDE.md §Versioning cascade).
- 사용자 도그푸딩 중 데이터 누적 후, 본격 튜닝 (e.g. timeout 조정) 에 따라 `chore: bump 0.20.x → 0.20.y` 별 commit 가능 (dogfood 결과 반영 시).
---
## §11 Contract stability
**Locked sections** (design contract 의 일부, future changes require spec §N update):
- wire schema `ocr_stats.v1` field list.
- wire schema `ocr_failures.v1` field list.
- `[logging]` config fields.
- CLI `inspect ocr-stats / ocr-failures` output format.
**Flexible sections** (implementation detail, future refactor OK):
- `extract_image_dimensions` helper location (intra-crate 이동 가능).
- `cleanup_old_logs` 정책 (더 sophisticated 알고리즘 가능).
- SQLite index strategy (추가 index 가능).
---
## Summary
v0.20.x ingest log round 2 는 round 1 의 file-only ndjson 을 4 가지로 확장:
1. **image dimension capture** — raster JPEG decode (trivial).
2. **SQLite mirror** — V008 migration + `pdf_ocr_events` table (medium).
3. **CLI inspect** — corpus-wide OCR statistics API (medium).
4. **log retention** — automatic cleanup (low).
모두 non-breaking additive changes. backward compat 보장, wire schema minor bump, 500-800 line spec (이 문서 ≈ 700 lines).