feat(mcp): Self-healing symbol lookup errors

[Xuepoo]

Motivation

The MCP ergonomics benchmark (docs reviews/2026-08-24-mcp-ergonomics-benchmark.md) identifies symbol-not-found as the dominant visible failure class: ×20 occurrences across the corpus (fabric 7, graphology 1, konva turn-2 alone burned 12 misses probing wrong names/kinds). Current misses offer at most 3 prefix/substring hints and no file overview, forcing extra agent round-trips to recover.

Design (commander spec, both layers deterministic + byte-stable)

1. Ranked suggestions

Up to 5 candidates scored by best tier:

• tier 0: case-insensitive prefix match
• tier 1: case-insensitive substring match
• tier 2: bounded Levenshtein distance <= 2

Ties broken alphabetically; duplicates removed. No new dependencies — a small inline bounded Levenshtein with early exit (None once any DP row minimum exceeds 2, plus length-difference shortcut).

2. Embedded mini-outline on miss

The same isError message appends a capped outline of the file's top-level symbols — kind name lines sorted by byte position, MAX 20 lines, then ... and N more when truncated. Nested definitions (e.g. methods inside impl blocks) are skipped via span containment on the position-sorted symbols. Message shape:

symbol not found: X
suggestions: A, B, C

struct NodeHandle
function unrelated
... and 3 more

First line stays exactly symbol not found: X (pinned by tests). Output depends only on file content + query → byte-stable across runs.

3. Range-parse teaching

Invalid lines ranges now state the accepted formats verbatim:
invalid range \X`: expected N, N-M, or N-(existinginvalid range` contract preserved for both CLI and MCP paths).

Scope

MCP-only enrichment (the CLI symbol error is unchanged); success path untouched. Tests written red-first in tests/mcp_test.rs.

Gates

cargo fmt --check ✓ · cargo clippy --all-features -- -D warnings ✓ · cargo test (285 tests) ✓