docs(spec): the two open measurements are settled — and one premise was wrong

[stellarfeline]

spec-0031 and spec-0032 both blocked on the same question, and the question was phrased wrongly. Both said "the pre-respawn death advancement". The engine's edge is not an advancement — it is a deathCount scoreboard criterion, guarded by unless data entity @s {Health:0.0f}, which is exactly what defers cp_respawn_check to the tick after respawn.

The answer is favourable

5 causes × 3 repeats on pinned 1.21.11, every repeat agreeing: the edge is armed on the corpse, pre-respawn, for void, fall, drowning, lava and a mob control; LastDeathLocation is written on the same tick; the corpse's position is the death position and does not drift (0.000 — a corpse stops falling).

So on_death can fire corpse-side for every cause, and the recovery stake can be placed for every death a souls-shaped delve produces. spec-0032's single load-bearing unknown clears.

The reusable lesson

An advancement would have been the wrong instrument. entity_killed_player covers 1 of 5 causes. entity_hurt_player fires on the first damage event with HP still 16–20 — it means "was hurt", never "died here". Implemented as the spec wrote it, four of five death causes in a souls delve would have gone unaccounted, and the gap would have surfaced as content that mysteriously fails to fire.

Fall damage across a mid-fall teleport

It carries and applies at the destination — Δ exactly 0.0000 in 46/46 trials, including teleports 143 and 157 blocks straight up. Landing damage is floor(fall_distance) − 3; survived at 20.279 (3 HP left), died at 23.435.

An arriving lift car does not catch a falling player past ~20 blocks — it is the surface they die on. That agrees with the owner's design of record, which already says a player who jumps the shaft with the car below takes ordinary fall damage, lethal or not by height.

Recorded as unmeasured rather than assumed: the exact boundary between 20.279 and 23.435 (free fall quantises in ~3-block steps, so floor − 3 is a fit); whether a real client matches the bot; and what does reset fall distance — only same-dimension /teleport was tested, so the mechanism that would make a lift safe is unmeasured.

Two incidental findings that constrain the verbs

• 1.21.11 rejects every legacy camelCase gamerule (all eight probed answer Incorrect argument for command and change nothing). The compiler already emits new names. The defect is not the rename — it is that neither offending site reads the command's response, so a rule that silently stopped applying looked identical to one that worked. Filed separately.
• A respawned player is invulnerable for 59 ticks (~3 s), and /kill answers Killed <player> while doing nothing. Directly load-bearing for the lethal volume and the shaft-bottom volume, and it explains two things previously read as instrument flakiness. Filed separately.

What CI proves

Doc-only. check-doc-dupes and check-reference-versions green. The measurements themselves landed in #349 with the rig and 4140 raw samples committed and re-runnable.