gate: lsp — the override-suppression registry

[Antawari]

What this gate adds

Substitutability (Liskov) enforcement has two surfaces; only one is new:

1. Credit-existing — substitutability itself stays the pinned mypy gauge: an override with an incompatible signature is an [override] error under configs/mypy-base.toml (mypy's own note names the Liskov substitution principle), surfacing as a NEW error through mypy | mypy-baseline filter per configs/BASELINE-CONVENTIONS.md. Nothing rebuilt.
2. Build delta — the override-suppression registry in cf-exemptions: every # type: ignore[override] token in the measured tree must trace to a five-field exemptions.json entry ({file, symbol_or_line, rule override, reason, approver}) or the gate fails UNREGISTERED_SUPPRESSION naming the site. Same mechanism and entry shape as the noqa/nosec families.

Budget

• 0 unregistered # type: ignore[override] suppressions — every site traces to a reasoned, approved registry entry, or the gate fails.
• Anchored to measurement, not invented: 4 live sites measured fleet-wide 2026-06-11 (re-verified at review time), to be registered at mount time in their own repos' registries. The count rides the frozen_count ratchet, shrink-only.
• The kit's own measured tree: 0 sites (self-compliance pinned by TestKitSelfCompliance).
• Signal-honesty note: the measured fleet is composition-dominant (inheritance depth 3-4 max), so the substitutability signal is honest but weak by construction — the gate's value is keeping it that way.

Control rods

• Scope rod: only the override code is registry-gated — the one token that silences the substitutability gauge entirely. Other type-ignore codes and bare # type: ignore stay the mypy gauge's jurisdiction (warn_unused_ignores), so the gate isolates exactly one factor and cannot creep.
• Bypass closure: a multi-code ignore ([override, misc]) is still gated — hiding the token in a code list is the obvious bypass. Tokens inside string literals never trip the gate (comments are read via tokenize, COMMENT tokens only).
• Loud pass: every covered suppression is printed with its site, entry, and approver — a visible exemption, never a silent one.
• Ratchet: override entries ride the same shrink-only frozen_count discipline as every other gated family; 1:1 entry coverage (EXEMPTION_ENTRY_OVERLOADED) blocks the collision undercount.

Verification (oracle outputs)

• Self-run against this repo: cf-exemptions --root . exits 0 (existing S603 entry covered, ratchet 1/1; no override tokens in the measured tree).
• Full suite: 276 passed. ruff check . clean. mypy src clean (0 errors).
• Wheel regression (the harsh oracle, per tests/test_packaging.py pattern): wheel built from this branch, installed into a clean venv with no source tree — the unregistered fixture fails exit 1 naming src/mod.py:7; the registered fixture passes exit 0 with the loud entry line.

Do not merge — pending maintainer review.