Task: B4: Collapse refdata and refdata-cpp into a single component

A single refdata component entry in the component catalogue discovers all ores.refdata.*.org entity files (excluding junctions and module index) and dispatches SQL, C++, messaging, and Qt profiles from one component. The refdata-cpp component entry is removed.

• compass codegen entity –component refdata generates SQL, C++, and Qt output identical to running –component refdata and –component refdata-cpp separately before this change
• Dead models_dir entries are removed from the catalogue
• No other component references refdata-cpp
• All existing codegen regression tests pass with zero diff

Investigation confirmed the refdata / refdata-cpp split in component_catalogue.org is vestigial: both rows already point at the same modeling_dir (projects/ores.refdata/modeling), and their models_dir (legacy JSON) no longer exists on disk for either. manifest.py's org discovery doesn't split by component at all — it picks up every codegen-typed org file in modeling_dir regardless of which row is used — so --component refdata and --component refdata-cpp already discover an identical file set today. --address, not --component, selects the technical space (SQL/C++/Qt/…); the component only scopes which entities are processed.

1. Add an exclude filter to manifest.py's org-discovery path so the merged refdata component skips ores.codegen.module and ores.codegen.junction typed files (junction bulk regeneration currently produces spurious errors on some profiles; module is a docs-only index, never a codegen target).
2. Collapse the two component_catalogue.org rows into one refdata row; drop the dead models_dir=/=entity_glob=/=exclude_suffix (legacy JSON) columns.
3. Update all --component refdata-cpp references to refdata across recipes (doc/recipes/codegen/*.org), analysis docs (codegen-model-unification.org, entity_commissioning_reference.org), claude_code_settings.org, and the compass codegen regenerate help string in compass.py.
4. Diff-check: regenerate every address via the old dual-component invocation vs the new single-component one and confirm zero diff (acceptance criterion 1).
5. Run the codegen regression test suite; confirm zero diff.

• module type (ores.codegen.module) was already excluded from org discovery pre-change — its type was never in manifest.py's _CODEGEN_ORG_TYPES set. Only junction needed an explicit new exclusion mechanism (exclude_org_types catalogue column).
• Verified via a script harness that refdata and refdata-cpp discovered an identical 73-file set before the change (both point at the same modeling_dir, and the legacy models_dir is dead for both) — confirming the split was already vestigial. After the merge, refdata discovers 65 files (73 minus the 8 junction-typed files).
• Zero-diff confirmed via compass codegen entity generate <name> --address <addr> --diff across SQL/C++/Qt addresses. One entity (rounding_type, ores.sql.schema) showed a diff, but it is pre-existing template drift unrelated to this change (documented as an open issue in entity_commissioning_reference.org: "sql profile dual-fires old and new SQL templates").
• Left refdata-cpp mentions untouched in historical docs (doc/agile/**, doc/plans/**) and in migrate_json_to_org.py's docstring, which narrates a past migration task by name — editing these would misrepresent history.
• Codegen regression suite (projects/ores.codegen/tests/) requires pystache on the path; the codegen venv's own interpreter is version-skewed against its installed packages (stale symlink to /usr/bin/python3 after a system Python upgrade). Ran via system pytest with PYTHONPATH pointed at the sibling ores.compass/venv site-packages (which has pystache for a matching interpreter) plus src. All 50 tests pass.

The refdata / refdata-cpp split in the component catalogue was already vestigial (both discovered an identical 73-file set via the same modeling_dir; the legacy JSON models_dir was dead for both). Collapsed into a single refdata row backed by a new exclude_org_types catalogue column, which excludes the 8 junction-typed files from bulk regeneration (65 files remain). All --component refdata-cpp references updated to refdata across recipes, analysis docs, and the compass help string; historical agile/plans records intentionally left untouched. Zero-diff confirmed across SQL/C++/Qt addresses (one pre-existing, unrelated drift noted, not introduced here). Codegen regression suite: 50/50 pass. All four acceptance criteria met.