fix: prevent retired human source state from history resurrection

[AllenXiao95]

Follow-up to #38 / PR #41.

Problem

retire_human_entries() deletes the current panel-authored human row, but an older tm_source_history row with the same (stable_identity, source_fingerprint) can still exist. After retirement, exact-history lookup can therefore resurrect the older machine/formal translation for the source state the operator explicitly retired.

Fix

• before deleting an origin='human' current row, capture its exact source fingerprint;
• in the same transaction, delete only the matching tm_source_history fallback for that coordinate + fingerprint;
• preserve history for sibling source fingerprints;
• leave normal stale-formal retirement/history archival unchanged.

Regression coverage

Covers A -> B -> human A -> explicit retire A:

• A does not fall back to the older historical machine translation;
• sibling B history remains reusable via coordinate_history.

Scope

No lineage model, no tombstone framework, no Adapter/TranslationUnit changes, and no change to ordinary stale-formal archival semantics.