updates on the order bulk fix

docs/balanced-assignment-plan.md

@@ -1,6 +1,24 @@
 # Balanced auto-assignment: implementation plan
 
-**Status:** planned, not started (2026-10-06).
+**Status:** Phases A and B built and tested, not committed or deployed (2026-10-06). Phase C is a server setting (`MILER_MAX_ACTIVE_BOOKINGS=20`).
+
+### Phase B: what was built
+- `internal/assignment/release.go` (new):
+  - **`releaseUnaccepted`** (runs at the start of every sweep): an assignment still `Assigned` on an order still `Miler_Assigned`, older than `ASSIGNMENT_ACCEPT_TIMEOUT_MINUTES` (default 10; 0 = off), goes back to the pool. The order returns to `Pending_Pickup` with no rider, and the assignment becomes `Reassigned` with a note. The customer stage is walked back (`cxstage.Release`), and the rider is set to Available if they hold nothing else. Every write is conditional, so a rider accepting at the same moment wins.
+  - **Only auto-assignments are released.** They're now labelled `Remarks = "Auto-assigned"` (`autoAssignedRemark`). A rider hand-picked by ops or hub staff is never taken away. (The hub console's manual assign records no "assigned by", so the label is the only reliable marker.) Assignments made before this change have no label and are never released.
+  - **`ridersToSkip` / `withoutRiders`** (applied in `selectMilerWithAI`): an order is never offered back to a rider it was released from, who rejected it, or who cancelled it.
+  - **`SweepNear`**: called from `MilerStartDuty` in a goroutine. It assigns the pending orders within 10 km of a rider who just started duty, instead of waiting up to 5 minutes for the next sweep.
+- `pendingForSweep` now also loads the pickup coordinates (needed by `pendingNear`).
+- Tests: `release_test.go` (settings, skip filter, distance) and `release_pg_test.go` (release after timeout; inside the timeout, accepted and manual orders never released; the rider freed when nothing else is held; a concurrent accept wins; a released or rejected order goes to another rider; the duty-start sweep only picks nearby orders).
+- End to end on the real backend: a waiting order was assigned **0.19 s** after the rider started duty. An unaccepted order was released after the timeout and given to the other rider in the same sweep. Accepted and manual orders were untouched after 3 sweeps. Bulk split was still 3/3.
+
+### Phase A: what was built
+- **In hand** = `openStopsToday` (today's open, non-cancelled orders), the same count the ceiling uses. Using all open records ever would let stale, abandoned orders skew the balance.
+- `selectMilerWithAI` keeps only the least-loaded riders (`leastLoaded`), then the AI or the fallback chooses among them. It now also returns the full pool.
+- `pickBestFromCandidates` → `betterChoice`: in hand ↑, then session ↑ (`sessionStops`, counted since the open `milerdutylogs.loginat`), then distance ↑, then rating ↓.
+- `finalizeChoice` (both commit paths): takes `pg_advisory_xact_lock(7270001)`, recounts the pool, keeps the original choice if it's still least loaded, otherwise picks the best of the least loaded. Returns `errNoRiderCapacity` if everyone is at the ceiling (the attempt is retried later). The AI decision id is dropped when the choice changes.
+- Tests: `balance_test.go` (ranking) and `balance_pg_test.go` (50 over 5 → 10 each; 23 → 5/5/5/4/4; **50 concurrent → within 1**; newcomers catch up; tie-break; all at the ceiling → waits).
+- End to end on the real backend: bulk upload of 9 orders with riders at 0, 2 and 5 km → **3/3/3**. A 4th rider then joins and 4 more orders arrive: the newcomer gets 3, the 4th goes to the nearest (all tied).
 **Goal:** however many orders and riders there are, split the orders **equally** among the riders, with riders logging in and out all day.
 
 ---