Howzer v2.5.0-DE: delivery-ready hardening

What's in v2.5.0-DE

v2.5.0-DE consolidates the hardening work since v2.4 into a delivery-ready release. The focus is not new analysis features but reliability: the pipeline returns correct and honest results under load, generated customer replies stay faithful to the source, first-time installation runs through without manual rework, and the automation workflows are fully wired.

Correct analysis

• Text normalization fixed so German umlauts, special characters, and channel labels are normalized consistently before any of the 12 analyzers run.
• Channel normalization (email, phone, app, web) reliably maps incoming feedback to the right intake channel.

Honest results under load

Large batches of several hundred feedbacks no longer block each other: language-model I/O is decoupled from the CPU-bound analysis. When an analyzer fails because of overload or a missing model, the pipeline reports that transparently instead of passing off a placeholder or constant value as a real result. A degraded component is marked as such, so downstream reports can rely on the output.

Reliable customer replies

• Entity guard: generated replies no longer invent names, case numbers, or facts that are not in the original feedback.
• Empathy in the reply scales with the severity of the case — urgent situations get the appropriate tone.
• Replies are grounded in the actual feedback; the response engine's examples were cleaned of test data.
• Honest fallback: when no dependable reply can be produced, the engine returns a clearly recognizable safe fallback instead of a fabricated answer.

Escalation and data safety

• Consistent escalation logic across all analyzers — a feedback flagged as needing escalation is handled that way end to end.
• Fail-closed persistence: if the database is unreachable at startup, the pipeline refuses to serve with a clear error instead of silently running without storage. A database outage during operation keeps requests available and is surfaced on the health endpoint.
• Configurable language-model timeout so a stalled model cannot hold up processing indefinitely.

Self-hosted operation and installation

• Air-gap capable: the model cache can be exported as a bundle and imported on the target system; in offline mode the pipeline contacts no external model hubs, and a missing model yields an actionable error instead of a crash.
• Hardened secret handling: a fresh install generates strong random secrets automatically and preserves existing secrets on re-run; placeholder values are rejected. The salt for the GDPR deletion routine is generated automatically and never rotated.
• The GDPR retention and deletion routine stays fail-closed — without a configured salt it deliberately does not run.

Automation

All workflow templates are published and ready to activate, with alert delivery over Slack and email (SMTP) wired in. Re-importing workflow templates is idempotent — no more duplicated workflows.