Retire Chronicle target profiles

[juaristi22]

Summary

• remove Chronicle-owned target profiles and target-resolution APIs now that Microcosm owns selection contracts
• limit consumer artifacts to fact rows and manifest hashes
• retire the target-profile agent lane and update the surrounding documentation and tests

Testing

• uv run --locked --extra dev ruff check .
• uv run --locked pytest -q --ignore=tests/test_chronicle_bundle.py --ignore=tests/test_chronicle_source_package.py (548 passed, 1 skipped)
• uv run --locked pytest -q tests/test_chronicle_source_package.py plus the focused consumer, contract, governance, and import tests (167 passed)
• uv run --locked pytest -q tests/test_chronicle_bundle.py -k 'not test_build_bundle_writes_merged_consumer_contract' (8 passed, 1 deselected)
• wheel build and import smoke test
• end-to-end build-consumer-artifact CLI smoke test

The 171,855-row full-bundle golden test was started locally but interrupted after eight minutes while computing the build ID. CI runs the complete test suite.

Fixes #172

[vahid-ahmadi]

Automated review pass (Claude Code, high effort, diff only — no execution). Deletion-dominant, so I pointed the review at what the retirement takes away rather than at new-code defects. Three findings, all of the same shape: a guarantee Chronicle used to enforce is now nobody's job, and its absence is silent.

1. policyengine_chronicle/consumer.py:~1640 — the schema version does not move with the contract

CONSUMER_ARTIFACT_SCHEMA_VERSION stays policyengine_ledger.consumer_artifact.v1 while the artifact's content contract changes incompatibly: the manifest no longer carries profiles, and artifacts that do are now hard-rejected.

The if "profiles" in manifest: raise guard protects readers of old artifacts, which is the easy direction. The unprotected direction is a downstream that pins schema_version == ...v1 and consumes a new artifact: it validates, finds no profiles, and resolves zero targets — compiling to nothing rather than refusing. Given microcosm compiles its target surface from this layer, an empty result there is a calibration that silently loses a family, not an error someone sees. Bumping the version is what makes the incompatibility loud on the consumer side.

2. chronicle/core.py:82 and the deleted resolve_profile_targets — the assertion policy is removed but the ambiguity it resolved is not

ASSERTION_POLICIES and DEFAULT_ASSERTION_POLICY = "observed_only" are deleted along with the resolver that enforced them, the per-series observation-vs-projection tie-break, and the ambiguous_assertion_at_period warning.

But ALLOWED_ASSERTIONS is unchanged, and the published feed still carries both observation and source_projection rows for the same series at the same period. So a consumer reading consumer_facts.jsonl directly now sums both — a double count that Chronicle refused by default until this PR. The ADR reassigns the responsibility to consumers, which is a legitimate architectural call, but nothing in the +271 added lines gives a consumer the tie-break the resolver implemented, and no fact-level marker makes the ambiguity visible at read time. A consumer that does not know to look will not fail; it will get roughly double.

This is the finding I would most want a ruling on, because it is a real semantic change shipping under a retirement label. If the intent is that consumers pin assertion explicitly, the feed refusing an unpinned read — or at minimum a documented, machine-readable statement of which series carry both — would keep the guarantee somewhere.

3. Deleted PeriodContractError / PeriodAlignmentDeclaration — the period-alignment guard has no substitute

The contract that hard-failed on consuming a fact at a period other than its reference period is deleted, with nothing Chronicle-side replacing it. Facts-only artifacts still carry period, but nothing enforces alignment, so the regression that guard was written for — microcosm#212, per its own docstring — is reachable again from the consumer's side.

Checked, not a finding

I checked pyproject.toml for stale packaging references to the fully-deleted policyengine_chronicle/target_profiles/: packages = ["chronicle", "db", "packages", "policyengine_chronicle"] is directory-level with no per-file globs, so nothing dangles.

---

The common thread is worth stating plainly: each of these three moves a check from "Chronicle refuses" to "the consumer should know", and in each case the consumer that does not know gets a plausible wrong answer rather than an error. If the retirement is the right call — and consolidating on facts-only artifacts may well be — then the replacement guarantees seem worth landing in the same PR, so there is no window where the old enforcement is gone and the new expectation is undocumented.