core/docs/handoff/wave-R3-briefs-2026-06-13.md

Wave R3 Briefs — Theater (the wow, honest)

Date: 2026-06-13 Plan: docs/workbench/wave-R-mastery-revamp.md § Wave R3. Predecessor: Wave R2 complete (#714–#721; all six routes real). Replay backend already merged (#716, GET /replay/{turn_id} → TurnReplayComparison).

Four pieces. The Replay Moment is first (its backend exists and the old artifact-keyed ReplayRoute is now stale against it). The others are independent and can be authored after.

Standing constraints (all briefs)

Same as Wave R2: worktree off fresh origin/main; green-local before push (pnpm build && pnpm test); STOP after checks green, Shay merges; token-only styling (hexScan); shrink-only NOT_YET_MIRRORED; conformance rows (ADR-0162 §6); no new mutation endpoints; no invented data; no force-directed / nondeterministic layout — ever.

---

Brief R3-Replay — The Replay Moment (first; reworks + retires)

The hero flow. Turn-keyed sealed replay made felt: same prompt, same genesis substrate → bit-identical envelope, rendered as hash≡hash with an honest leaf diff. Consumes the merged #716 backend.

This PR also retires the dead W-026 artifact-keyed replay machinery (confirmed zero serving uses) on both sides — the NOT_YET_MIRRORED comment already anticipates it.

Retirement surface (verified)

• Python (workbench/schemas.py): delete ReplayDivergenceSeverity, ReplayStatus, ReplayDivergence, ReplayComparison (the artifact-keyed block). Keep the TurnReplay* block. Zero word-boundary uses elsewhere.
• Snapshots: regenerate both — pnpm schema:snapshot (drops ReplayComparison/ReplayDivergence) and pnpm enum:snapshot (drops ReplayStatus/ReplayDivergenceSeverity).
• TS (types/api.ts): remove ReplayComparison/ReplayDivergence/ ReplayDivergenceSeverity; add TurnReplayComparison/TurnReplayDivergence (+ the TurnReplayDivergenceSeverity/TurnReplayBasis/TurnReplayOriginState unions). Remove both TurnReplay* entries from NOT_YET_MIRRORED.
• Badges (design/components/badges/{types,mappings,index}.ts*): remove ReplayStatusBadge + ReplayDivergenceSeverityBadge + their meta/types; drop their two cases from enumCoverage.test.ts. The hero renders the new severity (critical/informational) inline with a ≠ glyph — no new enum-tracked badge.
• API (client.ts/queries.ts): remove fetchReplayComparison/ useReplayComparison; add fetchTurnReplay(turnId) → /replay/<turnId> and useTurnReplay(turnId).
• Route dir (app/replay/): delete ArtifactList, ReplayComparisonPanel, ReplayDiffViewer, ReplayMetadataTable, and the old replay.test.tsx.

New ReplayRoute (turn-keyed hero)

• List journaled turns (reuse useTraceTurns) in the left pane (VirtualizedList
  • useListNavigation + SearchInput), same as Trace.
• Select → navigate /replay/<turnId> (replace) → useTurnReplay(turnId).
• Hero: original_trace_hash vs replay_trace_hash rendered big; a clear ≡ equivalent (when equivalent) or ≠ diverged verdict. Honesty card states comparison_basis (sealed_fresh_runtime_single_turn) and origin_state (unrecorded) — a divergence means nondeterminism OR origin-state influence, never disambiguated; never render divergence as a determinism-failure verdict.
• Leaf diff: each divergence as a row with path, original vs replay, and a ≠ glyph; critical weighted above informational; informational divergences (timestamp/cost/digest) explicitly labeled as expected.
• Publishes the turn subject for the inspector (identity → detail), like Trace.
• App route path="replay/:turnId?"; conformance row turn-keyed (loading "Loading turns...", empty "No turns recorded yet. Use Chat to create evidence." + core chat).
• Tests: hero equivalence (hash≡hash), a tampered-leaf divergence renders ≠ at the right path, informational-only divergence still reads equivalent, the honesty fields render, j/k spine, replace-mode URL selection.

Verification

cd workbench-ui && pnpm build && pnpm test
# plus the Python lane the snapshots feed:
uv run python scripts/dump-schemas.py | diff - workbench-ui/schema-snapshot.json
uv run python scripts/dump-enums.py | diff - workbench-ui/enum-snapshot.json

---

Brief R3-DAG — Deterministic DAG viewer

One hand-rolled component (design/components/Dag/): longest-path layering, lexicographic tie-break, ~150 lines, golden-file layout tests, no graph dependency (force-directed = doctrine violation). Pan/zoom/click-node → inspector. Consumers: proposal chains, the 8 PCCP proof-promotion scenarios (demos/proof_carrying_promotion), entailment traces. Pure layout function; golden tests pin node coordinates.

---

Brief R3-Demo — Demo Theater route

GET /demos, POST /demos/{id}/run (new read + a scoped run endpoint). Scenario results with evidence-class badges and "what this proves / what this does not prove" honesty cards; proposer-was-wrong scenarios visually highlighted. Backend brief first (Python), then the route.

---

Brief R3-Ledger — wrong=0 ledger view

Evals renders the correct/refused/wrong triplet as the PRIMARY visualization; refusal reasons inspectable; failures-first ordering. Additive to the existing EvalsRoute — no retirement, lowest risk of the four.