Task: Round trip the conventions configuration

Every conventions document in the corpus imports, exports and matches, with the two normalisations the mapper performs on purpose stated in the kind's comparison rather than hidden.

• The mapper reads and writes every field the modelled categories carry, so no element is dropped.
• Each convention category the corpus uses has an entity in ores.refdata and a mapper in both directions.
• The conventions kind is registered with the round-trip harness and all seventy-two corpus files round trip.
• The mapper reports any category it does not model, and a test asserts that it reports none.

1. Make the skip countable before changing anything. The mapper dropped seventeen categories in silence, which is why seventy-two files went unmeasured: nothing could say which categories the documents use. It now fills mapped_conventions::unmodelled and warns per category.
2. Measure which categories the corpus needs, and which files use only modelled categories. The probes are hidden tests, because the kind does not round trip and a test that passes while a document loses a category is worse than a measurement a person reads.
3. Fix the modelled categories first, because they are the cheap half and they are why nine files fail for a reason unrelated to the unmodelled categories.
4. Then model the sixteen categories, most used first, each with a refdata entity, both mappers, and its own round trip.
5. State the two normalisations the mapper performs on purpose in the kind's comparison, and only once the categories are modelled, because until then the kind cannot pass whatever the comparison says.

• The nine files that carry only modelled categories round trip, and are asserted green. The rest cannot, because a category the mapper does not model is content the export cannot write.
• The failure on those nine was a value spelling, not a lost element. A boolean comes back as True for true, and a day counter as A365F for A365, because the mapper collapses ORE's boolean spellings and enum aliases to the canonical codes the refdata columns hold. That is deliberate and carries the same information, so the kind's comparison states it. An earlier note here claimed <IndexBased> was dropped; dumping the export showed it present and respelled, and that note is corrected.
• The comparison is sound because it checks two things: no element appears fewer times in the export than in the document, and the same mapper reads the same conventions out of both. The first catches a category the mapper reads and does not model; the second tolerates the canonicalisation and still catches a value the mapper got wrong, because such a value survives neither direction.
• The mapper warns once per unmodelled category on every import, which will be noisy for the corpus until the categories land. That noise is the point: the skip was invisible for as long as it was silent.

Manual QA scenarios (scaffolded via compass add test_scenario) that verify this task. Link new ones here as they're created; the scenario doc itself links back via its "Verifies task" field.

Delivered in PR #2354, which closes this task with the modelled half of conventions proven and the unmodelled half scoped. The task's third acceptance bullet, that all seventy-two files round trip, is not met; the sixteen categories that stand between the corpus and it are carried by Model the convention categories the mapper skips, which continues this work.

What the mapper did. It ignored seventeen of the document's twenty-six categories without saying so, which is why seventy-two files passed a round-trip test that only compared element counts. It now counts what it skips into mapped_conventions::unmodelled and warns per category.

What now round trips. The nine files that carry only modelled categories do. They failed on a value spelling rather than a lost element: true came back as True and A365 as A365F, because the mapper collapses ORE's boolean spellings and enum aliases to the canonical codes the refdata columns hold. The kind's comparison answers it in two parts, and it is sound because the first part refuses a lost element and the second cannot hide one.

What does not. The other sixty-three files each carry at least one of the sixteen categories the mapper does not model, from CrossCurrencyBasis in forty-nine files to IntradayPowerLoad in one. A category the mapper cannot write is content the export cannot hold, so those files stay measured by a hidden probe rather than asserted.

A correction this task had to make. An earlier note claimed the export dropped <IndexBased>. It does not: the comparison quotes a window starting forty bytes before the difference, so the element at the top of the imported side was context rather than the difference. Dumping the export settled it, and the fix the claim implied, a mapper line and a new column, was not needed.