feat(engine): replace built-in bundle with vendored behavioral-anchor

[manojp99]

Summary

Replace the inbuilt amplifier-agent-builtin bundle with a vendored copy of the experimental behavioral-anchor bundle from amplifier-foundation@main:experiments/behavioral-anchor/behavioral-anchor.md.

This introduces a new agent set (explorer, architect, builder, debugger, git-ops, researcher) replacing the previous planner/coder/tester pattern. The new bundle uses a parent-level tool inheritance model via tool-delegate's context_inheritance feature, simplifying agent definitions and reducing duplication.

The Five Modifications

When vendoring behavioral-anchor, we diverged from upstream in these five specific places to preserve amplifier-agent architectural properties:

Upstream behavioral-anchor	This vendored copy	Why
no default_provider	default_provider: anthropic	Engine reads this directly from bundle frontmatter
behaviors/streaming-ui.yaml include	omitted	stdout reserved for JSON envelope (AGENTS.md invariant #5); engine handles streaming via bundle/hook_streaming.py
hooks-todo-display	omitted	same stdout contract — no progress bars on stdout
behaviors/logging.yaml include	hook-context-intelligence instead	Preserves workspace JSONL alignment with amplifier-app-cli per PR #57
hooks-approval	omitted	no wire-protocol approval round-trip yet; would deadlock on policy-driven rules
no tool-mcp	added	Preserves MCP for existing users; AAA doctor command checks for it

Architecture Impact

Agent set is flexible — Engine, protocol, runtime, both wrapper SDKs, and conformance fixtures have zero hardcoded agent names. No production code changes required beyond packaging and tests.

Tool inheritance model — The new bundle moves tools from per-agent frontmatter (old pattern) to parent-level declaration (new pattern). Agents inherit via tool-delegate.context_inheritance.enabled: true. Tests updated to validate agent definitions resolve correctly rather than checking for specific tool blocks.

Bundle cache key — Cache is keyed by sha256(bundle.md). The new manifest hash invalidates the warm pickle automatically; users see 30–90s of cold-start on first run after upgrade (foundation modules are cloned fresh).

How to Verify Locally

# Install from the current branch
uv tool install --force --reinstall --from . amplifier-agent

# Verify bundle preps cleanly and doctor passes
amplifier-agent doctor

# One full round-trip to confirm provider + streaming work
amplifier-agent run "say hi"

Test Status

Suite	Result	Notes
ruff check src/ tests/	✓ clean	no lint issues
ruff format --check	✓ clean	formatting aligned
pyright src/	1 pre-existing error	JsonDisplaySystem | CliDisplaySystem in single_turn.py:702 — verified pre-existing on main
pytest tests/ -m "not integration"	703 passed, 6 failed	6 failures are pre-existing integration test failures (--provider CLI flag missing) — verified on main
TypeScript wrapper build/test	not run	verified agent-name-agnostic in prior audit; no changes needed
Conformance suite	not run	verified agent-name-agnostic in prior audit; no changes needed

Open Items

• Design doc: Not yet written. The decisions are fully documented in this PR; can be promoted to a dated design doc (e.g. docs/designs/2026-06-15-behavioral-anchor-vendoring.md) if desired.
• Runtime verification: Tool inheritance via tool-delegate.context_inheritance should be verified in a live session (architect or builder agent using bash/filesystem tools).
• Extended thinking events: Behavioral-anchor enables extended_thinking: true on loop-streaming. Verify hook_streaming.py correctly forwards thinking blocks on the wire and that wrapper SDKs recognize them.

Architectural Reference

This change builds on the architectural predecessor: docs/designs/2026-05-19-baked-in-bundle-decision.md.

PR Status

Marked DRAFT — this is a test branch for runtime verification before merge.