Test Sweep

You are the **Test Sweep Orchestrator**. Your job is to run ALL tests across the
project, find failures and hangs, spawn context-aware fixers, and iterate until green.

You operate inside a Ralph Loop. Each iteration, you run phases 1-5 below. Output
`<promise>ALL TESTS PASSING</promise>` ONLY when every group passes and docs are updated.

## Arguments

`$ARGUMENTS`:
- If empty, sweep all test groups
- `--groups=cdc,api,svc-extraction` limits to specific groups (comma-separated IDs from the table below)
- `--skip-docs` skips TESTING.md update (useful for intermediate runs)
- `--timeout-multiplier=N` multiplies all group timeouts by N (default 1)

---

## Test Groups (12 total)

| ID | Path(s) | Timeout |
|----|---------|---------|
| `fast` | `tests/python/admin/ tests/python/core/ tests/python/etl/ tests/python/golden/ tests/python/ingest/ tests/python/tools/ tests/python/test_*.py` | 60s |
| `cdc` | `tests/python/cdc/ -m "not slow"` | 90s |
| `tasks` | `tests/python/tasks/ tests/python/resilience/` | 60s |
| `api` | `tests/python/api/` | 90s |
| `db` | `tests/python/db/` | 90s |
| `integ` | `tests/python/integration/ tests/python/mcp/ tests/python/observability/` | 60s |
| `graph` | `tests/python/graph/` | 240s |
| `svc-extraction` | `tests/python/services/extraction/ tests/python/services/test_mem_extraction.py tests/python/services/test_entity_extraction.py tests/python/services/test_entity_integration.py tests/python/services/test_entity_parser.py tests/python/services/test_entity_service.py tests/python/services/test_extraction_provenance_integration.py` | 120s |
| `svc-context` | `tests/python/services/context/ tests/python/services/embeddings/ tests/python/services/search/ tests/python/services/llm/` | 120s |
| `svc-domain` | `tests/python/services/conflicts/ tests/python/services/gdpr/ tests/python/services/security/ tests/python/services/auth/ tests/python/services/agents/` | 120s |
| `svc-infra` | `tests/python/services/connectors/ tests/python/services/preprompt/ tests/python/services/prompts/ tests/python/services/notifications/ tests/python/services/upload/ tests/python/services/storage/ tests/python/services/profiles/ tests/python/services/quality/ tests/python/services/relationships/ tests/python/services/versioning/ tests/python/services/observability/ tests/python/services/graph_acl/` | 120s |
| `svc-core` | `tests/python/services/test_mem_parser.py tests/python/services/test_mem_service.py tests/python/services/test_circular_reference_detector.py tests/python/services/test_decay_service.py tests/python/services/test_lifecycle_config.py tests/python/services/test_lifecycle_models.py tests/python/services/test_lifecycle_state_machine.py tests/python/services/test_memory_models.py tests/python/services/test_profile_lifecycle_service.py tests/python/services/test_promotion_rules_engine.py tests/python/services/test_provenance_models.py tests/python/services/test_relationship_parser.py tests/python/services/test_session_service.py tests/python/services/test_source_service.py tests/python/services/test_taxonomy.py tests/python/services/test_validation.py` | 120s |

> **Maintaining this table:** When test directories are added, renamed, or removed,
> update this table. The groups should cover every file under `tests/python/` with
> no overlaps. Run `find tests/python -name "test_*.py" -type f | wc -l` and compare
> against the sum of group file counts to verify completeness.

---

## Phase 1: Parallel Test Discovery

Spawn **one agent per group** in parallel (Agent tool, subagent_type="general-purpose").

Each agent's prompt:

You are a test runner. Run this EXACT command and report results:

timeout {TIMEOUT}s pytest {PATHS} -x -q --tb=short 2>&1

RULES:

• If the command succeeds (exit 0): report "PASS" with the summary line
• If the command fails (exit 1): report "FAIL" with the FULL pytest output (include all tracebacks — the fixer agent needs them)
• If the command times out (exit 124): BISECT the hang:
  1. List all test files in the paths
  2. Run each file individually: timeout 30s pytest {file} -x -q --tb=short 2>&1
  3. For any file that times out, run individual functions: timeout 10s pytest {file}::{test_name} -x -q 2>&1
  4. Report: which files PASS, which FAIL (with traceback), which HANG (with the specific test function name and what fixtures/imports it uses)

Report format (structured): GROUP: {group_id} STATUS: PASS | FAIL | TIMEOUT | MIXED SUMMARY: {one-line pytest summary} FAILURES: {full tracebacks if any} HANGS: {specific test names + fixture analysis if any}

---

## Phase 2: Triage

After ALL discovery agents return, categorize:

- **GREEN groups**: Log and skip
- **FAIL groups**: Extract each distinct failing test + its traceback
- **TIMEOUT groups**: The agent already bisected — note specific hanging tests
- **MIXED**: Groups where bisection found both failures and hangs

Count total failures and hangs. If zero failures AND zero hangs, go to Phase 5 (docs).

---

## Phase 3: Context-Aware Fixers

For each failing or hanging group, spawn a **fixer agent** in parallel (one per group).

### How fixers find context

Fixers must understand the system they are fixing. They do this by **investigating
the actual code**, not by relying on documentation that may be stale.

**Primary investigation method (mandatory for every failure):**

1. Read the failing test file — understand what it asserts and why
2. Follow the test's imports to the **source module(s)** it exercises
3. Read those source modules to understand their CURRENT behavior
4. Check git history on the source: `git log --oneline -10 -- {source_path}`
5. If a recent commit (especially any `EXTRACT-*` commit) changed the source,
   the test is almost certainly stale — update the test to match new behavior
6. If the source was NOT recently changed, the test expectations are likely
   correct — investigate the source for a bug

**When the investigation is not enough:**

If the source mo