Add Eurostat COFOG adapter and Irish budget pilot

ROADMAP.md

@@ -288,6 +288,24 @@ the polygon in the same document. The adapter links to the alert's own endpoint,
 workaround rather than a mapping. A publisher that inlines geometry currently has nowhere to
 put it. See [examples/pilot-nws-alerts/README.md](examples/pilot-nws-alerts/README.md).
 
+## Eurostat COFOG adapter — DONE
+
+`tools/adapters/eurostat-cofog.py` converts Eurostat general government expenditure by
+function. Ireland 2022: 79 lines across 10 COFOG divisions.
+
+One defect, the same shape as the UK Contracts Finder finding: the budget arithmetic check
+demanded exact equality between a parent line and the sum of its parts, which **no statistical
+publisher can satisfy**. Four of ten Irish divisions failed by 0.002% - rounding, because
+Eurostat reports in millions to one decimal place. The check now infers the rounding unit from
+the greatest common divisor of the amounts and allows half a unit per figure; real
+discrepancies are still caught.
+
+Deliberately not added: a precision or unit-multiplier field on `amount`. The granularity is
+already recoverable from the data, so a declared field would restate what the numbers say and
+add something a publisher can get wrong.
+
+See [examples/pilot-eurostat-budget/README.md](examples/pilot-eurostat-budget/README.md).
+
 ## The honest risk
 
 Seven profiles with one fictional example city is a demo, not an adopted standard. Profiles

examples/pilot-eurostat-budget/README.md

@@ -0,0 +1,76 @@
+# Pilot — Eurostat COFOG (Ireland)
+
+**Real data.** Generated by `tools/adapters/eurostat-cofog.py` from the Eurostat
+dissemination API, which publishes general government expenditure by function for every EU
+member state.
+
+```bash
+python3 tools/adapters/eurostat-cofog.py --geo IE --year 2022 \
+  --out examples/pilot-eurostat-budget
+```
+
+79 budget lines across 10 COFOG divisions, €120bn of Irish general government expenditure.
+
+This is the most direct test available of the `budget` profile: Eurostat classifies by COFOG,
+which is the classification the profile adopted, and it is the authority that maintains it.
+
+## The defect it found
+
+**The arithmetic check demanded exact equality, and no statistical publisher can satisfy it.**
+
+Four of Ireland's ten COFOG divisions failed the parent-equals-sum-of-children rule — by
+€100,000 to €200,000 on figures of €10 billion. That is 0.002%, and it is **rounding, not
+error**: Eurostat reports in millions to one decimal place, so independently rounded
+sub-groups cannot sum exactly to an independently rounded division.
+
+The rule was right in principle and wrong in practice, the same shape of defect the UK
+Contracts Finder pilot found in `areaServed`. The check now infers the publisher's rounding
+unit from the greatest common divisor of the amounts and allows half a unit per figure:
+
+```
+parent 1,000,000 but children total 250,000 (difference 750,000; rounding tolerance 250,000)
+```
+
+Real discrepancies are still caught — verified against a fixture 75% out — and the tolerance
+is reported so the reader can judge it.
+
+## Why no precision field was added
+
+Eurostat reports in **millions**; the adapter multiplies to whole euro, which implies
+euro-level precision that the source does not have. The obvious response is a
+`precision` or `unitMultiplier` field on `amount`.
+
+It was not added, because **the granularity is already recoverable from the data**. Every
+figure in this dataset is a multiple of €100,000, which is exactly what the tolerance
+inference reads. A declared field would restate what the numbers already say, and would be one
+more thing a publisher can get wrong or omit.
+
+## Notation, not semantics
+
+Eurostat writes COFOG as `GF01` and `GF0101`; the profile uses `01` and `01.1`. A pure
+notation difference, but one that has to be handled: a consumer matching `GF0101` against a
+code list of `01` finds nothing. The adapter normalises.
+
+## Two structural traps
+
+**A `TOTAL` row sits alongside the breakdown.** Imported naively it would double every figure.
+It becomes the budget's `totalAmount` rather than a line.
+
+**Divisions and their sub-groups arrive in the same response.** Emitting both as flat lines
+would also double-count, so sub-groups nest under their division through `isPartOf` /
+`hasPart` — which is what surfaced the rounding defect in the first place.
+
+## What Eurostat does not publish
+
+**Named organizations.** The publisher is ESA 2010 sector S13, "general government" — a
+statistical sector, not a body. There is no ministry breakdown here, so
+`administrativeClassification` is empty across all 79 lines. The profile treats that field as
+the thing that joins a budget to the org chart; against statistical sources there is nothing
+to join to.
+
+**Economic classification.** This dataset is functional only. GFSM economic codes would need a
+different Eurostat table.
+
+**Phase.** These are outturn statistics, so everything maps to `executed`. A proposed or
+approved budget for the same country would come from the national ministry of finance, not
+from Eurostat — which is precisely why `budgetPhase` is required.