Rebuild the analyze tool as a standalone platform extension, as
suggested by DOsinga in Oct 2025 (issue #5155). Single unprefixed
'analyze' tool with three auto-selected modes:

- Structure: directory overview with LOC/function/class counts,
  language breakdown
- Semantic: per-file functions with signatures and return types,
  classes with fields and inheritance, imports, call frequency (•N)
- Focused: symbol call graph with BFS incoming/outgoing chains,
  cross-file resolution, test/production caller separation

10 languages: Rust, Python, JavaScript, TypeScript, TSX, Go, Java,
Kotlin, Swift, Ruby. Tree-sitter queries extract functions, classes,
imports, and call sites. Rayon parallelism for directory analysis.

Key design decisions:
- 1 tool, 3 auto-selected modes (vs narsil-mcp's 90 tools)
- Compact LLM-optimized output format
- Function signatures with params and return types inline
- Class inheritance shown for all languages
- File-scoped call graph nodes with language isolation
- Deduplicated edges (HashSet) prevent BFS chain explosion
- Tree-formatted chain output with shared-prefix grouping
- Test callers separated from production callers
- Multiline format for large files (>10 symbols)
- Subagent delegation pattern for token budget management

Architecture: 5 files, ~2k LOC, no file over 650 LOC
- mod.rs: client, McpClientTrait impl, mode selection, tests
- parser.rs: tree-sitter parsing, signatures, inheritance extraction
- languages.rs: language registry (pure data, one entry per language)
- format.rs: output formatting, tree dedup, test separation
- graph.rs: call graph, BFS traversal, scoped resolution

Validated by:
- 3 rounds of crossfire review (final: APPROVE 8/10)
- 2 waves of agent experience (AX) testing across 7 popular OSS
  repos (FastAPI, Zod, Fiber, Rails, Spring Boot, Ktor, Vapor)
- AX understanding metric: 55-70% from analyze alone
- Academic validation: aligns with RepoGraph (ICLR 2025),
  CodeCompass (2602.20048), AST-derived graph benchmarks

Also includes analyze_cli binary for ad-hoc testing.

…ze guard error

Same-file callee resolution (graph.rs): when multiple definitions
share a name in the same file, pick the nearest by line proximity
instead of returning all matches. Fixes false cross-linking in OO
files with repeated method names.

JS/TS import normalization (parser.rs): 'import React from react'
now extracts 'react' instead of 'React from react'. Handles all
JS/TS import forms: default, named, namespace, type-only.

Swift init/deinit (parser.rs): initializers and deinitializers now
captured as function definitions with correct parent class, signatures,
and graph node registration. Calls inside init/deinit properly
attributed.

Size guard (mod.rs): output exceeding 50k chars now returns
CallToolResult::error instead of success, so agents recognize the
failure state and can retry with force or narrower scope.

Class inheritance was baked into Symbol.name (e.g. 'Child(Base)'),
breaking graph lookups — --focus Child returned 'not found' because
the graph key was 'Child(Base)'. Now inheritance lives in detail
while name stays canonical. Display still shows Child:4(Base).

Module-scope callers (graph.rs): top-level calls like Python script
invocations were silently dropped because <module> had no graph node.
Now registers a <module> pseudo-node per file and falls back to it
when no enclosing function is found.

Go receiver types (parser.rs, languages.rs): Go methods like
func (s *Server) Handle() now show as Server.Handle instead of
bare Handle. Extracts receiver type from method_declaration's
parameter list.

Rust trait impl parent (parser.rs): impl Display for MyType now
correctly attributes methods to MyType, not Display. Scans for
the 'for' keyword and picks the type after it.

Call frequency (format.rs): Self::foo and module::foo now count
toward the bare foo symbol's •N marker. Uses the same rsplit('::')
normalization as graph callee resolution.

JS inheritance (parser.rs): class Foo extends React.Component now
shows Foo(React.Component) instead of Foo(React). Takes the full
heritage expression text instead of the first descendant identifier.

Ref count in FOCUS header now always uses depth=1 graph queries
regardless of the user's follow_depth. Previously --follow 0
showed '0 refs' because BFS returned no chains to count from.
Now all depths report the same accurate count.

…m-extension-pr

* origin/main:
  Update CODEOWNERS for team restructuring (#7574)
  Add snapshot test with platform extensions (#7573)
  Handle Bedrock 'prompt is too long' error (#7550)
  feat: make pctx/Code Mode an optional dependency via 'code-mode' feature (#7567)
  chore(release): release version 1.26.0 (minor) (#7512)
  feat: allow goose askai bot to search goose codebase (#7508)
  Revert "Reapply "fix: prevent crashes in long-running Electron sessions""
  Reapply "fix: prevent crashes in long-running Electron sessions"
  Revert "fix: prevent crashes in long-running Electron sessions"
  fix: replace unwrap() with graceful error in scheduler execute_job (#7436)
  fix: Dictation API error message shows incorrect limit (#7423)
  fix(acp): Use ACP schema types for session/list (#7409)
  fix(desktop): make bundle and updater asset naming configurable (#7337)
  fix(openai): preserve order in Responses API history (#7500)
  Use the correct Goose emoji 🪿 instead of Swan in README.md (#7485)
  fix: prevent crashes in long-running Electron sessions

Chain links now show relative paths (e.g. platform_extensions/
analyze/mod.rs:call_tool:259) instead of bare filenames (mod.rs:
call_tool:259). Disambiguates repos with repeated basenames like
mod.rs, index.ts, lib.rs across different directories.

DEF lines also updated to include symbol name for consistency
with chain link format.

…m-extension-pr

* origin/main:
  Restore old system prompt behavior around subagents (#7576)