sentiment-engine/prod/docs/POST_FINISH.md

# POST_FINISH — the parking lot (parked with love, not killed)

**Standing rule (operator + Fable, 2026-07-10):** until the FINISH LINE is
crossed, new scope lands HERE, not in code. Fable answers new scope with
"parked" — even to the operator. Especially to the operator.

## THE FINISH LINE (frozen — only these three count)
1. **Parity RESIDUAL = 0** over N windows (codex dossier machinery,
   `uv/parity-certification`)
2. **72h clean on VST** on the merged stack (`a2b0fac`+): no §4 anomalies —
   no orphan orders, no journal gaps, tripwire quiet
3. **Operator gate:** small real capital, separate account, BLUE untouched

## SEQUENCED (above parked — operator 2026-07-11: "it is costing us money")
- **Smart OB exec: PINK ExecutionRouter integration** — FIRST task upon parity
  (item 1) completion, before/alongside items 2–3. Asset exists certified
  (265 tests incl. live VST E2E, maker_both ran production). Slots below the
  kernel at the venue-adapter seam — parity-invisible. Economics: taker RT
  ~0.10% → maker-both ~0.04%; smoke-run day-1 flips breakeven → +0.9%
  equal-weight. Research: 90%+ of orders fill as maker ("move to us").
  **CRITICAL READING:** `prod/docs/BingX_FILL_CHARACTERIZATION_AND_ADVANTAGES.md`.
  **SPEC WRITTEN (2026-07-13):** `prod/docs/SPEC_UV_SMART_EXEC_MM.md` — urgency
  ladder, contract, rollout; adjudication is its sub-task 1.
  **SOA ADJUDICATION (first sub-task, Fable):** TWO exec assets exist —
  (a) PINK ExecutionRouter: maker/taker POLICY, 265 tests, maker_both ran live;
  (b) BLUE `alpha_engine/execution/` **SmartPlacer**: OB-aware PLACEMENT
  (fill_simulator w/ 30s bookDepth, signal_confidence_adapter bridging Alpha
  conviction → placement, ob_reader, smartplacer_constants). Likely
  complementary (Router = maker-vs-taker; SmartPlacer = where-in-book).
  Adjudicate router ∪ placer ∪ union against the fill-characterization doc
  before integrating either.

## PARKED (specced where noted; zero build until finish)
- **IMPETUS build** — spec `IMPETUS_SPEC.md` + annex
  `IMPETUS_SPEC_ANNEX_UV_INTEGRATION.md` (steel-clock plug, ASEx placement,
  last-tick table). Paper done; build parked.
- **MALKHUT hose adaptation** — Fable's job when unparked (memory:
  malkhut-hose-adaptation-fable)
- **JIMINY / PINOCCHIO** — ASEx fleet scoreboard + false-green detector
- **TP/SL cadence ladder L2/L3** — pulse-host tick path, Rust reflex engine
  (spec `UV_TPSL_ENGINE_AND_REMAINING_DEV_SPEC.md` §1)
- **305M ring upgrade** — producer-side cached index in `uv_steel::ring`
  (PASS9, ~10 lines + bench)
- **wCQ wait-free MPMC** — L3 mailbox candidate (`305MM_OPS_ALGO_ANNEX_FIFO.md`)
- **★ IMPETUS WIRE-ATOMIC (operator 2026-07-11, "insane, write it down")** —
  exchange-wire→zinc-shm ATOMIC: the venue packet's tick becomes visible to
  shm readers as ONE indivisible act, no intermediating copies/queues/threads.
  Concretely: kernel-bypass ingest (AF_XDP / DPDK / io_uring registered
  buffers — NIC DMAs frames into user memory), parse-in-place off the wire
  buffer, single seqlock commit into `impetus_last` as the SAME pinned
  thread's next instruction. Latency floor becomes NIC+parse (~µs), and the
  atomicity contract extends end-to-end: a reader can never observe a tick
  the wire didn't deliver, nor miss one it did (A2/A5 doctrine stretched to
  the physical boundary). Prereqs: IMPETUS v1 shipped; WS/stream adapters
  (annex R2); pinned-emit thread (annex R4); flat-C last-tick table (R3) —
  each is a rung TOWARD this. TLS termination path (operator follow-up: the
  sought term was "homomorphic" — FHE itself is 1e6-1e9x too slow, wrong
  universe): SOA answers preserving wire-atomicity, in preference order:
  (1) NIC inline TLS RX offload (ConnectX/kTLS-offload — silicon decrypts
  AES-GCM at wire speed BEFORE DMA; buffer lands already-plaintext; CPU never
  sees ciphertext); (2) kernel kTLS (decrypt in-kernel, zero-copy up);
  (3) VAES/AES-NI decrypt-IN-PLACE (plaintext overwrites ciphertext in the
  same buffer, ~1-4 cycles/byte — "decoded where it lies").
- **`uv_steel` crate extraction** — pulse-host modules → shared workspace crate
- **PASS9 mass-test regeneration** — far-end assertions, no broken codegen
- **BLUE trade_events dedup DDL** — operator-blessed write into dolphin ns
- **Vendor reconciliation** — T0-DEVIATION port upstream + vendor_sync
  --from-committed-HEAD hardening
- **UvExecRunner consolidation** — route legacy direct async path through the
  ASEx owner or retire it (execution-gate dossier §6.6)
- **Journal EXIT-labeling cleanup** — no-entry scans mislabeled EXIT (cosmetic)
- **Fleet identity hygiene** — stale zellij names, fossil handles, split bus

- **★ SHARP TODO (operator 2026-07-11): 375-branch upside study** — realized
  upside under favorable-extension vs realized loss; left-tail cure hypothesis
  (losses huge incl. scan-cadence SL misses; winners cramped). Depends on:
  MAX_HOLD full-modulation port + ingress journaling
  (UV_EXIT_DOCTRINE_PARITY_DESIGN.md D1.2). OB-cascade cause tuning rides the
  same data.

- **★ RESEARCH (operator 2026-07-13): dvol quanta sweep × market regime** —
  sweep the vol-gate threshold across its entire range; per quantum measure
  trade result / ROI / DD / all metrics; then the same MODULO market regime
  (MARAS-or-whatever fingerprinting — regime tagging itself "more or less
  solved but not conclusively"). PREREQ (small, cheap, do before the sweep):
  persist the per-trade dvol value AT ENTRY (not just gate pass/fail) — same
  field the DDSketch self-tuning gate needs. Existing single data point:
  EXP_VOL_GATE_RELAX 2026-07-04 (relaxed −17,847 / 189 trades vs normal
  +3.54 avg / 7,883) + F5 relax window 2026-07-12/13 (4 ADVSL losers,
  MFE ≤ +0.013%).
- **★ RESEARCH (operator 2026-07-13): scan-cadence effects study (11.06s UV
  vs 5.97s NG7-pinned BLUE)** — cadence effects on EVERYTHING: ROI, SL/TP
  achieved gain/loss, opportunity cost. Includes the bar-denomination
  question: are bar-constants (ADVSL min_hold=20, MAX_HOLD 100/125/250/375)
  hold-TIME (wall-clock, convert via avg scan duration) or evidence-COUNT
  (N OBF observations)? Empirical, decides whether to time-normalize or
  cadence-match. Related: TP-overshoot mechanism — BLUE effective TP ~0.3
  vs 0.20 configured: bar-cadence miss (beneficial upswing capture) vs
  modulation×1.40; measure exit_pct − tp_effective_pct on BLUE TP exits.
  UV at-tick TP_FLOOR (arm at .20 sharp, exit on give-back) is the operator's
  current best expression; keep unless study finds better.
- **★ RESEARCH (operator 2026-07-13): win/loss streak wave-phase alignment** —
  treat win-loss/loss-win STREAKS as a wave (profit/loss as amplitude);
  test PHASE match (in-phase / out-of-phase) against vel_div regimes,
  quantized to the 11s/6s scan cadence. If found and stat-significant,
  operator conviction: these waves are the unlock for the SHORT/LONG
  regime-switch grail.

## Not parked (allowed work)
Anything that directly serves items 1–3, bug fixes on the live path, and
operator-ordered exceptions (which must say "unpark X" explicitly).

## DESIRABLE (north-star, operator 2026-07-11 — not parked, no build yet)
- **Full-universe picking**: system should someday select from the ENTIRE
  Binance (~500-asset, as minimum reference) universe, not the 50-symbol
  NG7 feed slice — research indicates full-universe selection improves
  ROI/WR/joy. MALKHUT asset directory (malkhut/assets/) is built to that
  scale; the feed is the current bottleneck, not the store.