Discriminate LocalRead from Refused in the G2 under-selection control

[gunbai-bot]

Follow-up to #8142 (merged): makes that control's red discriminating.

The gap

specimen_classifies_as_runtime_read returns false in two states that are opposites on the safety axis:

• Classified{LocalRead} — the strongest safe-to-skip verdict the selection seam has, so it licenses a skip. This is the harm: a skipped witness presents as a faster green floor rather than a failure (under-selection).
• Refused{cause} — fail-closed, licenses nothing.

One Bool cannot separate them, so on its own the under-selection claim rested on the note beside it. DESIGN §4c: an annotation is never evidence that a machine claim holds, because no Accepted program can read one. #8142's note says outright that the discriminating LocalRead-vs-Refused receipt "is NOT enrolled here."

The repair

specimen_does_not_classify_as_local_read carries the other bit. The pair has a unique signature per reachable state:

state	pair reads
Classified{LocalRead} (today)	(false, false) — both red, so the red is agreement
Classified{RuntimeRead} (after the traversal repair)	(true, true) — both rows delete together
Refused{cause}	(false, **true**) — this claim greens alone

That third row is what earns the second evaluation. A drift from LocalRead to Refused is a real change of state with a different remedy, and today it is invisible because the first claim stays red either way. Greening alone makes it a StaleKnownRed refusal against its own admission row — the loud, typed, located signal the floor already routes.

Why expected-red rather than an ordinary green witness

It shares its sibling's evaluation cost, and the long-eval budget is coupled to expected-red across today's population (exact_witness_admission_witness_test every_declared_long_budget_row_is_also_expected_red_today). Asserting the desired state keeps this inside that population instead of retiring that claim, which would have widened scope. Both rows carry one dissolution obligation deliberately: same terminal construction, same trigger, neither survives the other.

Evidence

Release claim_batch, source roots dag + src/v2:

FAIL specimen_classifies_as_runtime_read        cpu=14182ms
FAIL specimen_does_not_classify_as_local_read   cpu=5689ms

(false, false) is the LocalRead signature. Refused would have made the second claim PASS, so the pair establishes by execution what the note previously only asserted.

Admission walls: 18/18 PASS on this change, including the four that could have rejected the new row — every_declared_long_budget_row_is_also_expected_red_today, each_budget_roster_is_homogeneous_in_declared_budget, no_batch_mixes_a_pre_verdict_row_with_an_assertion_false_row, and every_live_admission_is_reached_by_its_own_row (the last confirming the row reaches the new function rather than dangling).

Cost: the discriminator is 5.7s, not a second full 14s — the manifest memoizes across the two witnesses — leaving the pair near 20s against the 180000ms long-lane ceiling.

Scope

Two files, +46/-1. No production code, no traversal change. The G2 traversal repair itself (following data initializers as well as function bodies) is a separate lane and is not in this PR.