fix(report): count distinct tests, and assert the catalog is well-formed

[BitHighlander]

Two defects a human auditor finds before anyone else does.

1. The header over-counted. total summed catalog rows, and two tests are deliberately catalogued twice — test_eip1559_requires_chain_id is the replayable-signature refusal in the 7.14.2 defect narrative (J9) and a guard in the EVM catalog (VG2); test_contract_handler_streamed_calldata_signs_full_data is J8 and VG6. Both entries earn their place: the same test carries two different arguments.

But summing rows claimed 374 tests where the run contains 372, so anyone reconciling the header against the JUnit finds a two-test shortfall that is pure double-counting — and a report whose own arithmetic doesn't survive a reconcile isn't evidence, whatever the tests did. Now counts distinct (module, method) and keeps both rows.

2. VG4 had no context. It rendered as a bare test name with no statement of what it proves — precisely the row an auditor cannot evaluate. Filled in: the 0x02 envelope prefix comes from msg.type but the fee fields from has_max_fee_per_gas, so a type-2 tx carrying only gas_price would hash a legacy fee into a 1559 field list. Refused, because a signature over a malformed field list is still a valid signature over something.

Then the check that finds the next one, run on every render: unique section letters, unique test ids, no entry missing a title or context. Fifteen lines, no new flag, no CI wiring — it runs because the report runs.