Cost model v2

Costs describe how far a result is from an ordinary, checked-in source build. They are fixed by the library and cannot be lowered by project configuration.

Cost per unit	Class
0	Native output, verification, receipts, and path normalization
1	Proven non-emitting declarations or compiler-state reservations
5	Emitting carriers, generated suppliers, or metadata normalization
10	Ordering existing compiler-produced units
25	Strict intact same-symbol compiler donor
50	Intact donor with resize, EH, or relocation-layout handling
100	Cross-unit donors or manifest-only source interventions
250	Instruction-level semantic rewrites
500	Mosaics, hybrids, repacks, or comparable binary surgery
10,000	Direct oracle-byte installation

Most interventions contain one typed intervention unit. A source_overlay_graph instead contains one source_overlay_edit unit for every declared top-level edit, one generated_translation_unit unit for every generated translation-unit placement, and one link_admission unit for every admission. Each of those units receives the class weight of 100. Two edits therefore cost 200 whether they are represented by one intervention ID or two; ID bundling cannot reduce the charge.

The project total sums typed units after deduplicating identical intervention IDs. Reports expose both intervention and unit counts, distinguish direct function cost from allocated shared cost, and retain non-additive exposure so shared work is never counted twice. The target and class breakdowns are two views of that same total, not extra costs to add together. Function attribution also conserves the total exactly: attributed function cost plus cost remaining at target or TU scope equals the project total, even when equal shares are fractions.

Every beneficiary must be a unique, canonically ordered function scope in the intervention's target (and in the same translation unit for a unit-scoped intervention). Project loading also requires that scope to exist in the authoritative function universe established by direct function interventions or oracle identities. Unknown “ghost” labels are rejected, so adding an unanchored beneficiary label alone cannot dilute a shared function allocation.

A shared donor is charged once. Reusing the same donor for another function widens its beneficiary list but does not charge the donor again; only newly added intervention work changes the project total. Each beneficiary's exposure still shows the donor's full cost, and remains a non-additive diagnostic.

An oracle install costs 10,000 points per intervention unit, not per referenced byte. Exact range and byte counts remain separate authenticity evidence in the report; the cost score deliberately ranks any reference-byte installation as a large departure from an ordinary build.

Runtime work is not intervention cost

The score above measures semantic distance, not elapsed time or compiler-process count. For a project-level source_overlay_graph, the closed validator derives a declaration counterfactual. Declaration-only leaves add no compiler process. Strict source leaves add one evidence-only counterfactual compile per exact source owner; strict headers conservatively select all ordinary compilers because reader exposure is not independently sealed. Counterfactual objects are never linked. These sparse invocations appear in producer progress and each obeys the compile timeout.