P1: Enforcement residue: granted_at stamping, kinds-filter probe, representations witness, red-script check mapping

[macanderson]

GOVERNANCE's freeze criterion 4 requires the conformance suite to fully enforce the documented requirements. #10's named items (F4 timestamps, B4 flood-frames) landed, but the surface shipped since (#37/#38/#41) has enforcement holes:

• Representation invariants are unenforced. representation_invariants() / representations_consistent() are unit-tested only — check_frames (contextgraph-conformance/src/lib.rs:467-529) never calls them, the fixture advertises representations: vec![], resolve: false (example-docs.rs:233-234), and no --misbehave mode trips a representation lie. PR Defer the frame-representations witness to unblock #33 #42 deferred this witness noting it was untracked; this issue tracks it. — caller landed in Pre-freeze: normative completeness + enforcement + host conformance (top-10 build) #60: check_frames now invokes the invariants. Residual (survivor): there is still no live witness — no fixture advertises representations and no --misbehave mode trips a representation lie, so the invariants are exercised against nothing.
• as_of is never probed. sample_query() sends as_of: None (lib.rs:459); the fixture has no disjoint validity windows, so a provider that ignores time-pinned retrieval passes. (Stella uses as_of for bi-temporal retrieval today — this is exercised surface, not speculation.) — landed in Pre-freeze: normative completeness + enforcement + host conformance (top-10 build) #60: the fixture now has disjoint validity windows and a probe that asserts the pin is honoured.
• Host-side max_frames audit is missing. contextgraph-host/src/host.rs:361 audits only respects_budget(max_tokens); there is no FrameCountLie-style audit, so only the conformance suite ever catches flooding. — landed in Pre-freeze: normative completeness + enforcement + host conformance (top-10 build) #60 via the host-conformance harness (B4), with a named drop reason.
• ConsentRecord.granted_at is never stamped (contextgraph-host/src/consent.rs:36,51). — still open (the last survivor of Close enforcement gaps: audit max_frames, validate temporal fields as RFC 3339, probe as_of #10).
• E1 (embedding-fingerprint exact-match gate) has no witness; kinds-filter honoring is unprobed (the sample query sends empty kinds). — E1 landed in Pre-freeze: normative completeness + enforcement + host conformance (top-10 build) #60: fingerprint helpers plus an embedding-fingerprint check with a raw-wire probe and the AcceptBadEmbedding misbehave mode. Residual (survivor): the kinds filter is still unprobed — sample_query() continues to send empty kinds, so a provider that ignores the filter passes.
• CI looseness: the red script accepts a misbehave mode caught by any check rather than the matching one; tests/docs_site_witness_test.py never runs in CI. — still open.

Related: #10 (predecessor — its named items are done and it can be trimmed to point here), #14 (the host-side harness these host-binding checks belong in), #7 (an anchored-query probe would ride the same fixture work).

---

Status (2026-07-26): four of the original items are done, four survive.

PR #60 merged four hours after this issue was filed and closed everything the old title advertised — representation witnesses (caller), the as_of probe, the host-side max_frames audit, and the E1 fingerprint. The title was left describing only finished work; it has been rewritten to name the survivors instead.

Survivors:

1. granted_at stamping — ConsentRecord.granted_at is still never set.
2. The kinds-filter probe — sample_query() sends empty kinds, so filter honoring is unverified.
3. A live representations witness — the invariants have a caller but nothing to catch: no representation-advertising fixture, no representation-lying --misbehave mode.
4. The red script's expected-check mapping — a misbehave mode caught by the wrong check still passes the red gate. (tests/docs_site_witness_test.py is still unwired in CI, in the same bullet.)

[macanderson]

Re-verified 2026-07-26 against origin/main @ 1ee9b1c.

This issue was overtaken by PR #60, which merged roughly four hours after it was filed. Most of the body describes a gap that closed almost immediately.

What landed: four of the seven items, all via PR #60 —

• representation invariants finally have a caller (they were dead code before),
• as_of is probed,
• host-side max_frames is audited,
• E1 has a fingerprint witness.

What survives: four items —

1. granted_at is still unstamped. This is inherited from Close enforcement gaps: audit max_frames, validate temporal fields as RFC 3339, probe as_of #10, now closed as superseded by this issue, so P1: Enforcement residue: granted_at stamping, kinds-filter probe, representations witness, red-script check mapping #51 is its only tracker.
2. The kinds filter is unprobed.
3. There is no live representations witness.
4. The red script has no expected-check mapping, so a red run's failures are not tied back to the checks they are meant to trip.

Recommended scope: rewrite the body down to those four survivors, and strike anything describing the post-#41 enforcement gap as open — it is not.