Task: Close M06, the synthetic TypeScript twin

Every synthetic protocol header has its TypeScript twin committed, both sides generated from an org model, and the twin-coverage check covers the component.

• Every synthetic protocol header has a committed TypeScript twin: 12 protocol modules under packages/wire-protocol/src/generated/synthetic/protocol/ and the 9 domain modules they import.
• Every one of those twins is generated from an org model. The four non-entity operation surfaces are operation models, so no hand-written protocol header remains in the component.
• check_protocol_twin_coverage.py passes with synthetic among the covered components.
• check_component_drift.py --component synthetic regenerates every address and reports no drift, so a twin cannot go stale silently.

Three steps, in the order the gates forced them.

• The scaffold's acceptance was wrong twice, and both corrections are worth keeping. It said nine twins, but the request that proved it named only the ores.ts.protocol address: the domain facet is required too, because a protocol twin imports its entity's interface.
• The scaffold also named the wrong blocker: it said the models needed the * Physical space address mechanism before their TypeScript could be generated. They do not, and a targeted run created all nine entity twins from the models as they stood.
• The drift gate needed no change. _committed_ts_facets scopes a TypeScript facet by whether its output directory exists, so committing the twins puts ores.ts.domain, ores.ts.protocol and ores.ts.trade_type into the address set on their own.
• An earlier claim that the drift gate cannot see untracked generated files was wrong. It diffs the tree with git diff, but it also records the untracked set before regenerating and fails on anything regeneration materialises (check_component_drift.py, "untracked file(s) materialized by regeneration"). The gate that let M06 hide is the twin gate's scoping, not this one: both gates agree there is nothing to check until a component's first twin lands.
• component_registry.py carried a stale paragraph saying synthetic "is not listed yet" while the tuple lists it. Corrected here to record the M06 closure and the four items still open.
• clever_dijkstra owns the clean-standard record, task_clean_synthetic.org, and its M06, P04, V05 and V06 rows still read as open or held. The file is on main, so the rows were struck here rather than handed over in a message: M06 is closed, P04 records that the operation surfaces are modelled now, and V05 and V06 drop the twin from their open lists. This is the one file this branch edits that another environment owns.
• V06's row claimed COMPONENTS_UNDER_TEST stays ("iam",), which is false: the tuple carries synthetic. Corrected to say that the registry lists a component with four items still open. That disagreement is a gate question in its own right – the registry asserts every item passes while the record says four do not, and nothing cross-checks the two – but it is out of scope here.
• Three gate failures surfaced on the merged tree and none was synthetic's. format was the one exception: the new line in ir_curve_preview_handler.hpp needed clang-format. The other two came from 79e1123446 ("Serve the sub-periods coupon type list"), which added a refdata entity but left the hand-written composition points behind: the service event registrar and the shell refdata composite each needed one include and one call. That left codegen-tests red on two cases, and the same refdata work had left ores_schema.{refdata,trading}.puml and the index stale. All three were fixed on this branch, because the story cannot land through a red gate and each fix is the one the check names. The diagrams were regenerated without the timestamp-only churn the generator writes into all 22 files, and the SVG redraws for diagrams whose source had not changed were left alone because the render is not byte-stable.

Complete. 21 TypeScript modules are committed under generated/synthetic/: 9 domain twins and 12 protocol twins, all generated from the models. The four non-entity operation surfaces are operation models now, so no hand-written protocol header remains. The drift gate and the twin gate both cover the component.

Synthetic speaks TypeScript. 21 generated modules are committed: the 9 domain interfaces and the 12 protocol modules, each protocol module the twin of a generated C++ header. The four hand-written operation headers are gone, replaced by three operation models from which both sides generate, and the two IR operation surfaces share one model so parameter_spec is declared once. The drift gate regenerates every address and reports no drift, and the twin gate now covers the component. The M06 item on the clean standard is closed; P04 is superseded for synthetic.

Verified on the merged tree: the whole tree builds, compass test run -- -R synthetic passes 4 of 4 (ores.synthetic.api, .core, .service and ores.shell.synthetic), check_protocol_twin_coverage.py reports 418 headers in 18 components with no opt-out, and check_component_drift.py --component synthetic reports no drift. Three gate failures on the merged tree were fixed here; two were trunk's from 79e1123446, and the third was a clang-format tweak.