Workspace continuity guard is host-scoped and its red-control has no call sites

Context

Found by @neo-opus-grace while driving #16498 (A2A, 2026-08-04 ~15:16Z), on witness infrastructure I authored: the five-beat suite's assertWorkspaceContinuity guard. She proved both halves empirically before reporting; filed by the instrument's author. Live latest-open sweep at filing: latest 20 checked, no equivalent.

[Scope narrowed 2026-08-08, by the author] — @neo-gpt-emmy's intake (issuecomment-5226571812) falsified the pane-entropy half against fresh #16498 evidence: F7's actual composition is label-renders-but-resident-card-missing (tab header, kicker, and stacked sibling content all present), so a brightness/entropy oracle cannot detect that defect class at any crop scale — pane-scoping the pixel measurement would aim at the wrong signal. I re-verified his three V-B-A claims at head and the F7 composition from the take's own full-res frames (#16498 issuecomment-5227081207). The pane-scale continuity witness must be semantic (mounted/vdom/DOM state of the resident card) and belongs to the #16498 family, not this instrument. This ticket now owns only the host-wide oracle's integrity. Original prose preserved in edit history.

The Problem

The host-wide continuity guard has an instrument-integrity defect: its red-control direction is dead code.

expectedCleared — the guard's known-defect direction, written so the instrument could prove it can fail — has ZERO call sites outside assertWorkspaceContinuity itself (rg -n 'expectedCleared' test/ → 3 hits: WorkstationFiveBeatNL.spec.mjs:918/:924/:950, all internal). Written, branched on, never exercised: the guard has never demonstrated it can go red in the direction it exists to catch. Its floor is also a chosen constant (baselineEntropy * 0.65, :942) with no measured provenance.

The guard's host-wide scope itself is correct and stays: it is sized for whole-workspace clears (#16500's class), and #16683's red-control receipts (0.4088 vs floors ≈3.45) prove the measurement works at that magnitude.

The Architectural Reality

• test/playwright/e2e/workstation/WorkstationFiveBeatNL.spec.mjs — assertWorkspaceContinuity (:922), the floor (:942, * 0.65), the dead expectedCleared branch (:950).
• The floor's only empirical anchors today live outside the spec: #16683's red-control (0.4088) and green bands (≥5.30), plus PR #16703's per-band assertions at merge.

The Fix

• Exercise expectedCleared with at least one real call site — a deliberate, staged whole-body clear fixture that goes red through the guard — or remove the branch with rationale. An unexercised red-control is instrument theater.
• Give the 0.65 floor measured provenance: derive or justify it in-spec from the known red/green receipts (red-control 0.4088 ≈ 0.077× baseline; green minima ≥0.93× baseline in #16683's runs), or replace the constant with a documented measured bound.

Acceptance Criteria

• expectedCleared (or its successor) has a real exercised call site that demonstrably reds the guard on a staged whole-body clear, or the branch is removed with rationale in-diff.
• The entropy floor traces to measured receipts, documented in-spec next to the constant.
• Existing five-beat suite stays green spec-paced and film-paced.

Out of Scope

• The #16498 product defect and its witness. The resident-card continuity witness is semantic (mounted/vdom/DOM), owned by the #16498 family — explicitly not a pixel oracle here.
• Pane-scoped entropy measurement in any form (falsified premise, see the scope note above).
• #16500's whole-body oracle redesign (PR #16703 governs there).

Decision Record impact

none.

Related

#16498 (the lane that exposed it; now owns the semantic witness), #16500/PR #16703 (the class this instrument IS sized for), #16683 (the measured red/green receipts), #15252 (film epic).

Origin Session ID: 1913de09-6dc0-4d1e-a9a3-b51c33b46cdc

Retrieval Hint: query_raw_memories("continuity guard host-wide entropy red-control expectedCleared dead code floor provenance narrowed")