[BREAKING] MAINT: remove deprecations scheduled for v0.15.0

[romanlutz]

Description

Now that v0.14.0 has shipped, this PR removes every API marked removed_in="0.15.0". The deprecation warnings have been live for a full release cycle, so users have had the heads-up they need.

What goes away

MessagePiece

• Fields: originator, scorer_identifier, scores, targeted_harm_categories (the kwarg warnings tracked by _DEPRECATED_KWARGS are gone too)
• Module-level Originator alias

AttackResult / AttackResultEntry

• attack_identifier property and the before-validator that promoted attack_identifier= to atomic_attack_identifier
• AttackResultEntry.attack_identifier column (and PromptMemoryEntry.targeted_harm_categories column) dropped via Alembic migration f1a2b3c4d5e6

Memory

• MemoryInterface.export_conversations, SQLiteMemory.export_conversations, SQLiteMemory.export_all_tables, MemoryExporter (module deleted)
• get_attack_results kwargs attack_class and targeted_harm_categories, plus the per-backend _get_attack_result_harm_category_condition helpers

Targets / converters / scorers

• HuggingFaceEndpointTarget (class deleted)
• AzureMLChatTarget(message_normalizer=...)
• OpenAIImageTarget: style param, DALL-E sizes (256x256, 512x512, 1792x1024, 1024x1792), DALL-E qualities (standard, hd), URL fallback in _get_image_bytes_async
• use_entra_auth on both Azure Speech converters and on AudioTranscriptHelper / AudioTrueFalseScorer / AudioFloatScaleScorer
• AzureSpeechAudioToTextConverter.recognize_audio
• Positional arg support + x_pos/y_pos on AddTextImageConverter and AddImageTextConverter

Models / scenarios

• ScenarioStrategy.normalize_strategies
• ContentHarms / ContentHarmsStrategy aliases (module deleted)
• ChatMessage.to_json/from_json, EmbeddingResponse.to_json, ScorerMetrics.from_json

Downstream rewires

• SelectorScope and compute_technique_stats no longer accept targeted_harm_categories. The adaptive-scenario harm-category filter loses its backing query; per discussion this is accepted feature loss until it is reimplemented over the labels filter on AttackResult.
• conversation_manager queries memory for prepended-conversation scores instead of reading the removed MessagePiece.scores field. Lookup is by conversation_id so it resolves the duplicated pieces written by add_prepended_conversation_to_memory_async.
• get_prompt_scores queries ScoreEntry directly by prompt_request_response_id instead of relying on the relationship-loaded piece.scores attribute.

Things worth a careful look

• The Alembic migration drops two columns inside a batch_alter_table (for SQLite portability). Existing rows lose any data still pinned to attack_identifier or targeted_harm_categories; the load path no longer falls back to those columns, so legacy rows that only had attack_identifier will hydrate with atomic_attack_identifier=None.
• conversation_manager extraction of prepended-conversation scores now requires the user to have actually persisted those scores via add_scores_to_memory (the in-memory "attach to piece.scores" pattern is gone). Tests were updated accordingly.

Tests and Documentation

• make unit-test runs cleanly: 9311 passed, 120 skipped.
• pre-commit run --all-files is green (ruff, ty, Alembic migration check).
• Tests targeting the removed deprecation paths are deleted; tests that still need the underlying functionality (e.g. analytics, conversation-manager score extraction) were rewritten against the new flows.
• No documentation pages reference the removed APIs after the cleanup; no JupyText runs were required.