sap2: Tier 2 reaches parity - cascades, summon cells, the faint queue

[pjsny]

Tier 2 is now behaviourally identical to the shipped build on every differential this project has, Spider excepted. Six measured rules, three harness bugs that were reporting green on the wrong question, and the first test here that crosses a phase boundary.

The rules

1. A faint cascade is one batch, not a sequence. Everyone at ≤0 fires BEFORE_DEATH, then every body leaves, then every DEATH fires. Four pets killed by one blast give four DeathEarly events and then four Death events, never interleaved. This is what a summon count turns on: a level‑3 Rat lands three Dirty Rats because the corpses are gone by the time its DEATH resolves.
2. The pending faints are a priority queue on attack, not a batch sorted once. A 5/1 Hedgehog against [Hedgehog 3/1, Flamingo 1/1, Otter 1/4] kills the Otter in 120 of 120 seeds; swap those two attacks and it lives in 120 of 120, because the 3‑attack Flamingo — not pending when the queue started — jumps the 1‑attack Hedgehog. Equal attacks: 0.518/0.482 over 600 seeds, i.e. the coin flip.
3. A summon takes the cell the body vacated, and pushes what filled it. Three Dirty Rats each take the front cell and push; a Cricket's token then claims cell 3 — its own — and pushes the two behind it again: {1: RatToken, 2: RatToken, 3: CricketToken, 4: RatToken}. No rank arithmetic reproduces that, which is why the battle line now carries real cells and compacts once per exchange the way the engine's EmptyFront does.
4. A body at ≤0 health is on the line but is not a legal target. A splash skips it, a positional finder steps over it without spending a target, and it cannot be healed back — a Crab taken to exactly 0 never casts.
5. Insert slides forward, toward the front, and only slides back when the front is full.
6. Faints resolve by identity, not by array index — resolving one can shift the other side's line (Rat summons onto the opponent's front), which had been leaving corpses fighting on at negative health.

The harness was wrong three times, in the same way

All three suites were asking a question the shipped game does not answer deterministically:

• --survivors demanded an identical line-up from boards that are a coin flip in the real engine too (two simultaneous faints with equal attack). It now re‑resolves a mismatch over 200 fresh seeds in both engines and compares distributions.
• --deterministic had the identical flaw — excluding Ant/Mosquito/Spider does not make a board deterministic.
• --dist counted Spider's unimplemented summon as a divergence; it now names it under UNIMPLEMENTED so the worst delta is quoted over implemented rules only.

Plus: the differential probe had been replicating the battle driver rather than asking the env for the surviving line, so the suite spent a while comparing the shipped engine against a loop this env does not run; and the shop drivers mapped real grid cell i to sap2 slot i where the verified battle mapping mirrors, which is what hid the insert direction. battle.team_cell is now the one place that mirror lives, with a standing guard against a second one.

Verification

Check	Result
Tier‑2 exact surviving line‑ups	6000/6000 (8 coin‑flip boards, compared as distributions)
Tier‑1 exact surviving line‑ups	300/300
Tier‑2 winners	1500 boards, 0 disagreements, 1 coin flip
Holed boards	200/200 Tier 1, 600/600 Tier 2
Scripted shop checks	26/26
Shop fuzz, turn 1 / turn 3	2997 + 1795 actions, no divergence
Shop→battle bridge	8 scripted + 55 random matches, no divergence
Random‑ability distributions	worst delta 0.053, 4 boards named Spider
Env tests	72

debug_resolve_battle is new: the match API cannot express a named five‑pet board, and the battle half was the most‑measured and least‑tested part of the env. It is documented as not agent‑facing and range‑checks every field — 14 malformed input shapes raise TypeError/ValueError/OverflowError, none crashes.

Pack scope, measured

A pack is a table swap, not a rules change: nothing in the build keys a constant on Pack, and ten playable packs generate boards identical in every scalar field, differing only in the roll pool. Pack1 is not disjoint — four species and seven foods are shared, and a shared item is the same row. And the real ladder is cross‑pack: GetPossibleOpponents(Pack1) returns six packs, BoardModel carries both Pack and OpponentPack. This env is a symmetric two‑seat match where both seats roll one pool, so a cross‑pack pairing is not representable — recorded as a deliberate scope difference, alongside matchmaking itself.

Left

• Spider summons nothing until Tier 3's roster exists. It is the only known gap in tiers 1–2.
• The level‑up reward (two higher‑tier pets offered at 3 gold) and Sloth (~1 roll in 9,091) are measured and unimplemented.
• RNG identity remains deliberate divergence: distributional equality, not seed‑identical replay.