Provenance and clean-room boundary

This project is implemented independently under Apache-2.0. The compatibility requirements were derived from public Piper configuration and behavior observations, the eSpeak NG public header/API, and the local 01_todo.md engineering guide.

Requirement	Public or black-box basis	Independent decision
Voice-specific ID maps	Piper voice configuration shape	Keep the loaded map authoritative and validate IDs locally.
Ordinary BOS/PAD/EOS framing	Observed frontend behavior	Implement a small codec strategy with no universal map.
NFD normalization	Existing project behavior and compatibility requirement	Use Python `unicodedata.normalize` at the frontend boundary.
Clause terminators	eSpeak NG public API and runtime contract	Use the `espeakng-runtime` public clause API and convert results at the Piper adapter boundary.
Native global-state locking	eSpeak public API global behavior	Delegate process-wide locking and lifetime to `espeakng-runtime`.
Raw blocks	Observed eSpeak/Piper behavior	Use a small state-machine parser and independent composition helper.
CLI fallback	Existing project behavior	Delegate CLI execution to `espeakng-runtime` while retaining Piper’s local splitter and composition.
Lexphon managed assets	Lexphon public data/profile APIs	Keep installation, verification, and normalization in the optional adapter.
G2Lex local assets	G2Lex public open/lookup APIs	Keep direct lookup narrow, ordered, and exact-key.

PiperG2P preserves phoneme_encoding, lexicon ID, data version, producer, transform, and generator identity from built-in assets. ipa is a generic override. espeak-ipa3 is a Piper raw pronunciation representation and is not normalized by Lexphon. | Lexicon overlay | User-provided implementation guide | Preserve Piper sentence/model encoding ownership in PiperG2P. | | KokoroG2P | Architectural reference only | Reuse lifecycle ideas without copying source or tests. |

eSpeak-derived test assets are generated during tests from declared source words and pronunciations. Production dictionaries are not bundled for test convenience. No Piper source, comments, tests, lookup tables, model files, bundled resources, or runtime import is used. The pinned upstream identity in the benchmark metadata is historical evidence, not code to port.