fix: harden media buy response normalization

[bokelley]

Summary

Follow-up to #842 for the non-blocking Argus cleanup items that were worth doing now.

• centralize legacy media-buy lifecycle status inference in one helper
• stop the MCP boundary from treating draft, cancelled, or ambiguous completed as inferred media_buy_status
• avoid stamping task-like responses that already contain task_id as status="completed"
• keep UpdateMediaBuyResponse3 public for source compatibility while preserving UpdateMediaBuySubmittedResponse as the semantic alias

Validation

• PYTHONPATH=src python3 -m ruff check src/adcp/types/media_buy_status_helpers.py src/adcp/types/aliases.py src/adcp/types/__init__.py src/adcp/__init__.py src/adcp/types/generated_poc/media_buy/create_media_buy_response.py src/adcp/types/generated_poc/media_buy/update_media_buy_response.py scripts/post_generate_fixes.py tests/test_server_dx.py tests/test_schema_validation_server.py
• PYTHONPATH=src python3 -m pytest tests/test_schema_validation_server.py tests/test_server_dx.py tests/test_import_layering.py tests/test_public_api.py tests/test_create_media_buy_response_types.py tests/test_type_guards.py tests/test_code_generation.py -q
• PYTHONPATH=src python3 -m pytest tests/test_schema_validation.py tests/test_protocols.py -q
• PYTHONPATH=src python3 -m mypy src/adcp/
• PYTHONPATH=src python3 -m mypy --strict tests/type_checks/
• python3 scripts/post_generate_fixes.py
• pre-commit on commit: black, ruff, mypy, bandit, whitespace/EOF/JSON/large-file/conflict/key checks

[aao-ipr-bot]

Clean follow-up. Centralizes the legacy lifecycle inference set in one place and aligns the dict-based MCP boundary with the typed-validator behavior — coherence win.

Things I checked

• MEDIA_BUY_LEGACY_STATUS_VALUES at src/adcp/types/media_buy_status_helpers.py:13-22 matches the upstream MediaBuyStatus enum minus completed. Dropping draft is correct (not in the upstream enum at all). Dropping British cancelled is correct (upstream uses American canceled only).
• _normalize_response_envelope at src/adcp/server/mcp_tools.py:67-71 collapses to the right shape now that completed is no longer in the legacy set — the prior inner if raw_status != "completed": guard is semantically equivalent to the new code, but only because the frozenset excludes it. Worth a code-side comment so a future re-add doesn't quietly break it.
• Generated-file regeneration audit: scripts/post_generate_fixes.py:1515-1537 media_header / update_media_header produce exactly what landed in src/adcp/types/generated_poc/media_buy/create_media_buy_response.py and update_media_buy_response.py. The Sequence import shows up only in the update variant, as expected.
• Layering: new media_buy_status_helpers.py has zero generated_poc/ / _generated imports; mcp_tools.py reaches it through adcp.types (public surface). Layering rule holds.
• Validator branch ordering at create_media_buy_response.py:35-53 / update_media_buy_response.py:37-55 has no fall-through bugs; each arm copies before mutating.
• New elif raw_status == "completed": arm is the enum→string normalizer for MediaBuyStatus.completed constructor inputs. test_typed_success_does_not_infer_completed_lifecycle at tests/test_server_dx.py:189-194 and :248-251 exercises it.
• UpdateMediaBuyResponse3 re-export at src/adcp/types/aliases.py:1880 is non-breaking — already in src/adcp/__init__.py.__all__ per git show HEAD~. Intermediate layer now matches.

Follow-ups (non-blocking — file as issues)

• Warn-mode wire-conformance gap. ad-tech-protocol-expert flagged two: (a) {"task_id": ...} with no status now falls through to a status-less response (mcp_tools.py:73-74), and (b) handler-returned status="draft" ships unchanged in warn mode (tests/test_schema_validation_server.py:278-298 codifies this). The previous behavior was also wrong (synthetic completed stamp, or rewriting draft into media_buy_status), but both flavors leave the MCP boundary willing to emit a status that isn't in TaskStatus. Decide the contract: either default status="submitted" when task_id is present, or fail in both modes on a non-TaskStatus value. Same question for British cancelled — file a small follow-up if a tolerance window is intended.
• Default-status loss for non-media-buy methods. mcp_tools.py:73 changed from result.setdefault("status", "completed") (global default) to a guarded form. The guard targets the task-envelope case, but any future tool returning {"task_id": ..., "status": <something>} with status accidentally omitted will silently miss the completed default. Worth a one-line comment scoping the guard's intent.
• Changelog note for the behavior change. Adopters returning status="draft" or status="cancelled" (British) previously got auto-normalized; they now hit validation. The PR description covers this — the released changeset should too so adopters know to fix their handlers.

Minor nits (non-blocking)

1. Comment the "no-op" completed branch. create_media_buy_response.py:43-45, update_media_buy_response.py:45-47, and the mirror sites in scripts/post_generate_fixes.py:1935-1937 / :2008-2010 — elif raw_status == "completed": data["status"] = "completed" reads as a no-op until you remember it's the MediaBuyStatus.completed enum→string coercion. One-line comment saves the next reader an audit.
2. __all__ re-export style. src/adcp/types/__init__.py:790-797 uses as NAME for MEDIA_BUY_LEGACY_STATUS_VALUES but not for unwrap_enum_value. Pick one — the explicit-re-export idiom for type checkers is to use as NAME on both.
3. Test-only. tests/test_schema_validation_server.py:278-298 confirms the draft payload survives unmodified but doesn't assert the validation warning was actually logged. A caplog assertion or a strict-mode counterpart would lock in the "validation is the enforcement point" contract.

LGTM. Follow-ups noted.