Fono

Cloud STT now self-heals from one-off language misdetections, the LLM
cleanup stage stops occasionally replying with a question instead of
the cleaned text, and every release tag is gated on a real Groq
equivalence check across five languages.

Added

• Cloud equivalence gate at release time: a new cloud-equivalence
  job in .github/workflows/release.yml calls Groq's
  whisper-large-v3-turbo against the existing multilingual fixture
  set (en × 4, ro × 3, es × 1, fr × 1, zh × 1; ~110 audio-seconds
  total) and diffs the per-fixture verdicts against a committed
  baseline at docs/bench/baseline-cloud-groq.json. Blocks artefact
  production on failure. Auto-skipped when GROQ_API_KEY is unset
  (forks, bootstrap tags) or the tag carries the -no-cloud-gate
  suffix (operator escape hatch). Cost per release: < 0.5 % of
  Groq's free-tier daily cap. See ADR
  0021-cloud-equivalence-via-real-api.md
  and docs/dev/release-checklist.md.
• fono-bench equivalence --stt groq accepts cloud Groq as an STT
  backend. Reads GROQ_API_KEY from env; default model
  whisper-large-v3-turbo, overridable via --model. New
  --rate-limit-ms <ms> flag (default 250 ms for --stt groq, 0
  otherwise) paces requests under Groq's 30-req/min ceiling. HTTP
  429 is a hard fail with code 3 and an explanatory message; never
  retried.
• New docs/dev/release-checklist.md documenting the bootstrap
  command for the cloud-equivalence baseline, the regenerate
  conditions, and the -no-cloud-gate override.

Fixed

• LLM cleanup occasionally returned a clarification reply
  (“It seems like you're describing a situation, but the details are
  incomplete. Could you provide the full text you're referring to…”)
  instead of the cleaned transcript. Reproducible across every
  cleanup backend — Cerebras, Groq, OpenAI, OpenRouter, Ollama,
  Anthropic, and the local llama.cpp path — because the failure mode
  is a property of how chat-trained LLMs interpret a bare short
  utterance, not of any single provider. The fix is correspondingly
  universal: the default cleanup prompt was rewritten with hard
  “never ask for clarification” rules; every backend now wraps the
  user message in unambiguous <<< / >>> delimiters so the
  transcript cannot be mistaken for a chat message; and a refusal
  detector rejects clarification-shaped replies and falls back to the
  raw STT text. Applied identically to OpenAiCompat, AnthropicLlm,
  and LlamaLocal. See
  plans/2026-04-28-llm-cleanup-clarification-refusal-fix-v1.md.

Changed

• [llm].skip_if_words_lt default raised from 0 to 3. One- and
  two-word captures (“yes”, “okay”, “send it”) now bypass the LLM
  cleanup roundtrip entirely — regardless of whether the configured
  backend is cloud or local — saving 150–800 ms and avoiding the
  short-utterance clarification failure mode at the source. Override
  in config.toml if you want every utterance cleaned.

• [stt.cloud].cloud_rerun_on_language_mismatch default flipped from
  false to true. Combined with the new in-memory language cache,
  cloud STT now self-heals from one-off language misdetections (e.g.
  Groq Turbo flagging accented English as Russian) at the cost of one
  extra round-trip per misfire. Set false to opt out.

Added

• In-memory per-backend language cache
  (crates/fono-stt/src/lang_cache.rs). Records the most recently
  correctly-detected language code per cloud STT backend; consulted
  only as a rerun target when post-validation fires. No file I/O,
  no persistence — daemon restarts rebuild within one or two
  utterances. OS locale (LANG / LC_ALL) seeds the cache at start
  if and only if its alpha-2 code is in general.languages.
• New crates/fono-core/src/locale.rs — POSIX-locale → BCP-47 alpha-2
  parser; used by both the cache bootstrap and the wizard.
• Tray Languages submenu (Linux): read-only checkbox display of
  the configured peer set plus a "Clear language memory" item that
  drops every entry from the in-memory cache.
• New ADR
  docs/decisions/0017-cloud-stt-language-stickiness.md
  documenting why the cache is rerun-only, in-memory only, and
  peer-symmetric (no primary/secondary).

Deprecated

• [stt.cloud].cloud_force_primary_language — superseded by the
  in-memory language cache. Field still parses for one release; will
  be removed in v0.5.
• LanguageSelection::primary() — renamed to fallback_hint(). The
  alias is retained as #[deprecated] for one release; usage is
  scope-restricted in its doc-comment to single-language transports.

See plans/2026-04-28-multi-language-stt-no-primary-v3.md.