fix(compiler,harness): the wave countdown is a liveness census, not a kill tally (task #178)

[stellarfeline]

The divergence

The owner played the staged hollow-vigil build (engine worker/spec-0026-valley @ 649a164, content @ ce82515) on 2026-08-05 and hit a hard softlock: after the first zombie wave, the campaign never advances and there is nowhere to obtain the barrow key. The #157 ladder on a byte-identical build was all green — delvec build clean, PackTest 19/19, bot passed: true with both encounters census-cleared.

Both were telling the truth about different games.

The delve's rule. #<wave> dw.wave was a tally of PLAYER KILL CREDITS: spawn_<wave> seeded it with the wave size, and only the player_killed_entity advancement decremented it. The party's rule. A kill objective asks for the wave to be down — and a mob has many ways to die that credit nobody. This campaign's opening beat had its roof and two walls carved off for the open-air rule, the world is pinned at time set noon, and the wave stands one door from the yard. The footmen chased the party out and burned.

Reproduced live on the shipped datapack (rcon, worker-isolated server):

[01:48:43] Named entity '空洞步卒' … x=6.28 z=-3.24 died: 空洞步卒 burned to death
[01:48:43] Named entity '空洞步卒' … x=2.97 z=-8.35 died: 空洞步卒 burned to death

both outside the carved north wall. Terminal state:

	shipped	fixed
wave mobs left in the world	0	0
kill countdown	2	0
obj/purge	never	completes
flag/halls-quiet	never	set
quest/the-barrow-key armed	never	armed
barrow-key chest placed anywhere	no	yes, with the trial_key

Why every rung was green

1. PackTest. The generated verb_kill kills the wave and then calls k_reward_<wave> once per mob by hand — it supplies the player-kill credits itself. It asserts the countdown arithmetic, never the causal link "the mob died ⇒ the countdown moved".
2. Bot. MineflayerExecutor.kill was the one objective-bearing step that never called requireObjective. talk-to, reach, collect, interact all end on the delve's own [dw:complete …] marker; kill ended on MOB-side evidence — confirmed kills, every engaged mob down, the server's live wave census — and returned. critical-path.json format 2 already documented the marker contract for kill; the harness simply did not implement it.
3. Compile time. Nothing relates "a mob can die here without a player" to "this objective's completion".

So the machine proved: the room can be emptied. The player needed: the campaign accepts that the room was emptied.

What changes

Emission — one line per wave a kill objective adjudicates, in tick, ahead of every completion check:

execute if score #<w> dw.wave matches 1.. store result score #<w> dw.wave if entity @e[tag=dw_wave_<w>]

Vanilla's own answer to "how many are left", so no new primitive and no bookkeeping to keep honest. matches 1.. does both guard jobs: before spawn_<wave> the counter is unset, so an empty selector can never read as "cleared"; after a legitimate zero the line stops firing, so a finished wave stays finished and a respawns_on_rest re-seat (whose kill+spawn are atomic) re-arms cleanly. The advancement decrement STAYS as a fast path — a mob the player just killed is matched by @e through its ~20-tick death animation, so without it the final blow would land a second before the objective noticed.

Harness — kill now ends on requireObjective(step.objective) like every other step. The census logic is untouched and still decides when to stop swinging; it no longer decides whether the step is proven.

Ladder — new generated PackTest kill_uncredited: verb_kill minus the hand-fed k_reward calls. The same wave, wiped the same way, with nobody credited for it — the objective must complete anyway. Atomic (one tick, no await), so the batch model needs no new reasoning about it.

What CI now proves

• Red on the shipped delve. Shipped datapack/ + this branch's packtest-datapack/:
  1 required tests failed :( - hollow-vigil:kill_uncredited: Expected #party dw.o_warden to match 1, but got 0 (19/20).
• Green on the fixed one. All 20 required tests passed :).
• Bot tier on the fixed build: critical path 'hollow-vigil' PASSED (9 steps, 2 advisory finding(s)).
• Harness: a new unit test pins the contract — a fight genuinely won whose marker never arrives reds the step, naming the objective; and the "soft lock, loudly" fixture now asserts the step reds rather than merely recording findings.
• Determinism: two builds byte-identical. The datapack diff against the staged build is exactly the two census lines.

docs/reference/compiler.md records both halves (the kill/spawn-wave row and the critical-path.json format-2 note).

Not fixed here

The wave can still be won by the sun rather than the party — a difficulty defect, not a completability one. The DSL already carries the sanctioned remedy (equipment.head, "never set-time"); the content fix is delvewright-campaigns#worker/hollow-vigil-daylight-helm. The compile-time version of it — a diagnostic for daylight-burning undead whose arena is walk-reachable from sky-open cells, buildable on the existing compiler::nav BFS and light.rs::sky_open — needs an owner-approved spec and is proposed, not built.

Merge gate

Engine emission + campaign content, so the owner's own playtest is the gate (batched, never per-PR). The machine red→green above admits it to her batch; it does not replace her.

Branch is based on worker/spec-0026-valley @ 649a164 so it stacks on the build she will replay; it carries a pick of the already-merged #263 so the bot tier can run 0.9.0 from this base.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RnrGdWS7ADsjCoJfiJc6GC