Bundle expectations: per-source deltas that compose across parallel PRs

[juaristi22]

Filed from #137's review: four UK-wave PRs (#137, #139, #141, #155) branch from the same base and each rewrites the absolute totals in tests/test_chronicle_bundle.py (fact_count, source_package_count, per-entity and per-source splits). Each is right on its own branch and stale the moment a sibling merges — so the current shape costs one hand-recomputed rebase per sibling merge, and the tempting conflict resolution (take one side's numbers) yields a green test that has stopped checking anything.

Proposal: keep the whole-bundle snapshot as a final tripwire, but move the load-bearing assertions to per-source (or per-package) counts:

• a by_source mapping asserted key-by-key, so a PR touching only slc and a PR touching only obr compose without conflict — their diffs to the mapping are disjoint;
• derived totals (sum of the mapping) asserted against the snapshot, so drift between the two representations still fails loudly;
• per-package fact counts inside each source's packages, so an accidental record-set drop inside one package can't hide in a source-level aggregate.

A sibling merge then rebases with zero constant edits when sources are disjoint (the common case), and a genuine collision surfaces as a real semantic conflict on one mapping line instead of four unrelated integers.

Current cost is real: this wave alone has done four hand re-measurements (~13-minute suite runs each) that per-source deltas would have made no-ops.

🤖 Generated with Claude Code

[vahid-ahmadi]

Endorsed — this is the shape I was asking for when I raised it on #137, and the three-layer design is right: per-source mappings carry the load, per-package counts stop a record-set drop hiding inside a source aggregate, and the whole-bundle snapshot demotes to a tripwire.

One nuance worth writing into the implementation so nobody relitigates it at the first sibling merge: the snapshot's absolute totals will still conflict between parallel PRs — that line is shared no matter what. What changes is the character of the conflict. Today resolving it means a ~13-minute hand re-measurement or a silently wrong "take one side" merge; under this design the resolution is mechanical (recompute the sum) and guarded — a wrong resolution fails loudly against sum(by_source). So the honest claim is "conflicts shrink to one mechanically-checkable integer per axis", not "zero edits" — zero holds for the mapping lines when sources are disjoint, which is the case that matters.

Go ahead before wave 4 — this shouldn't wait on the #166 adjudication, since it touches test structure only.