fix(represent): the margin and the registry share one typed vocabulary (BS2-F2)

[chrishayuk]

BS2-F2 — found by the real-bank integration smoke test that BALANCED-SEARCH-2's pre-registration demanded, before rung 4 starts.

The defect

The calibration registry keyed "route flip rate". ConstraintVector::of emitted "route flips". They never matched, so lookup missed and evidence_for fell through:

if let Some(e) = self.lookup(statistic, scale) { return e.verdict; }
if status.is_priceable() { Direct } else { Unusable }

A route-flips margin is a count, counts have no tail to be thin, so measurement_status is Measured, so is_priceable() is true — and the statistic ROUTE-CAL-1 established as ordering-only came back Direct.

That is the invariant the whole evidence ladder exists to protect, violated by the exact statistic the doctrine named it about: a proxy affects promotion class and never becomes price, or route_flip_rate slowly becomes a shadow behavioural contract.

Demonstrated, not argued — the same scenario run against pre-fix code:

PRE   margin.what="route flips"  evidence=Direct         priceable=true
POST  Statistic::RouteFlips      evidence=OrderingProxy  priceable=false

Two of the four registry keys did match (kl p99, routed mixture moved at p99), which is what made the mechanism look like it worked.

The fix

Statistic — a typed vocabulary — replaces the free string on all three joins:

join	was	now
Margin::what	String	Statistic
SearchCalibration::statistic	String	Statistic
ProxyObservation::for_criterion	String	Statistic

That third one had the same latent defect: proxy_for joins on the same string, and a mismatch silently yields "no proxy" — which ranks Uninformed, below ProxyRisky.

Statistic is deliberately finer than Criterion: RouteDisplacement carries both a p99 and a max limit that bind independently, and at 256 positions the p99 is a maximum wearing a percentile's name while the max is exactly what it says.

Admission::Refused keeps its Vec<String> — it reports, it does not join, and keeping it preserves the refusal report's alphabetical ordering.

Why the fixtures missed it

They passed their own literal to both sides of the join, so they were green throughout. The new regression test takes its key from a margin ConstraintVector::of actually built.

Not addressed: BS2-F1

No gate sets route_flip_max — kimi_logit_v3 sets it None and balanced-v1 inherits that — so in practice this margin is absent entirely. v3's doc says it "keeps the counts as DIAGNOSTICS and drops them as authority", but there is no diagnostic-side vehicle, so a criterion kept "as a diagnostic" produces no margin. That is a gate-design question, and it had to come second: setting route_flip_max first would have started pricing a proxy.

Verification

fmt --check, clippy -D warnings, E0 (10 passed), full suite (3618 passed, 0 failed), cargo check --workspace --all-targets. Coverage policy passed — search_evidence.rs 100.00%, constraint.rs 99.43%, assessment.rs 99.20%, promotion.rs 97.09%, quality.rs 96.52%; crate 93.45%.