feat(regen)!: flux stt numerals + aura-2 voices; remove AgentV1LatencyReport.stt_latency

[GregHolmes]

SDK regeneration for the 2026-07-20 Fern generator run (java-sdk generator 4.10.1, CLI 5.44.6, spec ff8fd2b7). Targets a 0.7.0 minor release.

⚠ BREAKING CHANGE

• AgentV1LatencyReport.getSttLatency() removed (plus the sttLatency(...) builder methods and constructor arg). The API spec removed stt_latency from the LatencyReport schema (deepgram-docs #1006). AgentV1LatencyReport is a server-emitted, read-only message, so there is no request/wire impact — but this is source-breaking for any caller of getSttLatency(). Inbound stt_latency (if a server ever sends it) now lands in additionalProperties rather than a typed getter.
  • Java allows breaking changes across minor bumps in the 0.x line (precedent: feat!: (regen) breaking flux timestamps req-double + close-stream/hints; diarize_model, profanity #66 → 0.6.0), so — unlike the Python SDK, which re-added the field as a read-side compat shim — we accept the removal here. Released as a minor bump via the ! in the commit type.

Features (additive)

• Flux STT numerals — new types/ListenV2Numerals.java (true/false, forward-compatible enum) plus a numerals query param on listen().v2().v2WebSocket().connect(...) via the new V2ConnectOptions. Connection-time only.
  • Wire coverage added: src/test/java/com/deepgram/ListenV2ConnectWireTest.java pins that /v2/listen serializes numerals=true when set and omits it when absent (the handshake previously had no wire test). Demonstrated in examples/listen/LiveStreamingV2.java.
• New Aura-2 multilingual TTS voices — SpeakV1Model and AudioGenerateRequestModel each grow 63 → 103 constants (es/de/nl/fr/it/ja). Additive only; zero removed constants/enum members.

Documentation

• Added docs/Migrating-v0.6-to-v0.7.md — migration guide covering the stt_latency removal (with before/after) plus the additive numerals + Aura-2 voice features.
• Linked it from the README migration list as current (and backfilled the previously-unlisted v0.5→v0.6 entry).

Manual patches reconciled

All 15 previously-frozen patches were still needed and re-applied byte-identical to main (the behavioral files do not appear in the net main...HEAD diff): ClientOptions (SDK identity + release-please markers, held at 0.6.1 for release-please to bump), ReconnectingWebSocketListener, both V2WebSocketClient forward-compat no-ops, and 11 fields-less hashCode() fixes. The listen V2WebSocketClient re-apply preserved the generator’s new numerals block. All .bak files deleted; AgentV1LatencyReport left unfrozen (generator-owned).

Generator input

• originGitCommit: ff8fd2b74fdd5c081e9f111d59d3619f7e286dd8, originGitCommitIsDirty: true — this regen was cut from an uncommitted API-definition state. Verified: the stt_latency removal is an intentional schema change (deepgram-docs #1006, merged 2026-07-14) and is an ancestor of ff8fd2b7, so it is in the regen’s committed spec base — the dirty flag reflects other uncommitted working-tree changes, not this removal. Note any spec change after this commit forces a re-regen.

Verification

• ./gradlew test compileExamples — BUILD SUCCESSFUL: 81 tests, 0 failures, 0 errors, 1 skipped (incl. forward-compat, reconnect, hashCode, and new numerals wire guards); examples compile.
• Numerals verified end-to-end against the live API (throwaway harness): the same TTS-synthesized audio through Flux rendered digits with numerals=true (23, 5 5 5 1 2 3 4) and spelled-out words without it.

[dg-coreylweathers]

Verified: build gate (81 tests, 0 failures) + live numerals behavior (23 / 5 5 5 1 2 3 4 with numerals=true, spelled-out without). Approving; S1 (metadata.json 0.6.2) and the Java/Python changelog divergence are non-blocking.