PunchThemAll

## [1.21.1-2.2.0] — NeoForge

A correctness pass over 2.1.0, two new authoring fields, and the first automated test suite.

### Added
- **`hand.consume.count`.** How much a successful interaction spends, on top of the existing
`chance`, which only ever decides *whether* anything is spent. The two are independent and
compose:

```json
"consume": { "mode": "shrink", "chance": 0.33, "count": { "min": 3, "max": 5 } }
```

= one click in three costs 3–5 items. It takes the same shapes as every other count (`3`,
`{ "count": 3 }`, `{ "min": 3, "max": 5 }`), defaults to `1`, and has a floor of `1` — spending
nothing is what `chance` and `mode: "none"` are for. Under `mode: "durability"` it is durability
points rather than items, and a tool that runs out still breaks exactly once. Holding fewer items
than the roll asks for never blocks the interaction; it just costs what is there. JEI shows an
*Amount* line on the hand slot whenever the value is not `1`, and EMI now carries the held-item
chance and amount on its arrow summary (it showed neither before). Example:
`pta_examples:hand_consume_count`.
- **`hidden`** on an interaction. `"hidden": true` keeps it out of JEI and EMI while it loads, syncs
and fires exactly as before — for secrets, and for the intermediate steps of a multi-stage recipe.
Distinct from `enabled: false`, which is the one that turns an interaction *off*. Hidden
interactions also stop stretching the category to fit their drop rows. Example:
`pta_examples:hidden_from_viewers`.
- **A test suite**: 244 tests over the NBT matchers, the count/weight arithmetic, the codec, the
resolver, the registry and the sync batching. `./gradlew test`. It boots the game's registries in
process, so it covers real `ItemStack` and NBT behaviour rather than stand-ins.
- **A warning when `type` and `conditions.requires_sneaking` contradict each other.** Pairing
`shift_left_click` with `requires_sneaking: false` yields an interaction that can never match; PTA
now names the file instead of leaving you to work it out. The warning immediately found one in our
own example datapack — see below. `requires_sneaking` is redundant with `type` in every case and is
now documented as such; prefer the `type`.
- **Unresolvable `sound`, `particles` and biome entries are reported.** They were dropped silently,
which is indistinguishable from never having written them.

### Fixed
- **A malformed id no longer takes down the whole datapack load.** `sound`, `particles`,
`effects[].id` and `rewards.fortune.enchant` used to throw on an id that was not a valid resource
location — inside the reload, so a single typo cost the pack *every* interaction. Bad ids are now
reported per file and skipped, and one interaction failing to resolve no longer aborts the rest.
The same applied to a malformed biome `#tag`, which threw on every click for as long as the pack
was installed.
- **`count: { "min": 0, "max": 3 }` on a drop now means "nothing to three".** It silently meant
"never anything": the entry produced no items, vanished from the viewers, and still consumed its
weight in the pool — while the schema and docs both advertised it as valid.
- **A shape mismatch in `nbt.whitelist` / `nbt.blacklist` no longer crashes mid-click.** Requiring a
list where the item or block entity holds something else threw a `ClassCastException` from inside
the interaction filter. It now simply does not match.
- **Which interaction wins is now deterministic.** Matches were collected into a hash set, so with
`max_matches_per_click`, or with several interactions competing to transform one block, the winner
varied between runs. Candidates are ordered by id.
- **Fortune can no longer produce an over-sized stack.** The bonus is clamped to the item's maximum
stack size; the surplus used to disappear the moment the drop entered an inventory.
- **The recipe viewers no longer churn on every reload.** An interaction now compares equal to itself
across a reload, so JEI hides and re-adds only what actually changed. Previously every entry looked
new, and JEI's permanent hidden-recipe set grew each time.
- **A client leaving a server forgets its interactions**, instead of showing the previous server's set
in JEI on the main menu and into the next world.
- **Cooldown bookkeeping no longer grows without bound.** Fake players never log out, so every machine
that clicked left an entry behind for the life of the server. Stale entries are pruned, and
everything is dropped when the server stops.
- **The interaction registry is published as one immutable snapshot.** In singleplayer the client and
the server rebuild the same registry from two threads while a third reads it on every click; the
old clear-then-refill could be observed half-empty.
- **A cost's maximum is clamped against its floored minimum**, so `amount: 0` no longer produces an
inverted range.
- **Slot counting matches what the viewers lay out**, so a drop grid can no longer overflow the
category background.
- **The `conditions_sneaking` example never worked.** It shipped `type: right_click` with
`requires_sneaking: true` — the same axis, contradicting — so it could not fire since 2.1.0. It is
now a plain `shift_right_click`, and the catalogue entry explains the trap instead of demonstrating
it.
- **Four examples never showed their particles.** `hand_off_hand`, `left_click_air`,
`target_block_entity_nbt` and `transformation_break` passed particle-*type* ids
(`minecraft:happy_villager`, `minecraft:cloud`, …) where `particles` wants a **block** id. The docs
always said so; the examples did not follow, and the failure was silent. Both halves are fixed —
the ids, and the silence.

### Changed
- **A client without PunchThemAll can now join a PunchThemAll server.** The network channel is
optional: every gameplay decision is server-side, and the sync only feeds JEI/EMI. Such a client
simply sees no PTA entries. *(Protocol version bumped to `2`; a 2.1.0 client and a 2.2.0 server will
not talk to each other.)*
- **The interaction sync is sent in batches.** The specs travel as NBT, which vanilla caps at 2 MiB
per tag — a large pack in one payload would have dropped clients at join with a decode error.
- **Registry lookups answer an unknown id with `null`** rather than the registry default, so a typo
can no longer read as `minecraft:air`.
- **Debug logging can no longer break loading.** Reading a config flag before the config file is
attached now falls back to its default instead of throwing.
- `kind: "any"` on `minecraft:water` / `minecraft:lava` resolves to the **block**, since those ids
exist in both registries and a target cannot mix the two. Documented; use `kind: "fluid"` if you
want the fluid.

---