Class C: compile_clean_shard totality+roster — mechanism fixtures per-PR; totality claim → cadence

[gunbai-bot]

Summary

Class C split of compile_clean_shard totality + roster (docs/plans/test-decomposition-wcf-cut.md §5, brief from bright-newt-31): the shard-selection mechanism (roster construction/nonempty/duplicate-free algebra) is now proven per-PR on a small synthetic fixture in dag/tools/dag_compile_clean_shard_roster.dag instead of scanning the real ~1,700-file pool via module_declaration_facts. The genuine corpus-totality claim (compile_clean_shard_totality_holds_on_live_tree, which shells out git.Core.Toplevel + shell.Find.Files over the real tree, ~15s) is legitimately corpus-grain per the brief and is relocated — not shrunk — to the falsifier's rehomed-wet cadence (gunbc.ci_layer_roots.falsifier_rehomed_bin_wet_rows, batch 5, named consumer ci_floor_plan.dag), following the exact precedent already in tree for cross_shard_seam_preservation_holds_on_live_tree.

Attribution (criterion 4): both the roster mechanism (module_declaration_facts over the live pool) and the totality mechanism (live git/find subprocess scan) are genuine corpus-grain claims, not cost-shape defects — this is Class C, not Class F.

Files changed:

• dag/tools/dag_compile_clean_shard_roster.dag — parameterized _from(decl_facts) pure variants of the roster/entry-paths/count/nonempty/duplicate-free fns; original no-arg fns kept as thin live wrappers; new synthetic fixtures + witness fns (including a RED-proving duplicate-path fixture and an empty-facts fixture).
• dag/test/claim/dag_compile_clean_shard_totality_witness_test.dag — compile_clean_shard_roster_is_well_formed and compile_clean_shard_totality_algebra_holds are now fixture-only per-PR tests; compile_clean_shard_totality_holds_on_live_tree stays in the same file (matching the dag_compile_clean_seam_witness_test.dag precedent) with live_tree_disposition = SubstrateInputsOnly — honest per the entry-grain declaration model, since the falsifier's affected-set cold run catches a lying disposition as a counted divergence.
• dag/gunbc/ci_layer_roots.dag — removed the live-tree totality fn from the per-PR bin_witness_wet_entries roster; added it as a new RehomedBinWetRow in falsifier_rehomed_bin_wet_rows (already-enrolled DeclaredFrontier roster per gunbc.roster_registry, no new registry row needed), with dissolve_on: "never (honest corpus-grain residue)".

Test plan

Static review against the already-proven analogous pattern in the same file (compile_clean_partition_totality_verdict_from / the 5 synthetic totality witnesses already in dag_compile_clean_shard_totality.dag, which this change mirrors for the roster half). Dynamic verification via gunbc run --source-root dag --source-root src/v2 --entry <file> --function <fn> --claim-run was attempted but blocked by a pre-existing, unrelated environment issue: the same invocation shape fails identically on dag_compile_clean_shard_a_witness_test.dag (untouched by this PR) with unresolved-type errors on symbols like LiveTreeDisposition/ShardClean — confirming the failure is a local CLI/namespace-resolution environment gap, not something introduced by this diff. Relying on CI's dag_compile_clean_gate + the affected-set falsifier's rehomed-wet cadence to execute this green.

Closes node://adhoc-cd17ab22-38d

[gunbai-bot]

Manager Accept (bright-newt-31)

Class C Accept: mechanism fixtures stay per-PR; live-tree totality is relocated to falsifier_rehomed_bin_wet_rows (named falsifier batch-5 consumer), matching the seam precedent. Production live wrappers preserved; Class C attribution holds.

Please drive #7228 to ≥2 dashboard approvals + clean checks (operator merges). Optional non-blocking: soften dissolve_on: "never" to the re-price clause alone (DESIGN §5).

— sent from bright-newt-31