Wet-evidence convergence: one shared wet-admission authority, and the local route migrated onto it (7 refusals in, 7 typed causes out)

[gunbai-bot]

The local wet route stops being its own authority and joins the shared wet-evidence one. Four claims land together because the migration is not separable from the decoder that reads its rows, but they are separable to review, and reviewing them in this order is the cheapest path:

1. The shape migration (src/v2/workflow/local_repo_wet_terminal.dag)

LocalRepoWetScheduled becomes WetScheduledClaim, the qualified identity string becomes a WitnessIdentity { module_path, function } record, and expected: LocalRepoWetExpectPassed becomes expectation: ExpectedToHold {}.

The identity change removes a duplicated fact rather than adding a field: every row's old identity ended in "." + function, so the qualified name was authored twice per row and could disagree with itself. It is now derived by witness_identity_qualified_name, the same projection every other identity consumer already uses, and a witness asserts the two spellings agree across the whole roster.

2. The seed decoder (src/v1/stage0/src/cli_run/required_floor_runner.rs)

The Rust seed reads these rows, so the shape change breaks it unless ported in the same commit. Every refusal is preserved and one is strengthened: the old entry_module was derived by stripping ".{function}" off the identity, which silently accepted a row whose identity and function disagreed; it now compares identity.function to the row's function directly and refuses with "one member, two names". ExpectedRed still refuses — the executor has no arm for it and says so rather than treating it as pass.

3. The 57 namespace admissions are deleted, and the deletion is charged correctly

NAMESPACE_TRANSITION_ADMISSIONS held 57 rows for #10077's required_floor -> floor_terminal_ledger relocation. #10077 merged as a4a6db175d2, so main carries the relocation and no run after it can produce those deltas.

The wave phase's own refusal wording states the charging rule: consumed rows are "due for deletion on this roster-touching change". Re-tensing their trigger touched the roster, so the obligation landed on the change that went looking for it, and it is paid here rather than left as a debt for a stranger. The roster returns to empty, and empty is not permissive — a run with a real delta still refuses it as UNADJUDICATED.

The evidence that the deletion took nothing load-bearing

Read off the merge ref, which is what CI judges — a census on the branch tip would answer about a tree nobody merges. Run 33700975802, base=53934adc676 head=f5b67e3e677:

modules_compared=4571  binding_rows_compared=691488  closure_rows_moved=172  deltas=18
  12 ExplicitlyEvaluatedZeroDelta
   4 AuthoredReferenceResolution
   2 SameDeclarationIdentityRebind
   0 TargetChanged        0 NewUnresolvedness

The deleted rows were all TargetChanged, and this branch produces no TargetChanged delta. So the answer is structural rather than incidental: it is not that the roster happened to go unqueried, it is that every delta this branch produces is in a class that auto-admits and never needed a row.

One limit of that green, stated rather than left for a reader to assume. The phase reports unadjudicated, stale and consumed only in its refusal line; on ADMITTED it prints the verdict alone. So unadjudicated = 0 is measured — a nonzero is precisely what would have made that line a refusal. stale = 0 and consumed = 0 are zero by construction, not by measurement: with an empty roster no row can fail to match a delta and none can be satisfied at base. They are sound and they carry no information about the mechanism, so this run must not later be cited as evidence that the stale detector works.

4. Two rows ported across the migration, not chosen against

#10055 (992fd4423f6) added two host-probe rows to this roster on main while the branch was migrating its shape. That conflict is a shape change meeting new members — both sides right about different things — so the resolution re-authors their rows in the new shape and preserves their module_path and function verbatim. Which witnesses that lane admits is #10055's authority; how the roster is spelled is this branch's.

Their module_path omits the .host segment their entry path carries, which looks like an entry-module disagreement and is not one: the file declares module test.claim.host_cli_dependency_wet_witness_test on line 1, and every sibling in dag/test/claim/host/ likewise omits the segment — paths are discriminators, not gospel (DESIGN §3). Predicted statically, then confirmed by execution:

[local-repo-wet] identity=test.claim.host_cli_dependency_wet_witness_test.observe_echo_wet_posix_command_v_check_does_not_crash expected=passed observed=passed
[local-repo-wet] identity=test.claim.host_cli_dependency_wet_witness_test.observe_npm_wet_posix_command_v_check_does_not_crash  expected=passed observed=passed

Membership prose and roster agree at 13 rows in three groups.

Verification

Run 33700975802, head-matched on f227058a: lane=build phases_run=2 failed=0, lane=witnesses phases_run=3 failed=0, zero FAILED PHASE lines, namespace-wave-admission ADMITTED.

required-floor: planned=3519 executed=3519 not_attempted=0 passed=3444 known_red_held=26 failed=0
                interrupted_cpu_deadline=0 interrupted_wall_deadline=0 stale_quarantine=0
required-floor: verdict=FloorClean unexpected_failures=0 verdict_incomplete=0

route_gap_held=49 is unchanged and is not this PR's subject — it belongs to #9725, the last PR in the sequence.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GvVoivi7L449wbh6rjeJY4