Bound and terminalize unresolved Realtime input transcription state

[andimarafioti]

Context

The per-item accumulator required by #474 is naturally released when a normal transcription completion arrives. Exceptional paths can still leave unresolved state behind: dropped terminals, direct-audio turns, handler failures, abandoned speculative revisions, or long-lived overlapping turns.

PR #476 experimented with completed tombstones, a 128-prefix cap, suppressing partials after eviction, and direct-audio cleanup. Those policies are broader than #474's wire-semantics fix and deserve an explicit lifecycle design.

Scope

• Enumerate every input-item terminal and abandonment path for STT and direct-audio pipelines.
• Define cleanup for dropped, failed, superseded, and disconnected turns.
• Decide whether completed tombstones are needed to reject duplicate late events.
• Bound unresolved transcript prefixes and routing records without silently corrupting incremental streams.
• Decide whether eviction should suppress later partials, emit a failure/abandonment signal, or use another recovery strategy.
• Add stress and late-event regressions.

Acceptance criteria

• Per-connection input-transcription state remains bounded on every pipeline path.
• Normal late authoritative completions retain correct item_id and audio duration.
• Eviction or abandonment cannot cause a cumulative hypothesis to be replayed as a wire delta.
• Direct-audio operation does not retain transcription-only state.
• Duplicate or late events have documented deterministic behavior.

Non-goals

• The basic incremental delta conversion and fixed content_index from Align Realtime input transcription delta semantics #474.
• Metadata-less terminal routing policy, tracked separately.
• Changing which STT backends emit progressive transcription.