new-creature-mount

name	new-creature-mount
description	Author a rideable creature mount end-to-end (assets, Monster/action/usage XML, C# behavior tree, validation) following the elephant+spider-proven workflow. Warg parity is law.

New Creature Mount

Thin entry point over the authoritative doc: docs/ai-includes/creature-mount-authoring.md — read it FIRST and follow its phases in order. It encodes both full campaigns (war elephant 2026-06, giant spider 2026-06) including the 16-gotcha index and the v1.4.6 lookup-hardening rules. This skill adds only the execution order, the gates, and the top traps.

Architecture (never deviate): vanilla cavalry spawn — rider troop + Horse-slot item whose HorseComponent names the Monster; engine does ALL mount work; TAOM layers attacks via a per-agent behavior tree. NO spawn patches, NO detached combatants (built twice, deleted twice). The warg (Alliance.Wargs) is the reference implementation — when in doubt, do what the warg does, byte-for-byte in shape.

Phase order (each gated before the next)

1. Assets (doc Phase 0–1): skeleton ≤64 bones; meshes ≤~38 bones each (split + recombine via <AdditionalMeshes>); clips in-place; every gait clip carries quad_movement + step points (Kit Clip usages, not Flags), gallop-pace runs also cyclic.
2. Monster XML (Phase 2): num_paces=6, family_type=1, Flags EXACTLY Mountable CanRear RunsAwayWhenHit CanCharge CanWander — CanAttack is forbidden (engine attack-AI path; 1.4.6 charge CTD). Rein surface + rider capsule/eye adders.
3. action_types (Phase 3): the typed-verb table verbatim — 12 actt_fall (+_continue), rear/kick/dash/quick-stops/hit_object/strikes typed; light strikes UNTYPED *_while_moving; jump_start action typed actt_dash, NEVER actt_jump; a dedicated actt_idle _1.
4. action_sets (Phase 4): bind every usage-referenced action to a VALIDATED clip; explicit act_horse_forward_canter binding; _map + _town_and_village children; the rider partial as_human_warrior at the TOP of the file (base_set snapshots at definition).
5. monster_usage_sets (Phase 5): all 10 verb attrs; per-pace direction="none" reference rows; jump table TOTAL — all 9 directions × all states = 45 rows (a missing lookup key CRASHES on 1.4.6; an extra row is inert); warg-exact falls + strikes matrices. Registration = project.mbproj standard soln_* ids ONLY (subfolder XML copies are dead decoys).
6. Item + troop (Phase 6), then C# (Phase 7): clone the elephant's MissionBehavior/BT wiring (attach keyed on Monster.StringId, never character id); pure attack service; mount-lock in TaomAgentStatCalculateModel; extend Patch47's monster predicate so riders dying on the new mount take the dismount-before-death path.

Validation gates (run BEFORE any battle test — parity-audit-first beats per-crash fixing)

1. Extend tools/audit_mount_parity.py (FILES + MOUNTS maps) for the new creature → zero unaccepted deltas vs warg/elephant/horse.
2. Animation-target sweep: every animation= byte-scanned against the real tpac inventories (module + Alliance.Wargs packs + vanilla Native\...\animation_clips.tpac) — a phantom target compiles a degenerate record that AVs later.
3. XML parse-validate every edited file; back up external-module files first (.bak-<topic> beside each, never overwrite existing backups).
4. In-game ladder: thumbnail → deployment → charge (jumps!) → melee → rider deaths → mount deaths. On any CTD: /native-crash-triage — never blind-retry.

Standard follow-through

External LOTRLOME edits get ledger entries (docs/reference/lotrlome-spider-mount-changes.md pattern); feature doc per docs/features/elephant.md/spider.md shape; CHANGELOG; GitHub issue; the /ship sequence for the C# delta.