diff --git a/AGENTS.md b/AGENTS.md
index 3e3b60c..ac212eb 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -164,7 +164,8 @@ Shipping simulation (`src/collections/logistics/Shipments.ts`, `src/scripts/`, `
 - **Base tick**: `bun run payload base-tick` — bin registered like the others (needs external cron). Runs `processBaseTick` (`src/lib/base/tick.ts`): pays staff salaries from the structure treasury (banking `applyTransaction` type `salary`), charges structure maintenance, flags upkeep shortages (see Base Management below). Emits `staff:salary-paid`/`staff:salary-unpaid`/`structure:maintenance-paid`/`structure:maintenance-unpaid` events. Notifies clients via the same SSE path as the other ticks (`source: "base-tick"`).
 - **Economy tick**: `bun run payload economy-tick` — bin registered like the others (needs external cron; cadence hint is GameRules `economy.tickIntervalHintMinutes`). Runs `processEconomyTick` (`src/lib/economy/tick.ts`): ensures a `market-state` ledger doc exists per tradeable resource/asset (copied from GM baselines), decays `realDemand`, drifts `artificialDemand` (deterministic per period via `periodSeed` → mulberry32 rng — a cron retry within the same period recomputes the same values), recomputes `priceModifier` from the demand/supply ratio clamped to the item's band. Master gate: GameRules `economy.enabled` (default **off** — no-op when disabled). Cold-start gate: items with fewer than `coldStartObservations` completed purchases keep `priceModifier = 1`. Period idempotency: docs with `lastComputedAt >= periodStart` are skipped. Emits a batch `economy:tick` event + one `economy:price-change` event per item whose modifier moved ≥ `PRICE_CHANGE_THRESHOLD` (0.05) — target `market-state` (logic: `src/lib/economy/` incl. `events.ts`, tests: `tests/int/economy.int.spec.ts` + `tests/int/economy-events.int.spec.ts`). Nothing consumes the modifier yet (P1d: market-tick new-listing pricing). Notifies clients via the same SSE path as the other ticks (`source: "economy-tick"`).
 - **Economy consumption (P1d)**: market-tick prices new NPC listings from the ledger — `price = max(1, round(baseBuyPrice × priceModifier))` (neutral 1 when no market-state doc) then `applyNpcPriceModifier` (vendor's own modifier); the old `autoPrice` ±15% randomness is **removed**. Restock quantity scales down under shortage via `scaleQuantityForShortage` (bounded to a 0.25 factor floor, never below 1). `buyListing` ingests real demand: purchases of NPC/vendor stock (`seller == null`) call `recordPurchase` (fire-and-forget, `src/lib/economy/consumption.ts` — quantity-weighted `realDemand` increment). GM tools: `resetEconomyState` server action (permission `market-state:update`, rebuilds the doc from baselines, emits `economy:reset`; per-listing "Reset economy" button gated by `canResetEconomy` on the market page) and `getReferencePrices` (modifier-applied reference price shown in `CreateListingDialog`'s tip when it differs from the resting valuation). Tests: `tests/int/economy-consumption.int.spec.ts`. Notifies clients via the same SSE path as the other ticks (`source: "economy-tick"`).
-- **Mission auto-completion**: `bun run payload mission-tick` — bin registered like the others (needs external cron). Sweeps missions whose status is `Scheduled`/`Active` and whose scheduled date (`classification.startDateTime`) has fully passed — from the start of the following **server-local** day — and sets them to `Completed`, emitting a `mission:auto-complete` event per mission (logic: `src/lib/intelligence/missionLifecycle.ts`, test: `tests/int/mission-lifecycle.int.spec.ts`). Draft statuses (Concept/Planning/Ready) and terminal statuses (Completed/Cancelled) are never touched; idempotent. Notifies clients via the same SSE path as the other ticks.
+- **Mission lifecycle auto-start / auto-completion**: `bun run payload mission-tick` — bin registered like the others (needs external cron). Each tick runs two sweeps (start first, then complete) in `src/lib/intelligence/missionLifecycle.ts`: `autoStartMissions` flips `Scheduled` missions whose `classification.startDateTime` has arrived to `Active` (emitting `mission:auto-start`), skipping missions whose whole window already elapsed so they don't get a spurious start event; `autoCompleteMissions` then sets missions whose scheduled window (`startDateTime` + `estimatedDuration`, default 240 min when missing/invalid) has fully passed to `Completed` with `completedAt` stamped at the sweep time (emitting `mission:auto-complete`). Draft statuses (Concept/Planning/Ready) and terminal statuses (Completed/Cancelled) are never auto-started; only `Scheduled`/`Active` are auto-completed. Both sweeps are idempotent. Notifies clients via the same SSE path as the other ticks (tests: `tests/int/mission-lifecycle.int.spec.ts`).
+- **Side-op workflow (Ready draft → Scheduled committed)**: `createFridaySideOp` creates a side op as a quiet `Ready` draft (no roll-call). The creator commits it with `markSideOpScheduled` (the "Mark as Scheduled" control on the mission page), which flips it to `Scheduled` and emits `mission:commit`. Only `Scheduled` ops post a Discord roll-call (about 3 days out) and are driven by the auto-start/auto-complete lifecycle above. The daily reminder (`readyReminders`, web notification + Discord DM, notification type `mission:ready-reminder`, label "Schedule reminder") nudges the creator to set the op to `Scheduled` while it is still Concept/Planning/Ready; the internal "ready" names are legacy and kept to avoid a migration.
 - **Server presence**: `bun run payload server-tick` — bin registered like the others (needs external cron, every minute). Flips `game-servers` docs that claim `status: "online"` but whose last heartbeat (`lastSeenAt`) is older than 180s to `offline`, including online docs with no `lastSeenAt` at all; emits a `server:offline` event per flipped server (logic: `src/lib/arma-bridge/presence.ts`, test: `tests/int/server-presence.int.spec.ts`). Idempotent. Notifies clients via the same SSE path as the other ticks.
 - **Arrival handling** (`src/scripts/processShipmentTick.ts`): destination storage rules are re-checked on arrival; rejected cargo bounces back to origin, shipment goes `failed`, `ShipmentFail` event logged.
 - **GameRules tuning**: `proximityThreshold` and `gameTickIntervalMinutes` live on the global `game-rules` doc.
@@ -310,10 +311,10 @@ Leaflet world map at `/map` (top-level sidebar entry, whole unit views it). Full
 - **Schema**: `tech-categories` collection (`src/collections/intelligence/TechCategories.ts`: `name` unique, `position` number = timeline order, optional hex `color`) + `Technologies.category` relationship (FK `ON DELETE set null`, so deleting a category just unfiles its techs) + `Technologies.treePosition` group (`x`/`y` numbers; manual canvas position from dragging, empty = automatic). Permissions mirror technologies (`tech-categories:*` + `admin:tech-categories:manage`, registered in `src/permissions/index.ts`); `tech-categories` is in `DIVISION_ADMIN_QUALIFICATIONS`, so Intelligence division members get admin-panel access exactly like `technologies`. Migrations: `20260930_230747_add_tech_categories`, `20261001_000320_add_tech_tree_position`.
 - **View wiring**: `payload.config.ts` `admin.components.views.techTree` (path `/tech-tree`, component `src/components/admin/tech-tree/TechTreeView.tsx`) and `admin.components.Nav` → `TechTreeNav.tsx`. Custom root views are public by default, so the view gates itself (logged-in + `technologies:read` OR Intelligence qualification). The nav replica injects the entry into the group containing `technologies` and reuses public exports (`NavWrapper`/`NavHamburger`/`DefaultNavClient` from `@payloadcms/next/client`, `groupNavItems` from `@payloadcms/ui/shared`); registered admin components must be **default exports** (importMap convention in this repo).
 - **Layout** (`layout.ts`, pure): a **propagating left-to-right tree**. A technology WITH prerequisites is placed one column to the right of its parents, at their average row, so linking A -> B drops B immediately beside A and B's dependents continue further right; nothing is teleported to a distant column when its depth changes. Technologies in a category stay pinned to that category's column (the category contract); uncategorized roots fill a catalog grid whose slots come from each tech's rank among ALL uncategorized technologies (`MAX_ROWS_PER_COLUMN` = 10 rows, spilling into more columns) so promoting a tech into the tree leaves an empty slot rather than re-shuffling the grid; uncategorized dependents follow their parents inside the uncategorized zone. Collisions search rows outward from the desired row (dependents may also grow upward, and their columns are not height-capped, so a link from a full column still lands beside its parent rather than far right). Category bands are the bounding box of their members' final positions. Prerequisites placed to the RIGHT of their dependent surface as amber warnings (toolbar chip + node icon + preview panel section).
-- **Edges**: AND prerequisites solid slate, OR prerequisites dashed cyan with an OR label. Connections are editable on the canvas: drag from a node's right handle onto another node (the target card OR its handle) to add it as an AND prerequisite; the drop-on-card case is resolved by an `onConnectEnd` fallback because React Flow only resolves drops onto a handle. Select an edge and press Backspace to remove the prerequisite. Duplicates, self-references, and cycles (direct or transitive) are rejected by the pure `prereqEdits.ts` helpers with an inline error in the toolbar.
-- **Interactions**: nodes are draggable; dropping one PATCHes its `treePosition` (persisted, survives reload) and the toolbar shows an "Auto-arrange" button (technologies:update gated) that clears every manual position and restores the ordered layout. A **Save positions** button (technologies:update gated), next to Auto-arrange, commits the current on-screen position of every node to `treePosition` in one pass (matching what React Flow rendered, not just dragged/pinned cards) so the exact arrangement survives a reload; Auto-arrange then clears them. **Connecting does not move anything**: if the ordered layout would relocate the connected technology, the canvas pins it to its current spot (`treePosition` written in the same PATCH as the prerequisite) so the edge appears without the card jumping; the catalog grid also uses stable slots (each tech's rank among all uncategorized technologies, so a tech promoted into the tree leaves its slot empty instead of re-shuffling every other root). Pressing "Auto-arrange" clears those pins and moves everything, including the connected node, to its ordered position. Node click opens the side panel, which is a **lite editor** over the fields it shows (name, type, approval status, summary, category with inline create, prerequisites with AND/OR grouping and add/remove, research costs with resource rows and min/max durations): dirty tracking, Save/Reset, and an "Edit in full view" jump to the standard admin edit route. Final approval states stay superuser-only (`canSetFinalApproval` = `system:admin-access`), mirroring the field access. Quick-create dialog (name, type, summary, category, research durations) creates via REST; the category picker is a native input + suggestion list (existing values selectable, new values creatable inline) because Radix Popover inside the admin Dialog proved layer-unstable. New categories get `position = max + 1`. The canvas uses React Flow's `<Background variant="lines">` so the grid scales with zoom (a CSS-background grid does not, and looked broken at low zoom), and `fitView` is capped at `maxZoom: 0.85`. Panel and toolbar controls carry dark tactical chrome (the admin panel's light shadcn defaults would render as white slabs on the dark surface; bare `<button>` elements in the admin additionally inherit a global light background AND border, so list rows need explicit `bg-transparent` + `border-0`), and the panel's dropdowns use the in-house `PanelSelect` listbox rather than native `<select>`: a native popup is browser/OS chrome that ignores dark styling (only `color-scheme` influences it) and cannot be screenshotted, whereas `PanelSelect` renders its list through a portal as a fixed-position overlay above the panel (so it never pushes the panel's content down), flips upward when there is more room above, closes on scroll/resize so it cannot detach from its trigger, and offers a search field for long lists. Both it and the panel's category picker use a mono header strip, hairline row separators, a cyan left accent on the selected row, and the `.tech-tree-scroll` dark scrollbar rules from `custom.scss`.
+- **Edges**: AND prerequisites solid slate, OR prerequisites dashed cyan with an OR label. Connections are editable on the canvas: drag from a node's right handle onto another node (the target card OR its handle) to add it as an AND prerequisite; the drop-on-card case is resolved by an `onConnectEnd` fallback because React Flow only resolves drops onto a handle. Select an edge and press Backspace to remove the prerequisite. Duplicates, self-references, and cycles (direct or transitive) are rejected by the pure `prereqEdits.ts` helpers with an inline error in the toolbar. Lines are also manually routable: selecting an edge reveals a "+" handle at each segment's midpoint (click to add a bend), bends drag to move, and double-clicking a bend removes it (`RoutedEdge.tsx` + `edgeRoutes.ts`). Routes are stored per browser in `localStorage`; making them shared would need a persisted field on the target technology. Edges also route around technologies they are not connected to: `edgeRouting.ts` detours each segment around the node rectangles it would cross (inflated by a clearance margin) and `buildRoutePath` rounds every corner. An unrouted edge keeps its automatic out/across/in shape unless that path would cut through a node, in which case it falls back to a direct line the router can bend around.
+- **Interactions**: nodes are draggable; dropping one PATCHes its `treePosition` (persisted, survives reload) and the toolbar shows an "Auto-arrange" button (technologies:update gated) that clears every manual position and restores the ordered layout. A **Save positions** button (technologies:update gated), next to Auto-arrange, commits the current on-screen position of every node to `treePosition` in one pass (matching what React Flow rendered, not just dragged/pinned cards) so the exact arrangement survives a reload; Auto-arrange then clears them. Dragging one card of a multi-selection moves and persists the whole selection (React Flow hands the full dragged set to `onNodeDragStop`), and a category band is draggable by its header (`dragHandle`), which moves every technology in that band. A **Snap: On/Off** toggle (default on, remembered per browser, 46px grid that divides `COLUMN_STEP`/`ROW_STEP`) snaps dragged cards to the grid. **Connecting does not move anything**: if the ordered layout would relocate the connected technology, the canvas pins it to its current spot (`treePosition` written in the same PATCH as the prerequisite) so the edge appears without the card jumping; the catalog grid also uses stable slots (each tech's rank among all uncategorized technologies, so a tech promoted into the tree leaves its slot empty instead of re-shuffling every other root). Pressing "Auto-arrange" clears those pins and moves everything, including the connected node, to its ordered position. Node click opens the side panel, which is a **lite editor** over the fields it shows (name, type, approval status, summary, category with inline create, prerequisites with AND/OR grouping and add/remove, research costs with resource rows and min/max durations): dirty tracking, Save/Reset, and an "Edit in full view" jump to the standard admin edit route. Final approval states stay superuser-only (`canSetFinalApproval` = `system:admin-access`), mirroring the field access. Quick-create dialog (name, type, summary, category, research durations) creates via REST; the category picker is a native input + suggestion list (existing values selectable, new values creatable inline) because Radix Popover inside the admin Dialog proved layer-unstable. New categories get `position = max + 1`. The canvas uses React Flow's `<Background variant="lines">` so the grid scales with zoom (a CSS-background grid does not, and looked broken at low zoom), and `fitView` is capped at `maxZoom: 0.85`. Panel and toolbar controls carry dark tactical chrome (the admin panel's light shadcn defaults would render as white slabs on the dark surface; bare `<button>` elements in the admin additionally inherit a global light background AND border, so list rows need explicit `bg-transparent` + `border-0`), and the panel's dropdowns use the in-house `PanelSelect` listbox rather than native `<select>`: a native popup is browser/OS chrome that ignores dark styling (only `color-scheme` influences it) and cannot be screenshotted, whereas `PanelSelect` renders its list through a portal as a fixed-position overlay above the panel (so it never pushes the panel's content down), flips upward when there is more room above, closes on scroll/resize so it cannot detach from its trigger, and offers a search field for long lists. Both it and the panel's category picker use a mono header strip, hairline row separators, a cyan left accent on the selected row, and the `.tech-tree-scroll` dark scrollbar rules from `custom.scss`.
 - **Styling**: tactical dossier (dark `#05070a` canvas + 28px grid via inline styles, squared corners, classification strips, mono readouts); shadcn form controls in the dialog stay conventional. React Flow controls/attribution are darkened via `.tech-tree-canvas` rules in `src/app/(payload)/custom.scss`.
-- **Gotchas**: React Flow in controlled mode needs `useNodesState` AND `useEdgesState` with change handlers that forward everything except `remove` (a no-op edge handler silently breaks edge selection, and a passed-through remove lets Backspace delete nodes that only exist in the data layer); category band nodes must be `pointerEvents: "none"` or they swallow edge/pane clicks; node handles must live OUTSIDE the card's `overflow-hidden` box or they become un-draggable; a second `next dev` for the same directory is impossible (Next 16 lock) so e2e verification either frees port 3000 or uses a copied repo (whose file watcher is unreliable: restart the copied server after every source change). technologies create requires `approvalStatus` + `researchCosts.minimumResearchDuration` (the field's `defaultValue: 0` violates its own `min: 1`, so creates must pass it explicitly); clearing `treePosition` means sending `{ x: null, y: null }` (a group cannot be nulled wholesale). The e2e seeds via cookie-authed REST (`page.request`) after frontend login, so it MUST run against the test DB: a dev server left on port 3000 is silently reused (`reuseExistingServer: true`) and its DEV database 403s every seed (the dev user has no create permission). Free the port first. Tests: `tests/int/tech-tree-layout.int.spec.ts` (pure layout: wrapping, overrides, band bboxes, warnings, cycles), `tests/int/tech-tree-prereq-edits.int.spec.ts` (pure add/remove/switch with cycle rejection), `tests/int/tech-categories.int.spec.ts` (CRUD, RBAC, division gating, FK set-null, position round-trip) + `tests/e2e/tech-tree.e2e.spec.ts` (nav, render, panel, edit nav, both quick-create paths, drag persistence + Auto-arrange, save positions + reload survival, connect/disconnect + lite-editor saves, wrapping at scale). Vivaldi crashes under automation in this suite; run with `E2E_BROWSER_PATH=<playwright chromium>`.
+- **Gotchas**: React Flow in controlled mode needs `useNodesState` AND `useEdgesState` with change handlers that forward everything except `remove` (a no-op edge handler silently breaks edge selection, and a passed-through remove lets Backspace delete nodes that only exist in the data layer); category band nodes must be `pointerEvents: "none"` or they swallow edge/pane clicks; node handles must live OUTSIDE the card's `overflow-hidden` box or they become un-draggable; a second `next dev` for the same directory is impossible (Next 16 lock) so e2e verification either frees port 3000 or uses a copied repo (whose file watcher is unreliable: restart the copied server after every source change). technologies create requires `approvalStatus` + `researchCosts.minimumResearchDuration` (the field's `defaultValue: 0` violates its own `min: 1`, so creates must pass it explicitly); clearing `treePosition` means sending `{ x: null, y: null }` (a group cannot be nulled wholesale). The e2e seeds via cookie-authed REST (`page.request`) after frontend login, so it MUST run against the test DB: a dev server left on port 3000 is silently reused (`reuseExistingServer: true`) and its DEV database 403s every seed (the dev user has no create permission). Free the port first. Tests: `tests/int/tech-tree-layout.int.spec.ts` (pure layout: wrapping, overrides, band bboxes, warnings, cycles), `tests/int/tech-tree-prereq-edits.int.spec.ts` (pure add/remove/switch with cycle rejection), `tests/int/tech-categories.int.spec.ts` (CRUD, RBAC, division gating, FK set-null, position round-trip), `tests/int/tech-tree-edge-routes.int.spec.ts` (rounded route path + obstacle avoidance) + `tests/e2e/tech-tree.e2e.spec.ts` (nav, render, panel, edit nav, both quick-create paths, drag persistence + Auto-arrange, save positions + reload survival, multi-select drag, category-band drag, snap-to-grid, edge routing, connect/disconnect + lite-editor saves, wrapping at scale). Vivaldi crashes under automation in this suite; run with `E2E_BROWSER_PATH=<playwright chromium>`.
 
 ## Wiki
 
@@ -371,7 +372,7 @@ A Discord bot living in `src/bot/`, run as a standalone long-running process via
 - **Evaluation reminders**: once a mission is evaluable (`Completed`, or `Scheduled`/`Active` past its start — same rule as `isMissionEvaluable`), the bot sends a one-shot DM (per user with a linked `discordId`) asking them to rate their leadership, plus a "rate your subordinates" section for leaders whose members RSVP'd yes. Links go to each ratee's profile page where the rating dialog lives. Recipient computation: `computeEvaluationReminderPlan` in `src/lib/evaluations/reminders.ts`; state marker: `evaluationRemindersSentAt` on Missions (one-shot, same pattern as `discordAttendanceSentAt`; withheld if the 40-DM/tick cap is hit — retries next tick). These DMs deliberately ignore `preferences.discord.enabled` (defaults false, not yet exposed in the web UI — honoring it would DM nobody). Staff can trigger manually via `/remind-evaluations` (optional `mission` option accepts an ID, name, or code name and re-sends even if the marker is set; omitted, it sweeps all pending missions and reports `{processed, sent}`).
 - **Sign-up / linking (feature 1)**: `/signup` is **DM-only** (the temp password flows through the DM). Creates the Payload user with `username` = `discordUsername` = the caller's Discord username, plus `discordId`, `displayName`, `steamId`, and a random temp password. The ephemeral reply carries the password as the guaranteed delivery path; `interaction.user.send()` is a best-effort persistent copy, so a blocked DM never orphans the account. `/link` works in servers **and** DMs (credential-free, ephemeral reply only) — matches `discordUsername` → sets `discordId`.
 - **DM gotcha**: a user with "Allow direct messages from server members" off in Discord privacy settings can neither receive the bot's DMs nor open a DM with the bot. The `/signup` rejection message explains how to enable it.
-- **Attendance (feature 2)**: `mission-attendances` collection + `src/lib/attendance/` (single write path, emits `mission:attendance-change`). The bot posts RSVP embeds (Yes/Tentative/No) for future, `Ready`/`Scheduled`, `visibility: "unit"` missions into the ops channel; stores `discordMessageId` + `discordAttendanceHash` on the mission; reconciles hash changes every poll tick (web ↔ Discord two-way sync, loop-safe). Web UI: `MissionAttendance` component on the mission detail page. `bun run payload generate-mission` clones the next weekly main mission.
+- **Attendance (feature 2)**: `mission-attendances` collection + `src/lib/attendance/` (single write path, emits `mission:attendance-change`). The bot posts RSVP embeds (Yes/Tentative/No) for future, `Scheduled`, `visibility: "unit"` missions into the ops channel; stores `discordMessageId` + `discordAttendanceHash` on the mission; reconciles hash changes every poll tick (web ↔ Discord two-way sync, loop-safe). Web UI: `MissionAttendance` component on the mission detail page. `bun run payload generate-mission` clones the next weekly main mission.
 - **Notifications / announcements (feature 3)**: `notificationBridge` polls `user-notifications` (cursor = last seen id; cap 5 DMs/tick) and DMs opted-in users (`preferences.discord.enabled`, not in `mutedTypes`). `/announce` (staff only) posts an announcement embed. New notify sites partially done: banking emits `finance:deposit`; shipments has no notify site yet.
 - **Remaining polish**: the web preferences UI does not yet expose the `preferences.discord` toggles — they exist on the Users collection and are read by the bridge, but are currently only settable in the Payload admin panel.
 
diff --git a/CHANGELOG.md b/CHANGELOG.md
index ba87104..f4936ac 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,7 +8,6 @@ Your in-game wallet and web bank account now talk to each other, and staff get a
 
 - **In-game money syncs to your bank account.** Shop purchases in Arma now run through the web banking system: what you spend in game is debited from your personal account and moved to the unit treasury, while refunds and rewards land back in it. The game can also read your web balance, so linked players see one consistent wallet. Retried updates can't double-charge anyone.
 - **The tech tree got a visual editor (admin).** Staff can arrange technologies on a canvas in the admin panel instead of editing fields by hand: drag technologies into place, connect prerequisites with AND or OR logic, group them into categories, and hit auto-arrange when the layout gets messy. Positions and connections save right from the canvas.
-- **We added site analytics.** Microsoft Clarity now tracks how the site gets used in production builds, which helps us see which features are worth polishing. Nothing runs locally or in the admin panel.
 
 ---