PersonaPlex Moshi CPU speech parity

.planning/architecture/mermaid/memory_streaming.mmd

@@ -0,0 +1,35 @@
+stateDiagram-v2
+  direction TB
+  [*] --> state_uninitialized
+  state_uninitialized --> state_empty : initialize [guard_configuration_valid_] / effect_initialize_
+  state_uninitialized --> state_uninitialized : initialize [guard_configuration_invalid_] / invalid_configuration__
+  state_empty --> state_empty : initialize [always] / already_initialized__
+  state_filling --> state_filling : initialize [always] / already_initialized__
+  state_full --> state_full : initialize [always] / already_initialized__
+  state_errored --> state_empty : initialize [guard_configuration_valid_] / effect_initialize_
+  state_errored --> state_errored : initialize [guard_configuration_invalid_] / invalid_configuration__
+  state_empty --> state_full : advance [guard_capacity_one_] / wrap__
+  state_empty --> state_filling : advance [guard_capacity_many_] / advance__
+  state_filling --> state_filling : advance [guard_filling_remains_partial_] / advance__
+  state_filling --> state_full : advance [guard_filling_becomes_full_] / wrap__
+  state_full --> state_full : advance [guard_full_position_available_before_wrap_] / advance__
+  state_full --> state_full : advance [guard_full_position_available_at_wrap_] / wrap__
+  state_full --> state_full : advance [guard_full_position_overflow_] / position_overflow__
+  state_full --> state_full : advance [guard_full_cursor_invalid_] / internal_error__
+  state_empty --> state_empty : reset [always] / effect_reset_
+  state_filling --> state_empty : reset [always] / effect_reset_
+  state_full --> state_empty : reset [always] / effect_reset_
+  state_empty --> state_empty : capture_view [always] / empty__
+  state_filling --> state_filling : capture_view [always] / filling__
+  state_full --> state_full : capture_view [always] / full__
+  state_uninitialized --> state_uninitialized : advance [always] / uninitialized__
+  state_uninitialized --> state_uninitialized : reset [always] / uninitialized__
+  state_uninitialized --> state_uninitialized : capture_view [always] / uninitialized__
+  state_errored --> state_errored : advance [always] / internal_error__
+  state_errored --> state_errored : reset [always] / internal_error__
+  state_errored --> state_errored : capture_view [always] / internal_error__
+  state_uninitialized --> state_errored : _ [always] / effect_unexpected_
+  state_empty --> state_errored : _ [always] / effect_unexpected_
+  state_filling --> state_errored : _ [always] / effect_unexpected_
+  state_full --> state_errored : _ [always] / effect_unexpected_
+  state_errored --> state_errored : _ [always] / effect_unexpected_

.planning/architecture/mermaid/speech_generator_moshi_personaplex_session.mmd

@@ -0,0 +1,72 @@
+stateDiagram-v2
+  direction TB
+  [*] --> state_uninitialized
+  state_uninitialized --> state_initialize_temporal_positions_result : initialize_run [guard_initialize_request_valid_] / effect_initialize_temporal_positions_
+  state_initialize_temporal_positions_result --> state_initialize_depformer_positions_result : completion_initialize_run_ [initialize_run__] / effect_initialize_depformer_positions_
+  state_initialize_temporal_positions_result --> state_failed : completion_initialize_run_ [initialize_run__] / memory_initialize_failed__
+  state_initialize_depformer_positions_result --> state_initialize_encoder_result : completion_initialize_run_ [initialize_run__] / effect_initialize_encoder_
+  state_initialize_depformer_positions_result --> state_failed : completion_initialize_run_ [initialize_run__] / memory_initialize_failed__
+  state_uninitialized --> state_failed : initialize_run [guard_initialize_request_invalid_] / effect_fail_initialize_invalid_
+  state_initialize_encoder_result --> state_initialize_decoder_result : completion_initialize_run_ [guard_encoder_initialize_succeeded_] / effect_initialize_decoder_
+  state_initialize_encoder_result --> state_failed : completion_initialize_run_ [guard_encoder_initialize_failed_] / codec_initialize_failed__
+  state_initialize_decoder_result --> state_initialize_executor_result : completion_initialize_run_ [initialize_run__] / effect_initialize_executor_
+  state_initialize_decoder_result --> state_failed : completion_initialize_run_ [initialize_run__] / codec_initialize_failed__
+  state_initialize_executor_result --> state_initialize_generator_result : completion_initialize_run_ [initialize_run__] / effect_initialize_generator_
+  state_initialize_executor_result --> state_failed : completion_initialize_run_ [initialize_run__] / executor_initialize_failed__
+  state_initialize_generator_result --> state_load_voice_result : completion_initialize_run_ [initialize_run__] / effect_load_voice_
+  state_initialize_generator_result --> state_failed : completion_initialize_run_ [initialize_run__] / generator_initialize_failed__
+  state_load_voice_result --> state_voice_prefill : completion_initialize_run_ [initialize_run__] / effect_publish_initialize_done_
+  state_load_voice_result --> state_failed : completion_initialize_run_ [initialize_run__] / voice_load_failed__
+  state_voice_prefill --> state_voice_prefill_result : advance_voice_run [always] / effect_prefill_voice_
+  state_voice_prefill_result --> state_voice_prefill : completion_advance_voice_run_ [advance_voice_run__] / effect_publish_advance_voice_done_
+  state_voice_prefill_result --> state_prompt_begin_result : completion_advance_voice_run_ [advance_voice_run__] / effect_begin_prompt_
+  state_voice_prefill_result --> state_failed : completion_advance_voice_run_ [advance_voice_run__] / voice_prefill_failed__
+  state_prompt_begin_result --> state_prompt_prefill : completion_advance_voice_run_ [advance_voice_run__] / effect_publish_advance_voice_done_
+  state_prompt_begin_result --> state_failed : completion_advance_voice_run_ [advance_voice_run__] / prompt_begin_failed__
+  state_prompt_prefill --> state_prompt_prefill_result : advance_prompt_run [always] / effect_prefill_prompt_
+  state_prompt_prefill_result --> state_prompt_prefill : completion_advance_prompt_run_ [advance_prompt_run__] / effect_publish_advance_prompt_done_
+  state_prompt_prefill_result --> state_live : completion_advance_prompt_run_ [advance_prompt_run__] / effect_publish_advance_prompt_done_
+  state_prompt_prefill_result --> state_failed : completion_advance_prompt_run_ [advance_prompt_run__] / prompt_prefill_failed__
+  state_live --> state_live_encode_result : live_frame_run [live_frame_run__] / live_frame_run__
+  state_live --> state_failed : live_frame_run [live_frame_run__] / invalid_request__
+  state_live_encode_result --> state_live_generate_result : completion_live_frame_run_ [live_frame_run__] / live_frame_run__
+  state_live_encode_result --> state_failed : completion_live_frame_run_ [live_frame_run__] / encode_failed__
+  state_live_generate_result --> state_live_decode_result : completion_live_frame_run_ [live_frame_run__] / live_frame_run__
+  state_live_generate_result --> state_live : completion_live_frame_run_ [live_frame_run__] / live_frame_run__
+  state_live_generate_result --> state_failed : completion_live_frame_run_ [live_frame_run__] / generate_failed__
+  state_live_decode_result --> state_live : completion_live_frame_run_ [live_frame_run__] / live_frame_run__
+  state_live_decode_result --> state_failed : completion_live_frame_run_ [live_frame_run__] / decode_failed__
+  state_live --> state_flush : begin_flush_run [always] / effect_publish_begin_flush_done_
+  state_flush --> state_flush_encode_result : flush_frame_run [flush_frame_run__] / flush_frame_run__
+  state_flush --> state_failed : flush_frame_run [flush_frame_run__] / invalid_request__
+  state_flush_encode_result --> state_flush_generate_result : completion_flush_frame_run_ [flush_frame_run__] / flush_frame_run__
+  state_flush_encode_result --> state_failed : completion_flush_frame_run_ [flush_frame_run__] / encode_failed__
+  state_flush_generate_result --> state_flush_decode_result : completion_flush_frame_run_ [flush_frame_run__] / flush_frame_run__
+  state_flush_generate_result --> state_flush : completion_flush_frame_run_ [flush_frame_run__] / flush_frame_run__
+  state_flush_generate_result --> state_failed : completion_flush_frame_run_ [flush_frame_run__] / generate_failed__
+  state_flush_decode_result --> state_flush : completion_flush_frame_run_ [flush_frame_run__] / flush_frame_run__
+  state_flush_decode_result --> state_failed : completion_flush_frame_run_ [flush_frame_run__] / decode_failed__
+  state_flush --> state_done : finish_run [always] / effect_publish_finish_done_
+  state_uninitialized --> state_failed : _ [always] / effect_mark_unexpected_
+  state_initialize_temporal_positions_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_initialize_depformer_positions_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_initialize_encoder_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_initialize_decoder_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_initialize_executor_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_initialize_generator_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_load_voice_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_voice_prefill --> state_failed : _ [always] / effect_mark_unexpected_
+  state_voice_prefill_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_prompt_begin_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_prompt_prefill --> state_failed : _ [always] / effect_mark_unexpected_
+  state_prompt_prefill_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_live --> state_failed : _ [always] / effect_mark_unexpected_
+  state_live_encode_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_live_generate_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_live_decode_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_flush --> state_failed : _ [always] / effect_mark_unexpected_
+  state_flush_encode_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_flush_generate_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_flush_decode_result --> state_failed : _ [always] / effect_mark_unexpected_
+  state_done --> state_failed : _ [always] / effect_mark_unexpected_
+  state_failed --> state_failed : _ [always] / effect_mark_unexpected_

.planning/architecture/mermaid/speech_tokenizer_moshi.mmd

@@ -0,0 +1,60 @@
+stateDiagram-v2
+  direction TB
+  [*] --> state_uninitialized
+  state_uninitialized --> state_ready : initialize [guard_configuration_valid_] / effect_initialize_
+  state_uninitialized --> state_uninitialized : initialize [guard_configuration_invalid_] / invalid_configuration__
+  state_ready --> state_ready : initialize [always] / already_initialized__
+  state_prepared_full --> state_prepared_full : initialize [always] / already_initialized__
+  state_prepared_generated --> state_prepared_generated : initialize [always] / already_initialized__
+  state_errored --> state_ready : initialize [guard_configuration_valid_] / effect_initialize_
+  state_errored --> state_errored : initialize [guard_configuration_invalid_] / invalid_configuration__
+  state_ready --> state_prepared_full : tokenize [guard_tokenize_full_] / effect_tokenize_full_
+  state_ready --> state_prepared_generated : tokenize [guard_tokenize_tail_] / effect_tokenize_tail_
+  state_ready --> state_prepared_generated : tokenize [guard_tokenize_empty_] / effect_tokenize_empty_
+  state_ready --> state_ready : tokenize [guard_tokenize_invalid_] / request_shape__
+  state_ready --> state_ready : tokenize [guard_tokenize_position_overflow_] / position_overflow__
+  state_prepared_full --> state_prepared_full : tokenize [always] / phase_order__
+  state_prepared_generated --> state_prepared_generated : tokenize [always] / phase_order__
+  state_prepared_full --> state_commit_full_zero : detokenize_run [guard_detokenize_valid_replace_] / effect_begin_detokenize_
+  state_prepared_full --> state_commit_full_generated : detokenize_run [guard_detokenize_valid_generated_] / effect_begin_detokenize_
+  state_prepared_generated --> state_commit_generated_zero : detokenize_run [guard_detokenize_valid_replace_] / effect_begin_detokenize_
+  state_prepared_generated --> state_commit_generated : detokenize_run [guard_detokenize_valid_generated_] / effect_begin_detokenize_
+  state_prepared_full --> state_prepared_full : detokenize_run [guard_detokenize_shape_invalid_] / request_shape__
+  state_prepared_generated --> state_prepared_generated : detokenize_run [guard_detokenize_shape_invalid_] / request_shape__
+  state_prepared_full --> state_prepared_full : detokenize_run [guard_position_overflow_] / position_overflow__
+  state_prepared_generated --> state_prepared_generated : detokenize_run [guard_position_overflow_] / position_overflow__
+  state_commit_full_zero --> state_output_decision : completion_detokenize_run_ [always] / zero__
+  state_commit_full_generated --> state_output_decision : completion_detokenize_run_ [always] / generated__
+  state_commit_generated_zero --> state_output_decision : completion_detokenize_run_ [always] / zero__
+  state_commit_generated --> state_output_decision : completion_detokenize_run_ [always] / generated__
+  state_output_decision --> state_ready : completion_detokenize_run_ [guard_before_output_delay_] / effect_publish_no_output_
+  state_output_decision --> state_output_validation : completion_detokenize_run_ [guard_past_output_delay_] / effect_collect_output_
+  state_output_validation --> state_ready : completion_detokenize_run_ [guard_output_complete_] / effect_publish_output_
+  state_output_validation --> state_ready : completion_detokenize_run_ [guard_output_incomplete_] / effect_publish_no_output_
+  state_ready --> state_ready : restore_cache [guard_restore_valid_] / effect_restore_column_major_cache_
+  state_ready --> state_ready : restore_cache [guard_restore_invalid_] / request_shape__
+  state_prepared_full --> state_prepared_full : restore_cache [always] / phase_order__
+  state_prepared_generated --> state_prepared_generated : restore_cache [always] / phase_order__
+  state_prepared_full --> state_ready : advance [guard_advance_position_available_] / effect_advance_
+  state_prepared_full --> state_prepared_full : advance [guard_advance_position_overflow_] / position_overflow__
+  state_prepared_generated --> state_prepared_generated : advance [always] / phase_order__
+  state_ready --> state_ready : advance [always] / phase_order__
+  state_ready --> state_ready : reset [always] / effect_reset_
+  state_prepared_full --> state_ready : reset [always] / effect_reset_
+  state_prepared_generated --> state_ready : reset [always] / effect_reset_
+  state_errored --> state_ready : reset [always] / effect_reset_
+  state_uninitialized --> state_uninitialized : tokenize [always] / uninitialized__
+  state_uninitialized --> state_uninitialized : detokenize_run [always] / uninitialized__
+  state_uninitialized --> state_uninitialized : restore_cache [always] / uninitialized__
+  state_uninitialized --> state_uninitialized : advance [always] / uninitialized__
+  state_uninitialized --> state_uninitialized : reset [always] / uninitialized__
+  state_ready --> state_ready : detokenize_run [always] / phase_order__
+  state_errored --> state_errored : tokenize [always] / internal_error__
+  state_errored --> state_errored : detokenize_run [always] / internal_error__
+  state_errored --> state_errored : restore_cache [always] / internal_error__
+  state_errored --> state_errored : advance [always] / internal_error__
+  state_uninitialized --> state_errored : _ [always] / effect_unexpected_
+  state_ready --> state_errored : _ [always] / effect_unexpected_
+  state_prepared_full --> state_errored : _ [always] / effect_unexpected_
+  state_prepared_generated --> state_errored : _ [always] / effect_unexpected_
+  state_errored --> state_errored : _ [always] / effect_unexpected_