[nicoloboschi]

Stacked on #5626 (the eval suite). Draft until the approach is agreed.

What

Prompt rules and server safety nets for consolidation, each one found by a failing scenario in #5626.

• Prompt rules:
  • recency follows mentioned_at;
  • a state change is an update, not a delete;
  • a change cascades to every observation it affects;
  • counts are kept as their own facet;
  • lookalike names stay separate;
  • chatter is skipped;
  • an update keeps the full text (current_text);
  • every durable fact lands somewhere.
• Server safety nets:
  • a delete's evidence is folded into the observation it merged into (merged_into), or re-queued;
  • an update aimed at a fact id becomes a create;
  • observation ids the model miscopied are repaired;
  • a lookalike-name guard;
  • facts the model skipped are asked again;
  • a cascade pass for observations the change left untouched;
  • a list of other observations about the same entities is added to the prompt.

It also puts retired-conventions back in the eval's must-pass set.

Results (Gemini 3.1 Flash Lite, 17 scenarios × 3)

	Passed	Gemini calls	Gemini cost
main	49/51	81	$0.136
this PR	51/51	82	$0.122

Tests

• Unit: test_consolidation_lookalike.py, plus updated delete tests.
• System: test_22_consolidation_safety_nets.py.