More planning

SPEC.md

@@ -115,7 +115,8 @@ npm link  # Makes `gi` available globally
 │  - Normalize artifacts to unified schema                        │
 │  - Extract searchable documents (canonical text + metadata)     │
 │  - Content hashing for change detection                         │
-│  - Build relationship graph (issue↔MR↔note↔file)               │
+│  - MVP relationships: parent-child FKs + label/path associations│
+│    (full cross-entity "decision graph" is post-MVP scope)       │
 └─────────────────────────────────────────────────────────────────┘
                               │
                               ▼
@@ -159,10 +160,16 @@ npm link  # Makes `gi` available globally
 Issues and MRs support efficient bulk fetching with incremental sync:
 
 ```
-GET /projects/:id/issues?updated_after=X&order_by=updated_at&sort=asc&per_page=100
+GET /projects/:id/issues?scope=all&state=all&updated_after=X&order_by=updated_at&sort=asc&per_page=100
-GET /projects/:id/merge_requests?updated_after=X&order_by=updated_at&sort=asc&per_page=100
+GET /projects/:id/merge_requests?scope=all&state=all&updated_after=X&order_by=updated_at&sort=asc&per_page=100
 ```
 
+**Required query params for completeness:**
+- `scope=all` - include all issues/MRs, not just authored by current user
+- `state=all` - include closed items (GitLab defaults may exclude them)
+
+Without these params, the 2+ years of historical data would be incomplete.
+
 ### Dependent Resources (Per-Parent Fetch)
 
 Discussions must be fetched per-issue and per-MR. There is no bulk endpoint:
@@ -178,10 +185,25 @@ GET /projects/:id/merge_requests/:iid/discussions?per_page=100&page=N
 
 **Initial sync:**
 1. Fetch all issues (paginated, ~60 calls for 6K issues at 100/page)
-2. For EACH issue → fetch all discussions (~3K calls)
+2. For EACH issue → fetch all discussions (≥ issues_count calls + pagination overhead)
 3. Fetch all MRs (paginated, ~60 calls)
-4. For EACH MR → fetch all discussions (~3K calls)
+4. For EACH MR → fetch all discussions (≥ mrs_count calls + pagination overhead)
-5. Total: ~6,100+ API calls for initial sync
+5. Total: thousands of API calls for initial sync
+
+**API Call Estimation Formula:**
+```
+total_calls ≈ ceil(issues/100) + issues × avg_discussion_pages_per_issue
+            + ceil(mrs/100) + mrs × avg_discussion_pages_per_mr
+```
+
+Example: 3K issues, 3K MRs, average 1.2 discussion pages per parent:
+- Issue list: 30 calls
+- Issue discussions: 3,000 × 1.2 = 3,600 calls
+- MR list: 30 calls
+- MR discussions: 3,000 × 1.2 = 3,600 calls
+- **Total: ~7,260 calls**
+
+This matters for rate limit planning and setting realistic "10-20 minutes" expectations.
 
 **Incremental sync:**
 1. Fetch issues where `updated_after=cursor` (bulk)
@@ -235,10 +257,15 @@ tests/unit/db.test.ts
   ✓ enables foreign keys
 
 tests/integration/gitlab-client.test.ts
-  ✓ authenticates with valid PAT
+  ✓ (mocked) authenticates with valid PAT
-  ✓ returns 401 for invalid PAT
+  ✓ (mocked) returns 401 for invalid PAT
-  ✓ fetches project by path
+  ✓ (mocked) fetches project by path
-  ✓ handles rate limiting (429) with retry
+  ✓ (mocked) handles rate limiting (429) with retry
+
+tests/live/gitlab-client.live.test.ts (optional, gated by GITLAB_LIVE_TESTS=1, not in CI)
+  ✓ authenticates with real PAT against configured baseUrl
+  ✓ fetches real project by path
+  ✓ handles actual rate limiting behavior
 
 tests/integration/app-lock.test.ts
   ✓ acquires lock successfully
@@ -256,6 +283,7 @@ tests/integration/init.test.ts
   ✓ fails if any project path not found
   ✓ prompts before overwriting existing config
   ✓ respects --force to skip confirmation
+  ✓ generates gi.config.json with sensible defaults
 ```
 
 **Manual CLI Smoke Tests:**
@@ -269,6 +297,7 @@ tests/integration/init.test.ts
 | `gi init` (config exists) | Confirmation prompt | Warns before overwriting |
 | `gi --help` | Command list | Shows all available commands |
 | `gi version` | Version number | Shows installed version |
+| `gi sync-status` | Last sync time, cursor positions | Shows successful last run |
 
 **Data Integrity Checks:**
 - [ ] `projects` table contains rows for each configured project path
@@ -332,6 +361,13 @@ tests/integration/init.test.ts
 }
 ```
 
+**Raw Payload Compression:**
+- When `storage.compressRawPayloads: true` (default), raw JSON payloads are gzip-compressed before storage
+- `raw_payloads.content_encoding` indicates `'identity'` (uncompressed) or `'gzip'` (compressed)
+- Compression typically reduces storage by 70-80% for JSON payloads
+- Decompression is handled transparently when reading payloads
+- Tradeoff: Slightly higher CPU on write/read, significantly lower disk usage
+
 **DB Runtime Defaults (Checkpoint 0):**
 - On every connection:
   - `PRAGMA journal_mode=WAL;`
@@ -387,13 +423,20 @@ CREATE TABLE raw_payloads (
   source TEXT NOT NULL,          -- 'gitlab'
   project_id INTEGER REFERENCES projects(id),  -- nullable for instance-level resources
   resource_type TEXT NOT NULL,   -- 'project' | 'issue' | 'mr' | 'note' | 'discussion'
-  gitlab_id INTEGER NOT NULL,
+  gitlab_id TEXT NOT NULL,       -- TEXT because discussion IDs are strings; numeric IDs stored as strings
   fetched_at INTEGER NOT NULL,
   content_encoding TEXT NOT NULL DEFAULT 'identity', -- 'identity' | 'gzip'
   payload BLOB NOT NULL          -- raw JSON or gzip-compressed JSON
 );
 CREATE INDEX idx_raw_payloads_lookup ON raw_payloads(project_id, resource_type, gitlab_id);
 CREATE INDEX idx_raw_payloads_history ON raw_payloads(project_id, resource_type, gitlab_id, fetched_at);
+
+-- Schema version tracking for migrations
+CREATE TABLE schema_version (
+  version INTEGER PRIMARY KEY,
+  applied_at INTEGER NOT NULL,
+  description TEXT
+);
 ```
 
 ---
@@ -411,7 +454,8 @@ tests/unit/issue-transformer.test.ts
 tests/unit/pagination.test.ts
   ✓ fetches all pages when multiple exist
   ✓ respects per_page parameter
-  ✓ stops when empty page returned
+  ✓ follows X-Next-Page header until empty/absent
+  ✓ falls back to empty-page stop if headers missing (robustness)
 
 tests/unit/discussion-transformer.test.ts
   ✓ transforms discussion payload to normalized schema
@@ -450,7 +494,7 @@ tests/integration/sync-runs.test.ts
 | `gi list issues --limit=10` | Table of 10 issues | Shows iid, title, state, author |
 | `gi list issues --project=group/project-one` | Filtered list | Only shows issues from that project |
 | `gi count issues` | `Issues: 1,234` (example) | Count matches GitLab UI |
-| `gi show issue 123` | Issue detail view | Shows title, description, labels, discussions, URL |
+| `gi show issue 123` | Issue detail view | Shows title, description, labels, discussions, URL. If multiple projects have issue #123, prompts for clarification or use `--project=PATH` |
 | `gi count discussions --type=issue` | `Issue Discussions: 5,678` | Non-zero count |
 | `gi count notes --type=issue` | `Issue Notes: 12,345 (excluding 2,345 system)` | Non-zero count |
 | `gi sync-status` | Last sync time, cursor positions | Shows successful last run |
@@ -543,7 +587,7 @@ CREATE TABLE discussions (
   gitlab_discussion_id TEXT NOT NULL,         -- GitLab's string ID (e.g. "6a9c1750b37d...")
   project_id INTEGER NOT NULL REFERENCES projects(id),
   issue_id INTEGER REFERENCES issues(id),
-  merge_request_id INTEGER REFERENCES merge_requests(id),
+  merge_request_id INTEGER,                   -- FK added in CP2 via ALTER TABLE
   noteable_type TEXT NOT NULL,                -- 'Issue' | 'MergeRequest'
   individual_note BOOLEAN NOT NULL,           -- standalone comment vs threaded discussion
   first_note_at INTEGER,                      -- for ordering discussions
@@ -686,6 +730,15 @@ CREATE INDEX idx_mr_labels_label ON mr_labels(label_id);
 -- Additional indexes for DiffNote queries (tables created in CP1)
 CREATE INDEX idx_notes_type ON notes(type);
 CREATE INDEX idx_notes_new_path ON notes(position_new_path);
+
+-- Migration: Add FK constraint to discussions table (was deferred from CP1)
+-- SQLite doesn't support ADD CONSTRAINT, so we recreate the table with FK
+-- This is handled by the migration system; pseudocode for clarity:
+-- 1. CREATE TABLE discussions_new with REFERENCES merge_requests(id)
+-- 2. INSERT INTO discussions_new SELECT * FROM discussions
+-- 3. DROP TABLE discussions
+-- 4. ALTER TABLE discussions_new RENAME TO discussions
+-- 5. Recreate indexes
 ```
 
 **MR Discussion Processing Rules:**
@@ -696,8 +749,8 @@ CREATE INDEX idx_notes_new_path ON notes(position_new_path);
 
 ---
 
-### Checkpoint 3: Document + Embedding Generation with Lexical Search
+### Checkpoint 3A: Document Generation + FTS (Lexical Search)
-**Deliverable:** Documents and embeddings generated; `gi search --mode=lexical` works end-to-end
+**Deliverable:** Documents generated + FTS5 index; `gi search --mode=lexical` works end-to-end (no Ollama required)
 
 **Automated Tests (Vitest):**
 ```
@@ -707,76 +760,67 @@ tests/unit/document-extractor.test.ts
   ✓ extracts discussion document with full thread context
   ✓ includes parent issue/MR title in discussion header
   ✓ formats notes with author and timestamp
-  ✓ truncates content exceeding 8000 tokens
+  ✓ excludes system notes from discussion documents by default
+  ✓ includes system notes only when --include-system-notes enabled (debug)
+  ✓ truncates content exceeding 8000 tokens at note boundaries
   ✓ preserves first and last notes when truncating middle
   ✓ computes SHA-256 content hash consistently
 
-tests/unit/embedding-client.test.ts
-  ✓ connects to Ollama API
-  ✓ generates embedding for text input
-  ✓ returns 768-dimension vector
-  ✓ handles Ollama connection failure gracefully
-  ✓ batches requests (32 documents per batch)
-
 tests/integration/document-creation.test.ts
   ✓ creates document for each issue
   ✓ creates document for each MR
   ✓ creates document for each discussion
   ✓ populates document_labels junction table
   ✓ computes content_hash for each document
+  ✓ excludes system notes from discussion content
 
-tests/integration/embedding-storage.test.ts
+tests/integration/fts-index.test.ts
-  ✓ stores embedding in sqlite-vss
+  ✓ documents_fts row count matches documents
-  ✓ embedding rowid matches document id
+  ✓ FTS triggers fire on insert/update/delete
-  ✓ creates embedding_metadata record
+  ✓ updates propagate via triggers
-  ✓ skips re-embedding when content_hash unchanged
+
-  ✓ re-embeds when content_hash changes
+tests/integration/fts-search.test.ts
+  ✓ returns exact keyword matches
+  ✓ porter stemming works (search/searching)
+  ✓ returns empty for non-matching query
 ```
 
 **Manual CLI Smoke Tests:**
 | Command | Expected Output | Pass Criteria |
 |---------|-----------------|---------------|
-| `gi embed --all` | Progress bar with ETA | Completes without error |
+| `gi generate-docs` | Progress bar, final count | Completes without error |
-| `gi embed --all` (re-run) | `0 documents to embed` | Skips already-embedded docs |
+| `gi generate-docs` (re-run) | `0 documents to regenerate` | Skips unchanged docs |
-| `gi stats` | Embedding coverage stats | Shows 100% coverage |
+| `gi search "authentication" --mode=lexical` | FTS results | Returns matching documents, works without Ollama |
-| `gi stats --json` | JSON stats object | Valid JSON with document/embedding counts |
+| `gi stats` | Document count stats | Shows document coverage |
-| `gi embed --all` (Ollama stopped) | Clear error message | Non-zero exit, actionable error |
-| `gi search "authentication" --mode=lexical` | FTS results | Returns matching documents, no embeddings required |
 
 **Data Integrity Checks:**
 - [ ] `SELECT COUNT(*) FROM documents` = issues + MRs + discussions
-- [ ] `SELECT COUNT(*) FROM embeddings` = `SELECT COUNT(*) FROM documents`
+- [ ] `SELECT COUNT(*) FROM documents_fts` = `SELECT COUNT(*) FROM documents` (via FTS triggers)
-- [ ] `SELECT COUNT(*) FROM embedding_metadata` = `SELECT COUNT(*) FROM documents`
-- [ ] All `embedding_metadata.content_hash` matches corresponding `documents.content_hash`
 - [ ] `SELECT COUNT(*) FROM documents WHERE LENGTH(content_text) > 32000` logs truncation warnings
 - [ ] Discussion documents include parent title in content_text
+- [ ] Discussion documents exclude system notes
 
 **Scope:**
-- Ollama integration (nomic-embed-text model)
-- Embedding generation pipeline:
-  - Batch size: 32 documents per batch
-  - Concurrency: configurable (default 4 workers)
-  - Retry with exponential backoff for transient failures (max 3 attempts)
-  - Per-document failure recording to enable targeted re-runs
-- Vector storage in SQLite (sqlite-vss extension)
-- Progress tracking and resumability
 - Document extraction layer:
   - Canonical "search documents" derived from issues/MRs/discussions
   - Stable content hashing for change detection (SHA-256 of content_text)
-  - Single embedding per document (chunking deferred to post-MVP)
+  - Truncation: content_text capped at 8000 tokens at NOTE boundaries
-  - Truncation: content_text capped at 8000 tokens (nomic-embed-text limit is 8192)
     - **Implementation:** Use character budget, not exact token count
     - `maxChars = 32000` (conservative 4 chars/token estimate)
+    - Drop whole notes from middle, never cut mid-note
     - `approxTokens = ceil(charCount / 4)` for reporting/logging only
-    - This avoids tokenizer dependency while preventing embedding failures
+- System notes excluded from discussion documents (stored in DB for audit, but not in embeddings/search)
 - Denormalized metadata for fast filtering (author, labels, dates)
 - Fast label filtering via `document_labels` join table
-- FTS5 index for lexical search (enables `gi search --mode=lexical` without Ollama)
+- FTS5 index for lexical search
 - `gi search --mode=lexical` CLI command (works without Ollama)
 
-**Schema Additions:**
+This checkpoint delivers a working search experience before introducing embedding infrastructure risk.
+
+**Schema Additions (CP3A):**
 ```sql
 -- Unified searchable documents (derived from issues/MRs/discussions)
+-- Note: Full documents table schema is in CP3B section for continuity with embeddings
 CREATE TABLE documents (
   id INTEGER PRIMARY KEY,
   source_type TEXT NOT NULL,     -- 'issue' | 'merge_request' | 'discussion'
@@ -806,7 +850,122 @@ CREATE TABLE document_labels (
 );
 CREATE INDEX idx_document_labels_label ON document_labels(label_name);
 
--- sqlite-vss virtual table
+-- Fast path filtering for documents (extracted from DiffNote positions)
+CREATE TABLE document_paths (
+  document_id INTEGER NOT NULL REFERENCES documents(id),
+  path TEXT NOT NULL,
+  PRIMARY KEY(document_id, path)
+);
+CREATE INDEX idx_document_paths_path ON document_paths(path);
+
+-- Track sources that require document regeneration (populated during ingestion)
+CREATE TABLE dirty_sources (
+  source_type TEXT NOT NULL,     -- 'issue' | 'merge_request' | 'discussion'
+  source_id INTEGER NOT NULL,    -- local DB id
+  queued_at INTEGER NOT NULL,
+  PRIMARY KEY(source_type, source_id)
+);
+
+-- Resumable dependent fetches (discussions are per-parent resources)
+CREATE TABLE pending_discussion_fetches (
+  project_id INTEGER NOT NULL REFERENCES projects(id),
+  noteable_type TEXT NOT NULL,            -- 'Issue' | 'MergeRequest'
+  noteable_iid INTEGER NOT NULL,          -- parent iid (stable human identifier)
+  queued_at INTEGER NOT NULL,
+  attempt_count INTEGER NOT NULL DEFAULT 0,
+  last_attempt_at INTEGER,
+  last_error TEXT,
+  PRIMARY KEY(project_id, noteable_type, noteable_iid)
+);
+CREATE INDEX idx_pending_discussions_retry
+  ON pending_discussion_fetches(attempt_count, last_attempt_at)
+  WHERE last_error IS NOT NULL;
+
+-- Full-text search for lexical retrieval
+-- Using porter stemmer for better matching of word variants
+CREATE VIRTUAL TABLE documents_fts USING fts5(
+  title,
+  content_text,
+  content='documents',
+  content_rowid='id',
+  tokenize='porter unicode61'
+);
+
+-- Triggers to keep FTS in sync
+CREATE TRIGGER documents_ai AFTER INSERT ON documents BEGIN
+  INSERT INTO documents_fts(rowid, title, content_text)
+  VALUES (new.id, new.title, new.content_text);
+END;
+
+CREATE TRIGGER documents_ad AFTER DELETE ON documents BEGIN
+  INSERT INTO documents_fts(documents_fts, rowid, title, content_text)
+  VALUES('delete', old.id, old.title, old.content_text);
+END;
+
+CREATE TRIGGER documents_au AFTER UPDATE ON documents BEGIN
+  INSERT INTO documents_fts(documents_fts, rowid, title, content_text)
+  VALUES('delete', old.id, old.title, old.content_text);
+  INSERT INTO documents_fts(rowid, title, content_text)
+  VALUES (new.id, new.title, new.content_text);
+END;
+```
+
+**FTS5 Tokenizer Notes:**
+- `porter` enables stemming (searching "authentication" matches "authenticating", "authenticated")
+- `unicode61` handles Unicode properly
+- Code identifiers (snake_case, camelCase, file paths) may not tokenize ideally; post-MVP consideration for custom tokenizer
+
+---
+
+### Checkpoint 3B: Embedding Generation (Semantic Search)
+**Deliverable:** Embeddings generated + `gi search --mode=semantic` works; graceful fallback if Ollama unavailable
+
+**Automated Tests (Vitest):**
+```
+tests/unit/embedding-client.test.ts
+  ✓ connects to Ollama API
+  ✓ generates embedding for text input
+  ✓ returns 768-dimension vector
+  ✓ handles Ollama connection failure gracefully
+  ✓ batches requests (32 documents per batch)
+
+tests/integration/embedding-storage.test.ts
+  ✓ stores embedding in sqlite-vss
+  ✓ embedding rowid matches document id
+  ✓ creates embedding_metadata record
+  ✓ skips re-embedding when content_hash unchanged
+  ✓ re-embeds when content_hash changes
+```
+
+**Manual CLI Smoke Tests:**
+| Command | Expected Output | Pass Criteria |
+|---------|-----------------|---------------|
+| `gi embed --all` | Progress bar with ETA | Completes without error |
+| `gi embed --all` (re-run) | `0 documents to embed` | Skips already-embedded docs |
+| `gi stats` | Embedding coverage stats | Shows 100% coverage |
+| `gi stats --json` | JSON stats object | Valid JSON with document/embedding counts |
+| `gi embed --all` (Ollama stopped) | Clear error message | Non-zero exit, actionable error |
+| `gi search "authentication" --mode=semantic` | Vector results | Returns semantically similar documents |
+
+**Data Integrity Checks:**
+- [ ] `SELECT COUNT(*) FROM embeddings` = `SELECT COUNT(*) FROM documents`
+- [ ] `SELECT COUNT(*) FROM embedding_metadata` = `SELECT COUNT(*) FROM documents`
+- [ ] All `embedding_metadata.content_hash` matches corresponding `documents.content_hash`
+
+**Scope:**
+- Ollama integration (nomic-embed-text model)
+- Embedding generation pipeline:
+  - Batch size: 32 documents per batch
+  - Concurrency: configurable (default 4 workers)
+  - Retry with exponential backoff for transient failures (max 3 attempts)
+  - Per-document failure recording to enable targeted re-runs
+- Vector storage in SQLite (sqlite-vss extension)
+- Progress tracking and resumability
+- `gi search --mode=semantic` CLI command
+
+**Schema Additions (CP3B):**
+```sql
+-- sqlite-vss virtual table for vector search
 -- Storage rule: embeddings.rowid = documents.id
 CREATE VIRTUAL TABLE embeddings USING vss0(
   embedding(768)
@@ -828,22 +987,6 @@ CREATE TABLE embedding_metadata (
 
 -- Index for finding failed embeddings to retry
 CREATE INDEX idx_embedding_metadata_errors ON embedding_metadata(last_error) WHERE last_error IS NOT NULL;
-
--- Track sources that require document regeneration (populated during ingestion)
-CREATE TABLE dirty_sources (
-  source_type TEXT NOT NULL,     -- 'issue' | 'merge_request' | 'discussion'
-  source_id INTEGER NOT NULL,    -- local DB id
-  queued_at INTEGER NOT NULL,
-  PRIMARY KEY(source_type, source_id)
-);
-
--- Fast path filtering for documents (extracted from DiffNote positions)
-CREATE TABLE document_paths (
-  document_id INTEGER NOT NULL REFERENCES documents(id),
-  path TEXT NOT NULL,
-  PRIMARY KEY(document_id, path)
-);
-CREATE INDEX idx_document_paths_path ON document_paths(path);
 ```
 
 **Storage Rule (MVP):**
@@ -879,6 +1022,12 @@ Agreed. What about refresh token strategy?
 Short-lived access tokens (15min), longer refresh (7 days). Here's why...
 ```
 
+**System Notes Exclusion Rule:**
+- System notes (is_system=1) are stored in the DB for audit purposes
+- System notes are EXCLUDED from discussion documents by default
+- This prevents semantic noise ("changed assignee", "added label", "mentioned in") from polluting embeddings
+- Debug flag `--include-system-notes` available for troubleshooting
+
 This format preserves:
 - Parent context (issue/MR title and number)
 - Project path for scoped search
@@ -889,14 +1038,28 @@ This format preserves:
 - Temporal ordering of the conversation
 - Full thread semantics for decision traceability
 
-**Truncation:**
+**Truncation (Note-Boundary Aware):**
-If content exceeds 8000 tokens:
+If content exceeds 8000 tokens (~32000 chars):
-**Note:** Token count is approximate (`ceil(charCount / 4)`). Enforce `maxChars = 32000`.
 
-1. Truncate from the middle (preserve first + last notes for context)
+**Algorithm:**
-2. Set `documents.is_truncated = 1`
+1. Count non-system notes in the discussion
-3. Set `documents.truncated_reason = 'token_limit_middle_drop'`
+2. If total chars ≤ maxChars, no truncation needed
-4. Log a warning with document ID and original token count
+3. Otherwise, drop whole notes from the MIDDLE:
+   - Preserve first N notes and last M notes
+   - Never cut mid-note (produces unreadable snippets and worse embeddings)
+   - Continue dropping middle notes until under maxChars
+4. Insert marker: `\n\n[... N notes omitted for length ...]\n\n`
+5. Set `documents.is_truncated = 1`
+6. Set `documents.truncated_reason = 'token_limit_middle_drop'`
+7. Log a warning with document ID and original/truncated token count
+
+**Why note-boundary truncation:**
+- Cutting mid-note produces unreadable snippets ("...the authentication flow because--")
+- Keeping whole notes preserves semantic coherence for embeddings
+- First notes contain context/problem statement; last notes contain conclusions
+- Middle notes are often back-and-forth that's less critical
+
+**Token estimation:** `approxTokens = ceil(charCount / 4)`. No tokenizer dependency.
 
 This metadata enables:
 - Monitoring truncation frequency in production
@@ -954,14 +1117,14 @@ tests/e2e/golden-queries.test.ts
 | `gi search "authentication" --author=johndoe` | Filtered by author | All results have @johndoe |
 | `gi search "authentication" --after=2024-01-01` | Date filtered | All results after date |
 | `gi search "authentication" --label=bug` | Label filtered | All results have bug label |
-| `gi search "redis" --mode=lexical` | FTS results only | Works without Ollama |
+| `gi search "redis" --mode=lexical` | FTS results only | Shows FTS results, no embeddings |
 | `gi search "auth" --path=src/auth/` | Path-filtered results | Only results referencing files in src/auth/ |
 | `gi search "authentication" --json` | JSON output | Valid JSON matching stable schema |
 | `gi search "authentication" --explain` | Rank breakdown | Shows vector/FTS/RRF contributions |
 | `gi search "authentication" --limit=5` | 5 results max | Returns at most 5 results |
 | `gi search "xyznonexistent123"` | No results message | Graceful empty state |
-| `gi search "auth"` (no data synced) | No data message | Shows "Run gi sync first" |
+| `gi search "authentication"` (no data synced) | No data message | Shows "Run gi sync first" |
-| `gi search "auth"` (Ollama stopped) | FTS results + warning | Shows warning, still returns results |
+| `gi search "authentication"` (Ollama stopped) | FTS results + warning | Shows warning, still returns results |
 
 **Golden Query Test Suite:**
 Create `tests/fixtures/golden-queries.json` with 10 queries and expected URLs:
@@ -991,8 +1154,12 @@ Each query must have at least one expected URL appear in top 10 results.
 - Result ranking and scoring (document-level)
 - Search filters: `--type=issue|mr|discussion`, `--author=username`, `--after=date`, `--label=name`, `--project=path`, `--path=file`, `--limit=N`
   - `--limit=N` controls result count (default: 20, max: 100)
-  - `--path` filters documents by referenced file paths (from DiffNote positions)
+  - `--path` filters documents by referenced file paths (from DiffNote positions):
-  - MVP: substring/exact match; glob patterns deferred
+    - If `--path` ends with `/`: prefix match (`path LIKE 'src/auth/%'`)
+    - Otherwise: exact match OR prefix on directory boundary
+    - Examples: `--path=src/auth/` matches `src/auth/login.ts`, `src/auth/utils/helpers.ts`
+    - Examples: `--path=src/auth/login.ts` matches only that exact file
+    - Glob patterns deferred to post-MVP
   - Label filtering operates on `document_labels` (indexed, exact-match)
   - Filters work identically in hybrid and lexical modes
 - Debug: `--explain` returns rank contributions from vector + FTS + RRF
@@ -1005,51 +1172,25 @@ Each query must have at least one expected URL appear in top 10 results.
   - Filters exclude all results: `No results match the specified filters.`
   - Helpful hints shown in non-JSON mode (e.g., "Try broadening your search")
 
-**Schema Additions:**
-```sql
--- Full-text search for hybrid retrieval
--- Using porter stemmer for better matching of word variants
-CREATE VIRTUAL TABLE documents_fts USING fts5(
-  title,
-  content_text,
-  content='documents',
-  content_rowid='id',
-  tokenize='porter unicode61'
-);
-
--- Triggers to keep FTS in sync
-CREATE TRIGGER documents_ai AFTER INSERT ON documents BEGIN
-  INSERT INTO documents_fts(rowid, title, content_text)
-  VALUES (new.id, new.title, new.content_text);
-END;
-
-CREATE TRIGGER documents_ad AFTER DELETE ON documents BEGIN
-  INSERT INTO documents_fts(documents_fts, rowid, title, content_text)
-  VALUES('delete', old.id, old.title, old.content_text);
-END;
-
-CREATE TRIGGER documents_au AFTER UPDATE ON documents BEGIN
-  INSERT INTO documents_fts(documents_fts, rowid, title, content_text)
-  VALUES('delete', old.id, old.title, old.content_text);
-  INSERT INTO documents_fts(rowid, title, content_text)
-  VALUES (new.id, new.title, new.content_text);
-END;
-```
-
-**FTS5 Tokenizer Notes:**
-- `porter` enables stemming (searching "authentication" matches "authenticating", "authenticated")
-- `unicode61` handles Unicode properly
-- Code identifiers (snake_case, camelCase, file paths) may not tokenize ideally; post-MVP consideration for custom tokenizer
-
 **Hybrid Search Algorithm (MVP) - Reciprocal Rank Fusion:**
-1. Query both vector index (top 50) and FTS5 (top 50)
+1. Determine recall size (adaptive based on filters):
-2. Merge results by document_id
+   - `baseTopK = 50`
-3. Combine with Reciprocal Rank Fusion (RRF):
+   - If any filters present (--project, --type, --author, --label, --path, --after): `topK = 200`
+   - This prevents "no results" when relevant docs exist outside top-50 unfiltered recall
+2. Query both vector index (top topK) and FTS5 (top topK)
+   - Apply SQL-expressible filters during retrieval when possible (project_id, author_username, source_type)
+3. Merge results by document_id
+4. Combine with Reciprocal Rank Fusion (RRF):
    - For each retriever list, assign ranks (1..N)
-   - `rrf_score = Σ 1 / (k + rank)` with k=60 (tunable)
+   - `rrfScore = Σ 1 / (k + rank)` with k=60 (tunable)
    - RRF is simpler than weighted sums and doesn't require score normalization
-4. Apply filters (type, author, date, label)
+5. Apply remaining filters (date ranges, labels, paths that weren't applied in SQL)
-5. Return top K
+6. Return top K results
+
+**Why Adaptive Recall:**
+- Fixed top-50 + filter can easily return 0 results even when relevant docs exist
+- Increasing recall when filters are present catches more candidates before filtering
+- SQL-level filtering is preferred (faster, uses indexes) but not always possible
 
 **Why RRF over Weighted Sums:**
 - FTS5 BM25 scores and vector distances use different scales
@@ -1125,17 +1266,23 @@ interface SearchResult {
   author: string | null;
   createdAt: string;       // ISO 8601
   updatedAt: string;       // ISO 8601
-  score: number;           // 0-1 normalized RRF score
+  score: number;           // normalized 0-1 (rrfScore / maxRrfScore in this result set)
   snippet: string;         // truncated content_text
   labels: string[];
   // Only present with --explain flag
   explain?: {
     vectorRank?: number;   // null if not in vector results
     ftsRank?: number;      // null if not in FTS results
-    rrfScore: number;
+    rrfScore: number;      // raw RRF score (rank-based, comparable within a query)
   };
 }
 
+// Note on score normalization:
+// - `score` is normalized 0-1 for UI display convenience
+// - Normalization is per-query (score = rrfScore / max(rrfScore) in this result set)
+// - Use `explain.rrfScore` for raw scores when comparing across queries
+// - Scores are NOT comparable across different queries
+
 interface SearchResponse {
   query: string;
   mode: "hybrid" | "lexical" | "semantic";
@@ -1186,7 +1333,7 @@ tests/integration/sync-recovery.test.ts
 | `gi sync` (no changes) | `0 issues, 0 MRs updated` | Fast completion, no API calls beyond cursor check |
 | `gi sync` (after GitLab change) | `1 issue updated, 3 discussions refetched` | Detects and syncs the change |
 | `gi sync --full` | Full sync progress | Resets cursors, fetches everything |
-| `gi sync-status` | Cursor positions, last sync time | Shows current state |
+| `gi sync-status` | Last sync time, cursor positions | Shows current state |
 | `gi sync` (with rate limit) | Backoff messages | Respects rate limits, completes eventually |
 | `gi search "new content"` (after sync) | Returns new content | New content is searchable |
 
@@ -1262,13 +1409,25 @@ gi sync-status
 
 **Orchestration steps (in order):**
 1. Acquire app lock with heartbeat
-2. Ingest delta (issues, MRs, discussions) based on cursors
+2. Ingest delta (issues, MRs) based on cursors
-   - During ingestion, INSERT into `dirty_sources` for each upserted entity
+   - For each upserted issue/MR, enqueue into `pending_discussion_fetches`
-3. Apply rolling backfill window
+   - INSERT into `dirty_sources` for each upserted issue/MR
-4. Regenerate documents for entities in `dirty_sources` (process + delete from queue)
+3. Process `pending_discussion_fetches` queue (bounded per run, retryable):
-5. Embed documents with changed content_hash
+   - Fetch discussions for each queued parent
-6. FTS triggers auto-sync (no explicit step needed)
+   - On success: upsert discussions/notes, INSERT into `dirty_sources`, DELETE from queue
-7. Release lock, record sync_run as succeeded
+   - On failure: increment `attempt_count`, record `last_error`, leave in queue for retry
+   - Bound processing: max N parents per sync run to avoid unbounded API calls
+4. Apply rolling backfill window
+5. Regenerate documents for entities in `dirty_sources` (process + delete from queue)
+6. Embed documents with changed content_hash
+7. FTS triggers auto-sync (no explicit step needed)
+8. Release lock, record sync_run as succeeded
+
+**Why queue-based discussion fetching:**
+- One pathological MR thread (huge pagination, 5xx errors, permission issues) shouldn't block the entire sync
+- Primary resource cursors can advance independently
+- Discussions can be retried without re-fetching all issues/MRs
+- Bounded processing prevents unbounded API calls per sync run
 
 Individual commands remain available for checkpoint testing and debugging:
 - `gi ingest --type=issues`
@@ -1298,7 +1457,8 @@ All commands support `--help` for detailed usage information.
 |---------|-----|-------------|
 | `gi ingest --type=issues` | 1 | Fetch issues from GitLab |
 | `gi ingest --type=merge_requests` | 2 | Fetch MRs and discussions |
-| `gi embed --all` | 3 | Generate embeddings for all documents |
+| `gi generate-docs` | 3A | Extract documents from issues/MRs/discussions |
+| `gi embed --all` | 3B | Generate embeddings for all documents |
 | `gi embed --retry-failed` | 3 | Retry failed embeddings |
 | `gi sync` | 5 | Full sync orchestration (ingest + docs + embed) |
 | `gi sync --full` | 5 | Force complete re-sync (reset cursors) |
@@ -1310,7 +1470,7 @@ All commands support `--help` for detailed usage information.
 | Command | CP | Description |
 |---------|-----|-------------|
 | `gi list issues [--limit=N] [--project=PATH]` | 1 | List issues |
-| `gi list mrs [--limit=N]` | 2 | List merge requests |
+| `gi list mrs --limit=N` | 2 | List merge requests |
 | `gi count issues` | 1 | Count issues |
 | `gi count mrs` | 2 | Count merge requests |
 | `gi count discussions --type=issue` | 1 | Count issue discussions |
@@ -1318,8 +1478,8 @@ All commands support `--help` for detailed usage information.
 | `gi count discussions --type=mr` | 2 | Count MR discussions |
 | `gi count notes --type=issue` | 1 | Count issue notes (excluding system) |
 | `gi count notes` | 2 | Count all notes (excluding system) |
-| `gi show issue <iid>` | 1 | Show issue details |
+| `gi show issue <iid> [--project=PATH]` | 1 | Show issue details (prompts if iid ambiguous across projects) |
-| `gi show mr <iid>` | 2 | Show MR details with discussions |
+| `gi show mr <iid> [--project=PATH]` | 2 | Show MR details with discussions |
 | `gi stats` | 3 | Embedding coverage statistics |
 | `gi stats --json` | 3 | JSON stats for scripting |
 | `gi sync-status` | 1 | Show cursor positions and last sync |
@@ -1401,6 +1561,8 @@ Common errors and their resolutions:
 | **Disk full during write** | Fails with clear error. Cursor preserved at last successful commit. Free space and resume. |
 | **Stale lock detected** | Lock held > 10 minutes without heartbeat is considered stale. Next sync auto-recovers. |
 | **Network interruption** | Retries with exponential backoff. After max retries, sync fails but cursor is preserved. |
+| **Embedding permanent failure** | After 3 retries, document stays in `embedding_metadata` with `last_error` populated. Use `gi embed --retry-failed` to retry later, or `gi stats` to see failed count. Documents with failed embeddings are excluded from vector search but included in FTS. |
+| **Orphaned records** | MVP: No automatic cleanup. `last_seen_at` field enables future detection of items deleted in GitLab. Post-MVP: `gi gc --dry-run` to identify orphans, `gi gc --confirm` to remove. |
 
 ---
 
@@ -1506,7 +1668,7 @@ CREATE TABLE note_positions (
   new_line INTEGER,
   position_type TEXT                          -- 'text' | 'image' | etc.
 );
-CREATE INDEX idx_note_positions_new_path ON note_positions(position_new_path);
+CREATE INDEX idx_note_positions_new_path ON note_positions(new_path);
 ```
 
 ---
@@ -1535,6 +1697,8 @@ Each checkpoint includes:
 | Search quality | Hybrid (vector + FTS5) retrieval with RRF, golden query test suite |
 | Concurrent sync corruption | DB lock + heartbeat + rolling backfill, automatic stale lock recovery |
 | Embedding failures | Per-document error tracking, retry with backoff, targeted re-runs |
+| Pathological discussions | Queue-based discussion fetching; one bad thread doesn't block entire sync |
+| Empty search results with filters | Adaptive recall (topK 50→200 when filtered) |
 
 **SQLite Performance Defaults (MVP):**
 - Enable `PRAGMA journal_mode=WAL;` on every connection
@@ -1552,23 +1716,24 @@ Each checkpoint includes:
 | sync_runs | 0 | Audit trail of sync operations (with heartbeat) |
 | app_locks | 0 | Crash-safe single-flight lock |
 | sync_cursors | 0 | Resumable sync state per primary resource |
-| raw_payloads | 0 | Decoupled raw JSON storage (with project_id) |
+| raw_payloads | 0 | Decoupled raw JSON storage (gitlab_id as TEXT) |
 | schema_version | 0 | Database migration version tracking |
 | issues | 1 | Normalized issues (unique by project+iid) |
 | labels | 1 | Label definitions (unique by project + name) |
 | issue_labels | 1 | Issue-label junction |
-| merge_requests | 2 | Normalized MRs (unique by project+iid) |
 | discussions | 1 | Discussion threads (issue discussions in CP1, MR discussions in CP2) |
 | notes | 1 | Individual comments with is_system flag (DiffNote paths added in CP2) |
+| merge_requests | 2 | Normalized MRs (unique by project+iid) |
 | mr_labels | 2 | MR-label junction |
-| documents | 3 | Unified searchable documents with truncation metadata |
+| documents | 3A | Unified searchable documents with truncation metadata |
-| document_labels | 3 | Document-label junction for fast filtering |
+| document_labels | 3A | Document-label junction for fast filtering |
-| document_paths | 3 | Fast path filtering for documents (DiffNote file paths) |
+| document_paths | 3A | Fast path filtering for documents (DiffNote file paths) |
-| dirty_sources | 3 | Queue for incremental document regeneration |
+| dirty_sources | 3A | Queue for incremental document regeneration |
-| embeddings | 3 | Vector embeddings (sqlite-vss, rowid=document_id) |
+| pending_discussion_fetches | 3A | Resumable queue for dependent discussion fetching |
-| embedding_metadata | 3 | Embedding provenance + error tracking |
+| documents_fts | 3A | Full-text search index (fts5 with porter stemmer) |
-| documents_fts | 4 | Full-text search index (fts5 with porter stemmer) |
+| embeddings | 3B | Vector embeddings (sqlite-vss, rowid=document_id) |
-| mr_files | 6 | MR file changes (deferred to File History feature) |
+| embedding_metadata | 3B | Embedding provenance + error tracking |
+| mr_files | 6 | MR file changes (deferred to post-MVP) |
 
 ---
 
@@ -1584,10 +1749,10 @@ Each checkpoint includes:
 | Labels uniqueness | **By (project_id, name)** | GitLab API returns labels as strings |
 | Sync method | **Polling only for MVP** | Webhooks add complexity; polling every 10 min is sufficient |
 | Sync safety | **DB lock + heartbeat + rolling backfill** | Prevents race conditions and missed updates |
-| Discussions sync | **Dependent resource model** | Discussions API is per-parent; refetch all when parent updates |
+| Discussions sync | **Resumable queue model** | Queue-based fetching allows one pathological thread to not block entire sync |
 | Hybrid ranking | **RRF over weighted sums** | Simpler, no score normalization needed |
 | Embedding rowid | **rowid = documents.id** | Eliminates fragile rowid mapping |
-| Embedding truncation | **8000 tokens, truncate middle** | Preserve first/last notes for context |
+| Embedding truncation | **Note-boundary aware middle drop** | Never cut mid-note; preserves semantic coherence |
 | Embedding batching | **32 docs/batch, 4 concurrent workers** | Balance throughput, memory, and error isolation |
 | FTS5 tokenizer | **porter unicode61** | Stemming improves recall |
 | Ollama unavailable | **Graceful degradation to FTS5** | Search still works without semantic matching |
@@ -1596,6 +1761,14 @@ Each checkpoint includes:
 | `gi init` validation | **Validate GitLab before writing config** | Fail fast, better UX |
 | Ctrl+C handling | **Graceful shutdown** | Finish page, commit cursor, exit cleanly |
 | Empty state UX | **Actionable messages** | Guide user to next step |
+| raw_payloads.gitlab_id | **TEXT not INTEGER** | Discussion IDs are strings; numeric IDs stored as strings |
+| GitLab list params | **Always scope=all&state=all** | Ensures all historical data including closed items |
+| Pagination | **X-Next-Page headers with empty-page fallback** | Headers are more robust than empty-page detection |
+| Integration tests | **Mocked by default, live tests optional** | Deterministic CI; live tests gated by GITLAB_LIVE_TESTS=1 |
+| Search recall with filters | **Adaptive topK (50→200 when filtered)** | Prevents "no results" when relevant docs exist outside top-50 |
+| RRF score normalization | **Per-query normalized 0-1** | score = rrfScore / max(rrfScore); raw score in explain |
+| --path semantics | **Trailing / = prefix match** | `--path=src/auth/` does prefix; otherwise exact match |
+| CP3 structure | **Split into 3A (FTS) and 3B (embeddings)** | Lexical search works before embedding infra risk |
 
 ---