Purpose of the change

Prevent asyncpg.CharacterNotInRepertoireError crashes caused by null bytes (\x00) in LLM-generated strings inserted into PostgreSQL TEXT columns.

Description

Implements defense-in-depth sanitization at two layers:

1. Pydantic model layer (early catch): field_validator on SemanticCommand and LLMReducedFeature strips null bytes as soon as LLM output is parsed, before it reaches any storage code.
2. PostgreSQL storage layer (backstop): A shared sanitize_pg_text() utility is called in add_feature() and update_feature() to strip null bytes from tag, feature, and value fields before insertion — catching any path that bypasses models (e.g. REST API direct inserts).

The sanitize_pg_text utility logs a warning when it strips null bytes so data-quality issues are visible in monitoring.

Fixes/Closes

Fixes #1147

Type of change

• Bug fix (non-breaking change which fixes an issue)

How Has This Been Tested?

• Unit Test
• Integration Test

Tests added:

• TestSanitizePgText — 10 unit tests covering clean passthrough, null byte stripping (single, multiple, leading, trailing, all-null), warning logging, and no-log for clean strings.
• TestSemanticCommandNullByteStripping — 4 tests verifying the Pydantic validator strips null bytes from feature, tag, and value fields on SemanticCommand.
• TestLLMReducedFeatureNullByteStripping — 4 tests verifying the same for LLMReducedFeature.
• test_add_feature_with_null_bytes_does_not_crash — Regression test confirming null bytes in feature values do not crash storage insertion.

Test Results: 1329 passed, 9 skipped, 0 failures. Ruff lint and format clean.

Checklist

• My code follows the style guidelines of this project (See STYLE_GUIDE.md)
• I have performed a self-review of my own code
• I have commented my code
• My changes generate no new warnings
• I have added unit tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have checked my code and corrected any misspellings

Screenshots/Gifs

N/A

Further comments

• Not in scope: Fixing the LLM itself (upstream model behavior), per-feature error recovery in ingestion (separate follow-up), and Neo4j storage sanitization (handles null bytes differently).
• The sanitize_pg_text utility is intentionally kept separate from the Pydantic validators so it can serve as a backstop for all PostgreSQL insertion paths, not just model-validated ones.