fix(policy): add craft category + narrow inventory_query keyword scope

[Pablomonte]

Summary

GemmaPolicy's L4 classifier had two collaborating bugs that made any craft-style intent unable to dispatch the actual crafting tools:

1. No 'craft' category existed. Captain-side intents like `Craft a stone shovel using cobblestone and sticks` had nothing to match in the keyword table.
2. 'inventory_query' keyword 'inventory' was too broad. Any intent containing the word "inventory" — including the perfectly natural "…from my inventory" clause that captains routinely emit — matched, narrowing tools to `[get_inventory, ask_clarification, report_execution_error]`.

Observed in production

End-to-end trace from a live Hermes-K2.6 + DaemonCraft session:

```
Captain (Kimi-K2.6) → embodied_plan(intent=
"Craft a stone shovel using cobblestone and sticks from my inventory."
)
→ policy.classify_category() loops CATEGORY_KEYWORDS
→ hits 'inventory_query' via keyword 'inventory'
→ allowed_tools = [ask_clarification, get_inventory, report_execution_error]
→ Gemma-Andy plan: ask_clarification("craft_item is not available
in this turn; should I wait?")
→ captain tells player: "no tengo capacidad de craftear ahora"
```

The user sees a hard wall on every crafting request despite Gemma-Andy and the embodied service being fully capable.

The fix (3 additive changes to `agents/gemma_policy.py`)

1. New 'craft' category

Inserted before mining/build so it wins on overlap. Keywords cover ES+EN imperative forms:

```python
("craft", ["craft ", "craftear", "crafteá", "craftea",
"fabricar", "fabricá", "fabrica",
"smelt", "fundir", "fundí", "fundi",
"cocinar", "cociná", "cocina",
"horno", "furnace", "forja", "forjar", "forge"]),
```

Tool subset:

```python
"craft": ["get_inventory", "view_craftable", "craft_item", "smelt_item",
"check_furnace", "take_from_furnace", "equip_item",
"place_block", "scan_nearby", "goto"],
```

Includes `place_block` for placing crafting_table/furnace first; `goto`/`scan_nearby` for finding a crafting station nearby.

Mapped through `STRATEGY_MAP["craft"] = "embodied_plan"` like every other category.

2. Tighten 'inventory_query' keyword scope

```diff

• ("inventory_query", ["inventario", "inventory", "decime qué tenés", ...

• ("inventory_query", ["inventario", "decime qué tenés", ...

•                   \"list inventory\", \"list my items\"]),
  

```

Drop bare `"inventory"` — too greedy. Keep `"inventario"` (Spanish-only, no overlap), `"show inventory"`, `"list inventory"`, and the existing Spanish phrases. Pure inventory-show intents still classify correctly:

Intent	Category
`mostrame el inventario`	inventory_query (3 tools)
`inventory check` (no "inventario" or specific phrase)	default (38 tools)
`Craft X from my inventory`	craft (12 tools) — was inventory_query before

3. Extend 'build' tools + keywords

Bonus cleanup that surfaced while fixing #1:

```diff

• "build": ["scan_nearby", "goto", "place_block", "equip_item", "get_inventory"],

• "build": ["scan_nearby", "goto", "place_block", "fill_volume", "ignite",

•       \"equip_item\", \"get_inventory\"],
  

```

`fill_volume` enables bulk construction without dropping category. `ignite` covers torch/lava placement scenarios.

Build keywords also extended with `"fill "`, `"rellena/rellená"`, `"construir"` (the infinitive — only imperative forms were covered).

Verification

Live test of the patched policy:

```python

p = GemmaPolicy()
intents = [
... "Craft a stone shovel using cobblestone and sticks from my inventory.",
... "craft 4 oak planks",
... "fundir 16 arena en vidrio",
... "construir una casa de 5x5",
... "fill volume con cobblestone",
... "mostrame el inventario",
... "inventory check",
... ]
for it in intents:
... cat = p.classify_category(it.lower())
... n = len(p.get_allowed_tools(cat) or []) or "ALL"
... print(f"{cat:18s} [{n:>3}] ← {it[:55]}")

craft [ 12] ← Craft a stone shovel using cobblestone and sticks from
craft [ 12] ← craft 4 oak planks
craft [ 12] ← fundir 16 arena en vidrio
build [ 9] ← construir una casa de 5x5
build [ 9] ← fill volume con cobblestone
inventory_query [ 3] ← mostrame el inventario
default [ 38] ← inventory check
```

Before this patch, the first three returned `inventory_query 3`.

Test plan

• `pytest tests/test_gemma_policy.py` — 29/29 pass.
• Manual end-to-end: `@AsciiProbe haceme una pala de piedra` now dispatches `embodied_plan(intent="Craft a stone shovel...")` → policy emits 12 tools → Gemma-Andy proposes `craft_item` → bot crafts.
• Follow-up PR: add regression test cases for the build/craft/inventory keyword collision matrix.

Upstream sync

This patch lives in `agents/gemma_policy.py` of the DaemonCraft canonical source. The vendored copy in `nicoechaniz/hermes-agent` (`tools/gemma_policy.py`) needs the same diff applied — that goes in a separate PR to the hermes-agent fork.