feat: add DSL output format for graph tools

[indrajeet0510]

Summary

Adds a format="dsl" parameter to get_impact_radius_tool,
query_graph_tool, and get_review_context_tool. When set, the graph
payload comes back as one-line strings instead of JSON dicts. Default
stays "dict" so nothing existing breaks.

Why

I was reading through graph.py and noticed that node_to_dict and
edge_to_dict (around line 1349) get called from roughly half of the 30
tools. Every node carries 9 JSON keys, every edge carries 8. On a
get_impact_radius that hits the default 500-node cap, that's roughly
30k tokens of structural metadata before any actual code shows up.

Two of those fields don't seem to do anything for the LLM consumer: the
internal database id, and confidence_tier (which is fully derivable
from the confidence float — same thresholds you already encode
elsewhere). And in the edge encoding, source almost always starts with
the same path as file_path, so that prefix gets duplicated on every
edge in the response.

Felt like there was real headroom to claw back without changing any
semantics:

# dict (unchanged, still the default)
{"id": 42, "kind": "Function", "name": "validateOrder",
 "qualified_name": "src/orders/validator.py::OrderValidator::validateOrder",
 "file_path": "src/orders/validator.py", "line_start": 142, "line_end": 178,
 "language": "python", "parent_name": "src/orders/validator.py::OrderValidator",
 "is_test": false}

# dsl (new, opt-in)
fn validateOrder@src/orders/validator.py:142-178 py parent=OrderValidator

Same information, about 4× fewer characters per node. Edges compress
roughly 2.5× once you strip the redundant file_path:: prefix from
source. Each response also carries a legend field documenting the
codes so the LLM doesn't need external docs to decode the lines.

Numbers

I wrote a reproducible benchmark in scripts/bench_dsl_output.py. It
builds a synthetic 50-file backend service with cross-module fan-out and
runs get_impact_radius against it. On a typical 71-node blast radius:

• dict response: ~14,800 tokens
• dsl response: ~6,700 tokens
• saves ~8,000 tokens per call

Worth flagging that per-row compression is more like 3-4×, but the
whole-response ratio comes out closer to 2.2× because the envelope
(summary, file lists, the legend itself) is identical in both modes.
On larger payloads the per-row encoding dominates and the ratio creeps
up toward the per-row number. Run the script to see for yourself.

Tests

26 new tests in tests/test_dsl_output.py covering the encoders, the
dispatch helpers, control-char sanitization (security parity with the
dict encoder — _sanitize_name is called from the DSL encoders too, so
the existing TestSanitizeName protections carry over), and the tool
integration. All pass. ruff check and mypy --ignore-missing-imports --no-strict-optional both clean on the modified files.

Heads up — there are 7 unrelated failures in tests/test_main.py
(TestApplyToolFilter and one in TestLongRunningToolsAreAsync). They
fail on clean main too, doesn't look related to this PR.

Notes

• Default of "dict" means every existing caller is byte-identical.
  Nothing changes unless someone explicitly opts in.
• I left detect_changes_tool and traverse_graph_tool alone even
  though they share the same encoder bottleneck — felt cleaner as
  separate follow-up PRs once we agree on the format here.
• Updated only the English README.md (added a short paragraph below
  the MCP tools table mentioning the new parameter). The four
  translated READMEs (Hindi, Japanese, Korean, Chinese Simplified)
  aren't updated — happy to take a follow-up for the Hindi sync if
  you'd like, but I'd defer to native speakers for the others.
• Open to bikeshedding on the field codes / legend wording / arrow
  character / etc. The current shape was a first stab at something
  compact-but-still-readable; happy to iterate if you have preferences.