feat(regen): add diarize_info, Flux force-end-turn, update-listen, word speaker confidence

[GregHolmes]

SDK regeneration for 2026-07-31 against fern-java-sdk 4.16.0, generated from the deepgram-docs sdk-gen-2026-07-31 branch (coordinate fix deepgram/deepgram-docs#1069 + Flux non-breaking reshape deepgram/deepgram-docs#1064 + cherry-picked spec PRs deepgram/deepgram-docs#1066, deepgram/deepgram-docs#1063, deepgram/deepgram-docs#1036, deepgram/deepgram-docs#1047).

Source-compatible with 0.7.x — no breaking changes

Verified by a full breaking-change sweep against 0.7.1 (public API diffed from compiled trees): zero deleted public types, zero renames, zero removed or retyped getters — 1414 public types and 10,999 public members all still present. The generator (4.16.0) introduced several source-breaking changes that are shimmed back below.

Two caveats on "purely additive", for the record:

• The new provider() overloads on AgentV1UpdateListenListen make a bare provider(null) ambiguous, so that one call no longer compiles. No working 0.7.x program is affected — it already threw NullPointerException from Objects.requireNonNull.
• ProviderStage gained two abstract methods, which technically breaks anyone implementing that generated builder-stage interface directly.

New functionality (additive)

• diarize_info in listen v1 response metadata (deepgram-docs#1036)
• ListenV2ForceEndTurn + sendForceEndTurn(...) on the listen v2 WS client (deepgram-docs#1047)
• UpdateListen can now change model/language — exposed additively (see shim below) (deepgram-docs#1066)
• speaker / speaker_confidence on pre-recorded words (deepgram-docs#1063)
• ListenV2Redact, retry-backoff tuning, and transparent gzip/deflate (ResponseDecompressionInterceptor)

Back-compat handling for generator-introduced breaks (STOPGAP)

fern-java-sdk 4.16.0 renamed/reshaped public types 0.7.x shipped. These are patched back and frozen; drop once fixed upstream (Fern report pending):

• Type renames (4.16.0 collision-based name shortening): DeepgramSpeakProviderModel→DeepgramModel, AnthropicThinkProviderModel→AnthropicModel, GoogleThinkProviderModel→GoogleModel, Cartesia*, Groq* — renamed back. (OpenAi* was unaffected — it's the only think+speak provider, so its names never collided.)
• AgentV1UpdateListenListen.getProvider() widened DeepgramListenProviderV2 → a V1|V2 union. Additive shim: getProvider() still returns DeepgramListenProviderV2 and provider(DeepgramListenProviderV2) still builds; the union is exposed via getProviderValue() and provider(DeepgramListenProviderV1) for the new model/language capability. Regression-tested by AgentV1UpdateListenShimTest.
• Agent conversation history (ConversationHistoryMessage, FunctionCallHistoryMessage) reparented onto history-specific types with the shared ContextMessagesItem* types deleted — restored to 0.7.x shapes, deleted types re-added (return-type changes can't be shimmed additively in Java).

Patch reconciliation

• ClientOptions retired from .fernignore — the runtime-version flag is on docs main, and the Maven-coordinate fix is deepgram/deepgram-docs#1069, which is still open. This regen was generated from a branch that includes it, so the headers in this PR are correct, but the unfreeze depends on #1069 landing: #1069 must merge before 0.7.2 is cut, or a later regen from main would restamp the wrong X-Fern-SDK-Name (com.deepgram:deepgram-sdk) with no freeze left to catch it. Added maven-jar-plugin manifest stamping to pom.xml so the published JAR carries Implementation-Version from the release-please-managed <version>.
• Re-applied clean: ReconnectingWebSocketListener, 11 fields-less hashCode() types, listen/speak WS query-param + additionalProperties + forward-compat patches. listen v2 merge preserves the new redact + sendForceEndTurn.

Verification

• Unit tests (./gradlew unitTest) — pass, including the new AgentV1UpdateListenShimTest (shim API surface, outbound JSON shape, V1/V2 round-trip) alongside the existing patch guards (RegenTypesTest hashCode, QueryStringMapperTest, connect wire tests, ListenV2/SpeakV2ForwardCompatTest, ReconnectingWebSocketListenerTest, RetryInterceptorTest).
• Integration tests (./gradlew integrationTest, live API) — pass (transcribe URL/file, speak REST, read analyze, manage projects; invalid-key error handling).
• Examples — all compile (compileExamples); the safe / changed-area set runs green against the live API.
• spotlessCheck clean; CI green.

Review follow-ups (added after the initial regen)

• Version-less agent listen providers were being dropped. version is optional on both provider variants, so {"provider":{"type":"deepgram","model":"nova-3"}} is valid and is what 0.7.x emitted — but Fern's defaultImpl = _UnknownValue sent it to an unknown variant whose value was never populated, so the getter returned null and re-serializing emitted {"provider":null}, silently dropping the provider. Now defaultImpl = V2Value on all three agent listen-provider unions. AgentV1UpdateListenListenProvider is new here so that was a regression; the two Settings unions shipped the same defect in 0.7.1 and are fixed too, so all three agree. Guarded by AgentV1UpdateListenShimTest and the new AgentSettingsProviderDefaultTest.
• ListenV2ForceEndTurn was missing hashCode() — the only one of twelve fields-less types not carrying the manual patch. Added, frozen, and asserted in RegenTypesTest.
• The new retry-tuning options were unreachable from DeepgramClient.builder(). initialRetryDelayMillis/maxRetryDelayMillis/retryJitterFactor landed on the parent builder only, so using one mid-chain downgraded the static type and build() returned a DeepgramApiClient. Overridden in both hand-maintained wrappers, matching the existing apiKey/timeout/maxRetries pattern.
• Freezes added for every newly hand-patched file — the two Settings unions were unfrozen, so the next regen would have reverted them silently.
• release-please: dropped the now-dead ClientOptions.java entry (the file has no x-release-please-version markers left, so the updater silently no-opped), and added jar-manifest stamping to build.gradle so Package.getImplementationVersion() resolves on the Gradle path too, not just the published Maven artifact.
• README: documented the backoff knobs and the per-request RequestOptions.maxRetries(...) override. All snippets were compiled against the built SDK — which is how the builder gap above was found.

Gate after these changes: ./gradlew clean unitTest spotlessCheck compileExamples — 104 tests, 0 failures. Each new guard was checked by reverting its fix and confirming the test fails.

[dg-coreylweathers]

Approved — clean pre-regen prep. Verified:

• All 17 .bak snapshots are byte-identical to the patched originals on main (hash-compared each pair), so the reconciliation source-of-truth is faithful and complete.
• Unfreeze is complete: no original (non-.bak) path lingers in .fernignore that would silently keep a file frozen.
• Blast radius is exactly .fernignore + 17 .bak files — no original .java modified/deleted; patches (no_delay escape-hatch + arraysAsRepeats multi-value serialization) still intact on the originals.
• Permanent freezes untouched; functionally inert (.bak isn't compiled, no withSourcesJar() so nothing leaks to a published artifact).

One optional nit: the .bak files sit under src/main/java/ with no .gitignore coverage — carry a cleanup item into the reconciliation PR to delete them and restore the 17 .fernignore entries to their .java paths.

The real risk lives in the post-regen reconciliation, not here. When Fern overwrites the now-unfrozen originals it drops every manual patch; the .bak snapshots are the only safety net. The deepgram-sdk-review-regen pass must confirm all four patch families survive — especially the no_delay escape-hatch + multi-value WS serialization, plus the Flux oneOf non-breaking reshape (#1064) that the regen source carries. Leaving merge to Greg.

[dg-coreylweathers]

Approved — full Fern 4.16.0 regeneration review (this supersedes the earlier prep-PR approval that was auto-dismissed when the head moved).

Independently verified (build gate run in Docker, not trusting CI):

• ./gradlew clean unitTest spotlessCheck compileExamples -> BUILD SUCCESSFUL, 104 tests, 0 failures; all new guard tests ran (shim, provider-default, regen-types hashCode, additionalProperties wire, forward-compat).
• Source-compat with 0.7.x: 0 removed/renamed public-type files vs v0.7.1; compileExamples (real usage code) green corroborates the surface.
• All 3 documented shims present and correct: type-rename reversal (short-named orphans not in tree), additive UpdateListen union shim, conversation-history reparent restored.
• Version-less provider drop fixed (defaultImpl=V2Value on all 3 agent listen-provider unions).
• hashCode on all 12 fields-less types incl. ListenV2ForceEndTurn; retry-tuning options reachable from both wrapper builders.
• no_delay escape hatch survived: wire test asserts it lands in the real upgrade URL on all 4 WS clients.
• New freezes complete + documented; .bak prep leftovers cleaned up; ClientOptions headers correct.

Two conditions, both about the RELEASE not the merge (details in Slack):

1. Merge preserving the fix: commits (not squash-as-chore, or release-please cuts nothing) + add a 0.7.2 changelog stanza for the new features (diarize_info, speaker_confidence, force-end-turn, redact, gzip) so the release isn't undersold.
2. Land docs #1069 (one line) so the now-unfrozen ClientOptions doesn't restamp the wrong Maven coordinate on the next regen. Order vs this PR doesn't matter for today's artifact; #1069 just must land before the next regen-from-main.

Cutting 0.7.2 today is safe: release CI does mvn deploy from committed source (no regen), and the correct headers are already committed here.