Address Discord feedback: Reforged skin data, custom model rendering, shortcuts, terrain palette

[StephenSHorton]

Addresses four issues from a Discord user testing wc3-forge against a heavily-customized Reforged map. All four were reproduced/root-caused in the source; two were confirmed data bugs (not just World-Editor-familiarity gaps).

1. Custom units all showed as base "Peasant" (name/model/icon) — confirmed bug

Reforged splits per-map object overrides across two tables per kind: war3map.w3* (gameplay) and war3mapSkin.w3* (art/skin: Name unam, Model File umdl, Icon uico, Tooltip utip). wc3-forge read only the primary half, so custom units loaded the right row count (from w3u) but displayed the base name + default model — the real name/model/icon live only in the never-read war3mapSkin.w3u.

• Session.Open now reads the war3mapSkin.w3* companion for every kind (read-only, keyed by kind).
• MergedObjects layers it under the primary shadow (primary wins on any field it sets; skin fills the art fields the primary leaves unset).
• Skin tables are never marked dirty → Save preserves them verbatim via the lossless copy-through (no data-loss risk).
• New objects_skin_test.go covers the overlay + primary-wins precedence.

Precedence note (for review): "primary wins, skin fills gaps" is faithful for real maps (name/model live only in the skin table) and keeps the edit path coherent — edits route to war3map.w3u and are never masked by the skin overlay. Writing edits back into the skin table (full WE round-trip semantics) is intentionally left as a follow-up.

2. Placed units rendered with the wrong (base) model — confirmed bug

Separate code path from #1: the viewport's unit merge (typeindex.go) carried only name + icon, never the model file, so custom-skin units drew the base model. Doodads already did this correctly.

• unitFieldMap: umdl → file, usca → modelscale; applyUnitOverrides sets the model (extension-stripped) + scale.
• mergeUnitIndex layers the skin companion then the primary (same precedence as feat: native macOS support (needs Windows regression triage) #1).
• Frontend: the unit type index is now per-map — scene-instances.ts drops the process-wide cache on each loadMap (it was cached for the process lifetime, which would have hidden the fix). (Tint uclr/g/b deliberately skipped for now to avoid a 0-vs-255 default regression.)

3. F4–F12 did nothing; the menu advertised dead F4/F6 labels

• Wired the advertised keys: F4 → Trigger Editor, F6 → Object Editor.
• Added F2/F3/F8 for Doodad/Terrain/Region mode (matching the existing Tab-cycle + header buttons).
• Added a keyboard-shortcuts cheat sheet (? / Help menu; Esc/backdrop/× to close) — the discoverable index a WE veteran reaches for.

4. Terrain palette showed "No ground tiles" on a map full of terrain

Root cause turned out to be a missing mount-fetch (not the folder-extract theory): the panel mounts already-open but refresh() only ran from the FAB toggle or a map-changed event that had already fired before Terrain mode was entered.

• refresh() now runs on mount.
• The empty state distinguishes "This map has no terrain file (war3map.w3e)." vs "Couldn't load tiles." + a Retry, instead of always claiming "No ground tiles."

Verification

• go build ./..., go vet ./... clean; go test ./... — all packages pass except a pre-existing mpq fixture-drift test (TestExtractMatchesStormLib_survival, unrelated personal fixture, fails identically on main).
• frontend: npm run check (0 errors) + npm run build (green).

🤖 Generated with Claude Code