v0.20.0

Fix lexical scopes in GPU callbacks and editor navigation (#266)

## TLDR
Fix lexical scope regressions in GPU callbacks and editor navigation,
refresh Osprey benchmarks, and add the Python-flavor AI guide.

## What Was Added
- Regression tests for named GPU callbacks and lexical
definition/reference resolution across Default, ML, and Python syntax.
- A Python-flavor guide for AI agents, linked from the documentation
index and included in website page coverage.

## What Was Changed/Deleted
- Named GPU helpers use their defining scope; lambdas retain lexical
captures.
- Editor navigation uses binding visibility ranges so nested bindings
and arm/lambda parameters do not capture unrelated references. Local
function navigation remains supported.
- Refresh Osprey-only benchmark results and generated tables for the
scope fixes: all 96 measurements passed and all 168 other-language
records remain unchanged.
- Consolidate repository agent instructions in AGENTS.md and clarify
GPT-2 backend usage in its README.

## How Tests Prove It Works
- Confirmed the GPU callback and nested-binding regressions fail before
the fixes and pass afterward.
- Syntax/codegen suites and all 228 LSP tests pass, along with
formatting, clippy, lint, Hawk, and duplication checks.
- Rust workspace and C runtime tests and coverage gates pass. Native
default/GC/ARC corpus, WebAssembly corpus, website tests (180), book
checks, bank E2E, docs HTML, and mobile domain tests pass.
- Local CI was not rerun for the final documentation/benchmark commits,
as requested. Hosted checks must pass before merging. Earlier local
verification is recorded above; the initial extension run had one
debugger startup timeout, and that case passed in the clean-profile
rerun.

## Spec/Doc Changes
Add the Python-flavor AI guide and refresh benchmark
documentation/results.

## Breaking Changes
None intended.

---------

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>