Agent benchmarks, Python indexer improvements, bump to 4.13.0

- Add agent-level benchmark suite: 25 tasks across 5 repos measuring
  Claude Code token consumption with vs without wonk (37.4% reduction)
- Update README benchmarks section with agent benchmark results
- Index Python class-level annotated assignments (e.g. pydantic fields)
- Split Python import_from_statement into per-name references
- Return Vec<Reference> from match_import_ref for multi-name imports
- Add .gitignore entry for __pycache__/

Author: etr

.gitignore

@@ -2,3 +2,4 @@
 .worktrees
 bench/repos/
 bench/results/
+__pycache__/

Cargo.lock

@@ -1,6 +1,6 @@
 [package]
 name = "wonk"
-version = "4.12.1"
+version = "4.13.0"
 edition = "2024"
 
 [lib]

Cargo.toml

@@ -82,17 +82,19 @@ Wonk pre-indexes your codebase with Tree-sitter so it understands code structure
 
 ## Benchmarks
 
-25 queries across 5 repos (ripgrep, tokio, Flask, Django, Express):
+25 code-understanding tasks across 5 real-world repos (ripgrep, tokio, httpx, pydantic, fastify), 5 runs each, median reported. Measures Claude Code token consumption with vs without wonk.
 
-| Comparison | Avg token reduction |
-|------------|---------------------|
-| wonk toon vs rg --json | **60%** |
-| wonk json vs rg --json | **57%** |
-| wonk grep vs plain rg  | **3%** (grep format is already compact) |
+| Category | Baseline (avg) | Wonk (avg) | Reduction | Quality (B→W) |
+|----------|---------------:|-----------:|----------:|--------------:|
+| symbol_location | 100k | 61k | 33% | 0.85→0.85 |
+| reference_tracing | 96k | 57k | 28% | 0.92→0.88 |
+| architecture | 162k | 101k | 29% | 0.90→0.96 |
+| multi_step | 143k | 104k | 23% | 0.93→0.93 |
+| structural | 130k | 69k | 46% | 0.95→0.88 |
 
-Definitions found in 92% of queries. Average of 73 re-export dedup annotations per query.
+**Overall:** 37.4% total reduction (median per-task 29.7%, best 68.5%). Quality maintained at 0.90 vs 0.91 baseline.
 
-Full results: [`bench/results/report.md`](bench/results/report.md)
+Full results: [`bench/results/agent_report.md`](bench/results/agent_report.md)
 
 ## Installation
 

README.md

@@ -0,0 +1,60 @@
+# Token Savings Benchmark
+
+Compares `wonk search` against raw `rg` (ripgrep) across 5 real open-source
+codebases. Both tools run from the repo root, producing `file:line:content`
+output so byte/token comparisons are apples-to-apples.
+
+## Prerequisites
+
+- `wonk` on PATH (or set `WONK=/path/to/wonk`)
+- `rg` (ripgrep) on PATH (or set `RG=/path/to/rg`)
+- `git` on PATH
+- ~2 GB disk for cloned repos
+
+## Usage
+
+```bash
+# From the repo root:
+cargo build --release
+cd bench
+./token_bench.sh
+```
+
+The script will:
+
+1. Shallow-clone 5 repos to `bench/repos/` (skips if already present)
+2. Run `wonk init` on each repo
+3. Run 5 queries per repo through both `rg` and `wonk search`
+4. Run a budget sweep (500/1000/2000/4000 tokens) for each query
+5. Print markdown tables to stdout and save to `bench/results/report.md`
+
+## What it measures
+
+| Metric | Description |
+|--------|-------------|
+| Token count | `ceil(bytes / 4)` — same heuristic wonk uses internally |
+| Line count | Number of result lines from each tool |
+| Reduction % | `(rg_tokens - wonk_tokens) / rg_tokens * 100` |
+| Defs found | Whether wonk's `-- definitions --` section appeared |
+| Dedup count | Number of `(+N other locations)` annotations |
+| Budget fit | What % of raw rg output fits in a given token budget |
+
+## Repos and queries
+
+| Repo | Language | Queries |
+|------|----------|---------|
+| BurntSushi/ripgrep | Rust | search, match, regex, parse, printer |
+| tokio-rs/tokio | Rust | spawn, runtime, task, poll, waker |
+| pallets/flask | Python | route, request, response, app, Blueprint |
+| django/django | Python | Model, QuerySet, view, middleware, Field |
+| expressjs/express | JavaScript | Router, middleware, request, response, next |
+
+## Runtime
+
+Expect ~5 minutes on first run (dominated by cloning django at ~250K LOC).
+Subsequent runs skip cloning and take ~2 minutes.
+
+## Output
+
+Results are written to `bench/results/report.md` (gitignored) and also
+printed to stdout.