Three workstation preview-surface e2e witnesses fail at dev head

Context

Found during the R2 readiness audit for #15252 (the film's case→capability→witness matrix at head). Three workstation e2e witnesses fail deterministically on the film host at dev head 5703e0d18f — all three pre-existing (falsifier below), none caused by the L2/L3 merges (#16319 / #16332 / #16335).

The Problem

Suite run at head (NEO_E2E_PORT=8117 npx playwright test e2e/workstation/ -c test/playwright/playwright.config.e2e.mjs --workers=1): 15 passed, 3 failed, 2 skipped. Each failure reproduces in isolation (suite-contention ruled out) and at a worktree pinned to 09f71b186d~1 (pre-#16332 — the merge-time window ruled out):

1. WorkstationNL.spec.mjs:482 ("the real tour keeps density, data, Canvas output, themes, rails, and identities live") — the cross-zone showcase beat errors with indicator 'preview:audit:scale-tabs:edge-bottom' never became active (crossZoneReceipt.errors expected []). The edge-bottom preview indicator never activates during the visible tour's final live-candidate commit.
2. WorkstationDragAffordancesNL.spec.mjs:388 ("same-gesture tear-out re-entry resumes proxy motion without reacquisition") — 1 failed / 5 passed in isolation.
3. WorkstationHumanPopupOverlapNL.spec.mjs:597 ("large-over-small: one local proxy plus readable target choices") — 1 failed in isolation.

The shared theme is preview/proxy activation timing across three surfaces; whether they share one root is the bisect's question, not an assumption — the receipts are bundled because the film matrix treats them as one readiness cell (preview-surface truth on the film host).

Falsifier run (pre-existing proof): identical failures at 09f71b186d~1 in an isolated worktree — WorkstationNL 1 failed, DragAffordances 1 failed / 5 passed, HumanPopupOverlap 1 failed.

The Architectural Reality

• All three surfaces ride the dockPreview / vessel-proxy machinery the film's drag choreography depends on — a take over a red witness is not a take (spec = authority).
• The FiveBeat spec itself is fully green at the same head (8/8 + 2 contracted skips) — the failures live in the surrounding workstation witnesses, not the five-beat spine.
• The film host is headed+GPU; these runs are this host's ground truth for readiness (per the film's evidence contract).

The Fix

Diagnose and repair the three witnesses: bisect the activation-timing window for the edge-bottom indicator first (its error names the exact predicate), then the two proxy-motion/overlap cases; split into per-root tickets if the bisect shows independent causes.

Acceptance Criteria

• WorkstationNL.spec.mjs green at head in isolation (the edge-bottom indicator activates during the cross-zone showcase beat)
• WorkstationDragAffordancesNL.spec.mjs green at head in isolation
• WorkstationHumanPopupOverlapNL.spec.mjs green at head in isolation
• Full e2e/workstation/ suite green at head on the film host
• Regression window named (first red commit per root cause, or "predates the bisect floor" with the floor SHA)

Out of Scope

• The L2/L3 transaction/perspective machinery (exonerated by the pre-merge worktree falsifier)
• The contracted FiveBeat skips (activated by their own gates: NEO_FILM_TAKE, cross-window wiring legs)
• The R2 witness matrix itself (#15252 comment)

Related

• #15252 (R2 readiness lane — these cells are NOT provable today until repaired)
• #16309 / PR #16321 (the prior stack-return preview repair — adjacent preview surface, already green)

Live latest-open sweep: checked latest 20 open issues at 2026-08-02T11:3xZ; no equivalent found. gh search on the indicator text and "preview indicator" returned epics only. A2A in-flight sweep: no claim on these defects in the last 60 min.

Origin Session ID: 1a7e3f91-8356-48bb-a353-9fd7da2647f5

Retrieval Hint: "workstation preview indicator never became active e2e film host"