[anth-volk]

Fixes #8

Summary

Copies Microcosm's Orrery projection into the collector, so the collector's future Orrery worker can turn a build's saved compiler schema (graph.schema.json) and execution overlay (graph.execution.json) into Orrery JSON without Microcosm's content store, numpy or pandas. Tests compare the copy's output with Orrery JSON saved from Microcosm's own projection and require identical bytes.

Nothing in the service calls the projection yet. The worker, its image, validation with Orrery's Node validator and publication are later changes.

What was copied, and from where

The source is PolicyEngine/microcosm#1295 at its head, 8d159a8c5653ad63affb1a5d2e39034f877db2ff (#1295 is still open), from packages/microcosm-graph/src/microcosm/graph/. #1295 is what makes the projection need only the standard library.

Module	State in the copy
orrery.py	Changed: keeps orrery_document_from_schema, orrery_json_from_schema and every helper they call, unchanged. It leaves out main and its file-reading and reference-rebasing helpers, because the command's --evidence/--store option reads the content store, and orrery_document/orrery_json, which compile a Graph object first.
orrery_evidence.py, execution_overlay.py, presentation.py, schema.py, serialize.py, decl.py, canonical.py, expansion.py	Unchanged. schema.py, serialize.py, decl.py and expansion.py are needed because the projection checks a schema by recompiling its embedded graph.

provenance.py records the repository, commit and directory, each source file's git blob ID, and what orrery.py leaves out. A test checks that every unchanged module still has its source's blob ID. The adapter identifier the projection writes at metadata.adapter, microcosm.graph.orrery.v1, is unchanged, and a test checks it. The README's new "Orrery projection" section says where a later version comes from (Microcosm's main, until Microcosm removes its copy) and how to take it.

Where it lives

telemetry_collector/orrery/projection/:

• The wheel ships one import package, telemetry_collector, and tests/test_packaging.py requires that, so carried-over calibration-diagnostics commits keep merging. A second top-level package would break that.
• telemetry_collector/orrery/ leaves room for the worker's validation and publication modules beside projection.
• The copied modules keep their Microcosm file names, so their relative imports work unchanged and each file can be compared with its source by blob ID.

The modules are in the container image, because the Dockerfile copies telemetry_collector, but the application doesn't import them.

Evidence that the output is byte-identical

tools/generate-orrery-projection-fixtures.py ran Microcosm's projection at 8d159a8 (Python 3.14.6) and saved each input with the Orrery JSON Microcosm wrote from it, under tests/fixtures/orrery_projection/. tests/test_orrery_projection.py loads the saved overlay with the copied reader, projects with the copy, and compares bytes.

Case	Overlay	Expected output	Result
toy-recorded-execution: Microcosm's toy country with grouped presentation, a scope boundary and source references; the overlay holds a computing phase and its replay, loaded from saved run evidence the way UK finalize loads it	yes	191,154 bytes	identical
toy-declarations-only: the same schema without an overlay	no	57,409 bytes	identical
toy-failed-check: one phase whose threshold check fails; no presentation contract	yes	122,309 bytes	identical
declaration-features-titled: rewrites, expansion, reweighting, an artifact input, numbers outside JavaScript's safe range, and a title	no	29,534 bytes	identical
uk-calibration-summaries: UK calibration nodes on a three-household fixture, with UK's artifact summaries and one failed summary provider	yes	131,853 bytes	identical
uk-full-dense-declarations: the complete UK dense graph's declarations and presentation contract (5,805 nodes, 21,620 edges)	no	18,899,459 bytes (stored gzip-compressed)	identical

Limits, found on toy-recorded-execution: the generator searched for the smallest value of each limit at which Microcosm still projects it. The copy writes the same bytes at that value and raises the same error one below it:

Limit	Smallest accepted	Error one below
_MAX_NODES	40	Orrery adapter: node limit.
_MAX_EDGES	71	Orrery adapter: edge limit.
_MAX_OUTPUT_BYTES	320,674	Orrery adapter: prospective metadata size.

A test also checks that the defaults match Microcosm's (20,000 nodes, 100,000 edges, 64 MiB output, and the schema reader's limits).

As a check that the comparison is sensitive, I changed one value the projection writes (metadata.truncated) locally and ran the file: 9 failed (all six cases and the three limits), and it passed again once the value was put back.

The inputs are synthetic: Microcosm's test fixtures and the UK graph declarations, with overlays from running those fixtures. They hold no build data. I checked them for local paths and machine names; there are none. The fixtures total 2.8 MB, 2.2 MB of it the gzip-compressed UK dense schema and output.

Other tests in the file, ported from #1295's: the copied overlay reader refuses another protocol, a non-object, duplicate keys, non-finite numbers, non-UTF-8 bytes, a FIFO and a file over the byte limit, and it reads each saved overlay back to the same canonical bytes. A static check confirms that the copied modules import only the standard library and each other.

Commits (red, then green)

Commit	Red	Green
Project saved schemas and overlays to Orrery JSON as Microcosm does	tests/test_orrery_projection.py failed to import (ModuleNotFoundError: telemetry_collector.orrery), so none of its 23 tests could run	23 passed
Record the Microcosm commit and files the projection was copied from	tests/test_orrery_projection_provenance.py failed to import (ModuleNotFoundError: ...projection.provenance), so none of its 5 tests could run	5 passed (and the 23 above still pass)
Name the projection's Microcosm source with the collector's own word	The existing check test_collector_uses_store_vocabulary (tests/test_deployment_workflow.py) refuses the word "repository" in the collector's own files. That file isn't changed here, so I didn't run it locally. A text search with the same exemptions found 6 occurrences in 4 of this branch's files.	The search finds none; the 28 tests above still pass. CI runs the check itself.

After each commit, I ran the new test files again: 23 passed, then 28, then 28.

Tests run locally

• uv run pytest -q tests/test_orrery_projection.py tests/test_orrery_projection_provenance.py: 28 passed, about 2.4 s.
• ruff check and ruff format --check on every changed Python file, which passed. The copied modules pass both unchanged.
• bash tools/check-changelog.sh, which passed.

I didn't run the rest of the suite; CI runs it.

Dependencies and deployment

No new dependencies: pyproject.toml, uv.lock, the Dockerfile and the workflows are unchanged, and the change reads no environment variables. Deployment stays switched off: COLLECTOR_DEPLOY_ENABLED is untouched.

Related pull requests

• Project Orrery artifacts without the content store, and write the execution overlay as its own file microcosm#1295 is the source. If it changes before merging, the copy and the expected outputs should be taken again from its final version (README steps). Once it merges, Microcosm's main is the source.
• Reconstruct declaration-only Orrery publications for UK builds before #1144 microcosm#1297's UK reconstruction tool imports Microcosm's own microcosm.graph.orrery.orrery_json_from_schema and canonical_json. This PR doesn't change that dependency. It matters only when Microcosm removes its copy.

🤖 Generated with Claude Code