[MaxGhenis]

Status: draft. This is part of the base-unit migration Max ruled on 2026-10-21 ("why not enter all chronicle facts as numbers (ie 1000000)?", then "Yes, migrate everything"). No data changes. It lands after #306 and before any row carries the pair.

What this PR does

Chronicle's Thesis ledger is moving to plain numbers: 7716000 in count, not 7.716 in millions. Each row keeps what it was recorded as in source.recorded_value and source.recorded_unit.

• SourceProvenance accepts the pair. arch/store.py and tests/test_policyengine_ledger.py build it with **source, so without this change they refuse the first row that carries it. fact_to_mapping leaves out an empty pair, so facts with no pair serialize as they did and the bundles don't move.
• A new test holds the ledger to its pairs. Every recorded pair must state the stored value in exact decimals, and every unit_restatement must repeat the pair of the row it names. The test covers tampered rows too.
• Two tests stop assuming a row is never restated. I found both by running the suite against a simulated post-migration ledger, and both would have turned Arch checks red at the migration:
  • The strip fixture test counted lines with a given week, and a restatement adds a second one.
  • The strict xfail that documents full-file mode rewrites line 1. Once a later row supersedes line 1, the rewrite breaks that link and the check catches it, so the xfail would pass. It now rewrites a row nothing names, which is the case it documents.

No gate-surface file changes.

Invariants

• For every row with a recorded pair, Decimal(value) == Decimal(recorded_value) × 10^power, and measure.unit is the base unit of recorded_unit.
• A unit_restatement names an earlier row in a scaled unit and repeats that row's value and unit as its recorded pair.
• A fact with no pair serializes without those keys, and a fact with a pair survives fact_to_mapping and fact_from_mapping.

Evidence

• On this branch: tests/test_policyengine_ledger.py, test_build_series_catalog.py and test_thesis_append_adversarial.py give 300 passed, 5 xfailed. test_thesis_append_shim_isolation.py and test_receipt_shim_transparency.py give 53 passed.
• Simulated post-migration ledger (this branch merged with Resolve catalog supersede links against every row's effective assertion id #306, plus 82 restatement rows built from the real ledger, with the catalog regenerated):
  • build_series_catalog.py --check passes.
  • The first three files above plus test_release_chain.py and test_arch_facts.py give 373 passed, 5 xfailed.
  • The shim tests fail in that simulation because it cuts no release for the new lines: "HEAD release lineCount 253 does not match working-tree line count 335". The real migration cuts one, and its append gate runs those checks before it merges.

🤖 Generated with Claude Code