fix: send list-of-primitive multipart fields as repeated form fields (#819)

[kraenhansen]

Summary

Interim fix for #819. speech_to_text.convert(keyterms=[...]) (and other multipart list fields) fails with invalid_keyword_length because the whole list is sent as one JSON-array string instead of repeated form fields.

Root cause (verified): generator fern-python-sdk 4.46.9 (commit cb4feb824, PR fern-api/fern#11284, 2026-01-08) added an _is_primitive_type() gate to multipart serialization: only scalar primitives are sent directly, everything else — including List[str] — is json.dumps(jsonable_encoder(...))'d. PR #817 bumped us onto that behavior (4.42.0 → 4.64.1). The change was undocumented — it shipped bundled inside a "fix python wire tests" PR. Fixed upstream in generator 5.22.2, but that's past the 5.5.4 min-Python-3.10 clamp, so we can't just upgrade forward.

⚠️ This is a temporary patch — when to revert

This hand-edits Fern-generated files, so it must be retired once the generator itself emits the correct code. Revert (via fern replay forget <patch-id>, then regenerate) as soon as either of these is true:

• The generator is upgraded to ≥ 5.22.2 — the version that fixes list-of-primitive multipart serialization (fix-multipart-list-of-primitives-repeated-fields.yml). ⚠️ Note: ≥ 5.5.4 raises the generated min-Python to 3.10 — only take this path once we're ready for that breaking change.
• A 4.x backport of the 5.22.2 fix lands (requested from Fern) and we pin to it — keeps Python 3.8 support and fixes the bug, at which point this patch is fully redundant.

Until then, Fern Replay re-applies this patch on every regeneration. Leaving it in place after the generator is fixed will cause a 3-way-merge conflict (both sides changing the same lines) surfaced in the regeneration PR body — that conflict is the signal to run fern replay forget and drop this patch.

Change

Reverts the 8 list-of-primitive (List[str]) multipart fields to raw values so httpx emits repeated form fields, matching generator 5.22.2 semantics:

File	Field(s)
speech_to_text/raw_client.py	keyterms
dubbing/project/raw_client.py	keyterms
music/raw_client.py	tags
music/finetunes/raw_client.py	tags
studio/projects/raw_client.py	genres, pronunciation_dictionary_locators, voice_settings
audio_native/raw_client.py	pronunciation_dictionary_locators

Object fields (labels, entity_detection, entity_redaction, webhook_metadata) keep json.dumps — correct for multipart. Now-unused json / jsonable_encoder imports removed.

Durability

These are generated files, so the edit is carried across regenerations by Fern Replay (auto-captured as a patch in .fern/replay.lock). See the revert conditions above for how it is retired.

Test plan

• ruff@0.11.5 (E/F/I) passes on all 6 files
• py_compile passes on all 6 files
• Verify speech_to_text.convert(keyterms=[...]) no longer returns invalid_keyword_length against the live API
• Confirm Fern Replay captures the patch on the next regeneration (patches detected: 1)

Fixes #819

🤖 Generated with Claude Code