feat: raise the sync and dictation config caps to match the service

[aurpsis-aai]

Companion to DeepLearning #20095 and docs #349. Gated on the DL PR merging — until it ships these caps would let through requests the service still rejects.

Why

The SDK mirrors the service's caps client-side so a too-long prompt fails locally instead of round-tripping. That mirroring cuts both ways: once the service raises its caps, the SDK is the thing rejecting valid requests.

What changed

Field	Was	Now
prompt (Sync) / stt_prompt (Dictation)	4096 chars	6000 chars
keyterms_prompt	2048 chars	8000 chars + 100-term cap
conversation_context	100 turns / 4096 chars	500 turns / 16000 chars

Both surfaces, not just dictation. SyncTranscriptionConfig in types.py mirrors the same caps and needed the same bump — worth noting since the follow-up was originally scoped as a dictation-only change.

The keyterms count cap is new on both sides. Term count, not total characters, is what the server's prompt-budget trim walks, and neither config had a count limit. 100 terms, matching the service and streaming.

Each cap lives in one constant per surface (_SYNC_MAX_*, _DICTATION_MAX_*), and the validators, field docstrings and README comments all follow from it.

Tests

pytest tests/unit — 650 pass.

Four new tests (over the count cap raises, exactly at it is accepted — for both Sync and Dictation).

Five existing tests restated the old caps as literals, so raising the numbers would have left them passing while testing nothing: a 3000-character keyterms list is now valid, 120 context turns no longer trims, a 5000-character stt_prompt is now under the cap. They read the constants now, so the next cap change can't quietly turn them into no-ops.

ruff check and ruff format --check --config ./ruff.toml clean on the changed files. (ruff check flags 8 other test files for unsorted imports on master; CI only lints a PR's changed files, so those are untouched here.)

Version

Patch bump to 1.5.2.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VK9A9Dj5YKn2dGJ1XDry9n