refactor(transcriber): move prerecorded transcription into prerecorded/v2

[ccampbell-aai]

Summary

Continues the SDK rearchitecture (following sync/v1, #214/#216, and streaming/v3): the async (prerecorded) product moves into a versioned subpackage, assemblyai/prerecorded/v2/ (v2 = the /v2/transcript API).

• transcriber.py → prerecorded/v2/{transcript,transcript_group,client}.py — one public class per module beside its private impl (Transcript/_TranscriptImpl, TranscriptGroup/_TranscriptGroupImpl, Transcriber/_TranscriberImpl), layered transcript ← transcript_group ← client with no circular imports. Class bodies are byte-identical to the originals; only module headers/imports changed.
• Transcript-only pieces of api.py (ENDPOINT_TRANSCRIPT + 10 functions) move to prerecorded/v2/api.py. Shared pieces (upload_file, ENDPOINT_UPLOAD, _get_error_message) and the LeMUR endpoints stay at root — LeMUR gets its own subpackage in a later phase.
• prerecorded/v2/__init__.py is the public facade (Transcriber, Transcript, TranscriptGroup, TranscriptionConfig).

Backwards compatibility

No behavior change; every existing import path resolves to the identical objects:

• flat assemblyai/transcriber.py is now a silent re-export shim (same idiom as assemblyai/sync/__init__.py), including the private _*Impl names;
• root api.py re-exports the moved names, so from assemblyai.api import ENDPOINT_TRANSCRIPT, ... keeps working;
• root assemblyai/__init__.py is untouched;
• new tests/unit/test_transcriber_backwards_compat.py pins old-path/new-path identity, the private surface, and warning-free imports.

Testing

• Full suite: 391 passed (385 at base + 6 new backwards-compat tests)
• Smoke-tested imports under warnings.simplefilter('error') from all entry orders (assemblyai first, assemblyai.prerecorded.v2.api first, assemblyai.api first)
• ruff check + format clean on all touched files

🤖 Generated with Claude Code

[ccampbell-aai]

Live backwards-compatibility verification

In addition to the in-repo compat tests, ran a live end-to-end comparison against the production API: the same set of prerecorded calls executed twice — once with the published PyPI package (0.64.32), once with this branch — using only pre-refactor import paths (import assemblyai as aai, assemblyai.transcriber, assemblyai.api), with results deep-compared as normalized JSON. A single shared transcript was created up front so read-path outputs could be compared byte-for-byte.

All ten relocated api.py functions were exercised:

check	covers	result
transcribe_e2e	upload_file + create_transcript + get_transcript polling	✅ identical
get_by_id	get_transcript (full word list incl. confidences)	✅ identical
word_search	word_search	✅ identical
subtitles	export_subtitles_srt / export_subtitles_vtt	✅ identical
sentences_paragraphs	get_sentences / get_paragraphs	✅ identical
list_transcripts	list_transcripts (before_id-pinned page)	✅ identical
transcript_group	TranscriptGroup.get_by_ids	✅ identical
redacted_audio (+ client-side guard)	get_redacted_audio via a redact_pii transcription	✅ identical (stable parts of signed URL)
delete_own	delete_transcript	✅ identical

10/10 identical. The only differences ever observed were server-side nondeterminism, reproduced within a single SDK version: per-word confidence drift between separate transcriptions of the same audio, and unstable ordering of word_search matches (the comparison pins confidence to the shared transcript and sorts matches by term). Module metadata (aai.Transcriber.__module__: assemblyai.transcriber → assemblyai.prerecorded.v2.client) differs by design and was excluded.

🤖 Generated with Claude Code