feat(decisioning): surface advertise_all=False on create_adcp_server_from_platform

[bokelley]

Closes #519

Summary

Adopters using the standalone "build the handler" path saw ~40+ entries in handler.advertised_tools after create_adcp_server_from_platform() because the ClassVar holds the full union across all PlatformHandler shims. serve() already defaulted advertise_all=False and passed it through get_tools_for_handler → advertised_tools_for_instance() to filter tools/list, but that filter was invisible to anyone inspecting handler.advertised_tools directly.

This PR adds advertise_all: bool = False to create_adcp_server_from_platform() and, when False, overrides handler.advertised_tools at the instance level with set(handler.advertised_tools_for_instance()). The ClassVar (and the _HANDLER_TOOLS registry populated at class-definition time) are unaffected. serve() now forwards advertise_all into create_adcp_server_from_platform() for consistency.

Before:

handler, _, _ = create_adcp_server_from_platform(platform=MySalesPlatform())
len(handler.advertised_tools)  # ~42 (full union)

After:

handler, _, _ = create_adcp_server_from_platform(platform=MySalesPlatform())
len(handler.advertised_tools)  # 9 (sales specialism set only)

What was tested

• pytest tests/test_decisioning_serve.py — 34/34 pass (3 new tests covering default filter, advertise_all=True escape hatch, and empty-specialism fallback)
• mypy src/adcp/decisioning/serve.py — clean
• ruff check src/adcp/decisioning/serve.py — clean
• pytest tests/test_decisioning_advertised_per_specialism.py tests/test_decisioning_handler.py tests/test_advertised_tools_gate.py tests/test_register_handler_tools.py — 50/50 pass

Pre-PR review

• code-reviewer: approved — no blockers; noted that get_tools_for_handler reads _HANDLER_TOOLS[cls.__name__] (ClassVar, class-definition time) and calls advertised_tools_for_instance() independently, so the instance attribute shadow and the MCP tools/list filter are separate paths.
• dx-expert: approved — advertise_all=False is the right default and name (consistent with serve()); noted behavior change for existing callers who may have relied on the full ~40-tool ClassVar (changelog note warranted).

Nits (not fixed — addressed here):

• Test imports _SALES_ADVERTISED_TOOLS (a private symbol). Acceptable: the test is specifically verifying the specialism-to-tool mapping; using the private frozenset is the most precise assertion available.
• # type: ignore[misc] on the ClassVar instance write: mypy classifies this as [misc], not [assignment]; [misc] is the narrowest suppressor available for this specific diagnostic.

Triage-managed PR. This bot does not currently iterate on
review comments or PR conversation threads (only on the source
issue). To unblock:

• Push fixup commits directly: gh pr checkout <num> →
  fix → push.
• Or re-trigger: comment /triage execute on the source
  issue.

See adcp#3121
for context.

Session: https://claude.ai/code/session_01VYgrLhaR7HWfBf3VxbMn5K

---

Generated by Claude Code

[bokelley]

Superseded by #529 (merged) which already adds advertise_all + handler.get_advertised_tools().

[bokelley]

Acknowledged — closing the loop here. Since #529 merged with the same advertise_all surface and handler.get_advertised_tools(), this PR is correctly superseded. No further action needed on this branch.

---

Generated by Claude Code

---

Generated by Claude Code