feat(mason): add code-first Python tools

[junchoi-db]

What

https://metric-reports-viewer-6051921418418893.staging.aws.databricksapps.com/20260829_mason-code-first-custom-tool-df1-e2e.html

• replaces Python tool scaffolding with code-first, framework-native @tool authoring
• makes exact agent.toml entrypoints the only activation source
• adds mason tools check/run, non-blocking MASON001 undeclared-tool warnings, and dev/deploy preflight validation
• extends the E2E matrix to prove warning, error, direct-run, local-dev, deployed execution, and cleanup behavior

Why

Writing customer code and also running a CLI generator duplicated ownership. This flow keeps behavior and schema in typed Python, exposure in agent.toml, and uses the CLI for validation, execution, development, and deployment.

Design doc

• Custom Agent Developer Interface Proposal
• SDK Driven CUJ for Custom Agents Tool Creation

Proof of work

• df1 validation/dev/deploy E2E: 16/16 semantic rows passed, covering CLI and direct-agent.toml authoring across mason dev and deployed Apps for sandbox, managed MCP, the custom Python tool, and a UC function
• Validation probes: 4/4 passed — non-blocking MASON001, broken-entrypoint rejection, valid contract check, and direct custom-tool run
• Custom Python tool: MASON_PYTHON_OK through both dev lanes and both deployed App lanes
• Cleanup: 3/3 passed; both Apps were subsequently observed absent and the UC function was absent from information_schema.routines
• Metric report
• Dogfood Volume path: dbfs:/Volumes/jun_choi_test/reports/metric_reports/20260829_mason-code-first-custom-tool-df1-e2e.html
• Published HTML round-trip: byte-identical SHA-256 2a374ea79789365f2a6d75d69344ab9740ba28d6f45b13a456120b0de09f8a9e

Testing

• Mason unit tests: 228 passed; JUnit reports errors=0 failures=0 skipped=0 tests=228
• Ruff check and format: passed
• Ty: passed
• Packaged/generated runtime copies: byte-identical
• Wheel build: passed (sha256 7351535f35e6dd6d35c8c1627ab507a06dd1433ce79ec9c400ff90dfaba0e0de)
• Generated template: 19 passed across all tests runnable in this environment
• Environment limitation: five unchanged thread-bridge tests cannot execute because a minimal asyncio.to_thread(lambda: 1) hangs in both project virtual environments; plain threads and plain asyncio both pass
• GitHub Actions: 49/50 checks pass, including both Mason test jobs, Mason Ruff, Mason formatting, and Mason typechecking. The sole OpenAI typecheck failure is baseline: the exact CI command on base SHA bffee76 produces the same eight httpx/httpx2 errors and two warnings.

This pull request and its description were written by Isaac.