STATE: harness armed + closing capture started; reverted my own non-close writes

specs/STATE.md

@@ -133,6 +133,48 @@ exist locally; harmless, the data restores completely.
 
 ---
 
+## 🟢 HARNESS ARMED + CLOSING CAPTURE STARTED (2026-07-20 ~16:10 UTC)
+Both verified by INDUCING the real code path — never by waiting for a slot.
+
+**A — Backtest harness runs on OUR scheduler.** `snapshotScheduler` fires it at
+`HARNESS_HOUR_UTC` (default 14), appends to `harness_results`, and
+`opsWatch.harnessStaleAlarm` pages if it stops (a validator that isn't running
+looks exactly like one that keeps passing). No external dependency.
+**INDUCED live:** `verdict INSUFFICIENT_HISTORY, scored 35, joined 283,
+snapshots 478` — the correct, passing output. `POST /api/internal/harness/run`.
+
+**B — Closing capture (CAPTURE ONLY; no CLV metric, as ordered).**
+- **C4 diagnosed:** `closing_line` is ONE overwritable field with no timestamp
+  and no provenance. `captureClosing` writes the current line and, on a prop
+  match failure, silently leaves the earlier value (= the lock) in place — so
+  "captured a real close" is indistinguishable from "never updated". It is
+  **92 % equal, not 100 %** (56 rows DID move): the defect is **provenance**.
+- **`closing_captures` (migration 029)** — append-only, never overwritten, BOTH
+  raw side prices retained so the existing de-vig engine computes the fair
+  closing probability later. Sharp (`pinnacle`) vs book close kept distinct:
+  "beat the market" and "beat our book" are different questions.
+- **JOIN drops `line` on purpose** — a close that MOVED off the graded line is
+  the whole point of CLV. Verified safe: all 164 identity groups have exactly
+  ONE line per `(sport, player_key, stat, side, game_date)`; zero ambiguity.
+- **Wired into the intraday refresh** (already polls ~20 min through slate
+  hours) with a capture-rate alarm — a missed close is unrecoverable.
+- **🔴 ODAWA IS NOT CAPTURABLE.** `odawa` exists only as a UI preference option
+  in onboarding/settings — no adapter, no `ALLOWED_BOOKS` entry, no feed. CLV
+  vs our own book cannot be measured until a real Odawa feed exists.
+- **Refusals proven first (test-first):** unbound game_time, doubleheader
+  ambiguity, missed window, or a one-sided price → `missed_reason` and **NO
+  price**. Induced live: 1,076 rows correctly refused `one_sided_price`.
+
+### ⚠️ MY OWN ERROR, CAUGHT AND REVERTED
+To force a capture I widened the window to 1,200 min and **persisted 7,114
+rows** — every one had a lead time of **392–602 minutes**, i.e. 6.5–10 hours
+before lock. **Those were not closes.** Leaving them in a table named
+`closing_captures` would have poisoned the first CLV computation with pseudo-
+closes. **All 7,114 deleted; table verified back to 0.** Verification is now
+done with `?dry=1`, which proves the pricing path (wnba: 486 rows priced, 0
+missed) while writing nothing. The real 45-min window currently yields 0 rows
+because no game is near lock — which is the correct answer, not a failure.
+
 ## 🟢 HEAL EXECUTED (Order 2, 2026-07-20 ~10:00 UTC) — record is clean
 Rollback point first: `vyndr-20260720-093821.dump` (856,890 B) **verified ON the
 Storage Box**, not just exit 0.