narratio/docs/releases/v2.2.0.md

Narratio v2.2.0

This release adds resumable per-speaker transcription, replayable WhisperX uploads, and explicit Notarius configuration profile selection.

Summary

Successfully checkpointed speaker transcripts survive later transcription failures. Ordinary retry reuses valid completed speakers and processes only the remaining work. WhisperX uploads can now be replayed by the HTTP transport without buffering whole audio files.

Compatibility

Existing CLI commands, speaker identities, transcript formats, and pipeline ordering are unchanged. Session manifests gain optional per-speaker completion evidence. Older manifests remain readable, but older Narratio binaries may reject manifests containing the new fields; downgrading after transcription is not transparent.

The optional pipeline.notarius.config_profile field selects a downstream Notarius configuration profile independently of Narratio's own profile. Using it requires a Notarius version supporting --config-profile; v0.7.0 is a known supporting version. Omitting the field retains downstream default selection.

Upgrade

No configuration migration is required. When transcribe is next selected for an older aggregate-only manifest, all current speakers are transcribed once to establish reusable evidence. Previously unrecorded results from failed runs are not automatically imported.

Use ordinary run/resume to retain valid checkpoints. Forcing transcribe resets reuse of prior results; an ordinary retry after a failed forced invocation can retain replacements already checkpointed by that attempt. See the operations guide and Notarius profile configuration.

Changes

• Persist validated speaker completions independently of aggregate stage status, with serialized saves for concurrent work and input/output integrity checks.
• Rebuild aggregate transcript references when speakers are removed or those references need repair, without retranscribing otherwise-current speakers.
• Keep removed and incomplete speaker results out of versioned merge inputs.
• Permit checkpoint evidence for zero-byte audio when WhisperX accepts it; service rejection still fails normally.
• Provide fresh streaming multipart bodies for HTTP transport replay while preserving application retries, cancellation, and resource cleanup. This addresses the missing-body-replay limitation, not persistent upstream protocol failures or exactly-once service execution.
• Pass explicit Notarius configuration profiles through the CLI and include their selection in extraction reuse decisions.