Gameplay variable audit: the tables were right, nothing was reading them

[themuffinator]

An audit of every gameplay number in the port — player and creature health,
weapon damage, armour, ammo — checked against the disc it was read from.

The data half is clean

No changes were needed there, and that is the headline. Re-run against
SLES_015.34:

q2psx-inspect weapons  -> PASS  11 weapons, 3 armour classes, 22 sounds, 0 mismatches
q2psx-inspect classes  -> PASS  every one of 651 spawns names a class in the table
q2psx-inspect items    -> PASS  1013 place records resolve; 0 with no handler
q2psx-inspect ai       -> PASS  130 checks, 0 mismatches
q2psx-inspect death    -> PASS  every check passed

The per-weapon damage, quad, pellet, spread, kick and muzzle figures in
weapon.c agree with FORMATS.md §13.1 row for row; the ammo capacity tiers and
all six armour constants land on PC Quake II's own numbers.

Every defect was in the wiring

Six of them, and they share one shape: a field that only ever held its default.

where	was	now
q2_actor_from_player	armour_class a literal 0	from the inventory
q2_actor_from_player	powerups never copied	inv->flags
q2_sim_fire	quad hardcoded false	level_time < inv.quad_until
combat rules	skill frozen at 1	q2_cre_skill()
client frame	sim.multiplayer written only by the save loader	follows mp_enabled
player thresholds	gib at −100 and −40	one copy, Q2_PDEATH_GIB_HEALTH

Armour was always the weakest kind. armour_class was written as 0, and 0
is jacket — so combat and body armour absorbed 0.30 of an ordinary hit instead
of 0.60 and 0.80, and neither absorbed anything from an energy weapon, where
jacket's column is genuinely zero. The projection rebuilds on every damage
attempt, so no pickup or save could get a real class past it. The power shield
was worse off: the two bits 0x80057AC4 tests live in the inventory's flag
word, which was never copied, so absorption returned at its first guard.

Quad multiplied nothing. Every fire function picks one of two immediates by
comparing the level clock against the player's expiry word, and the sim called
them with that comparison hardcoded false. The powerup was picked up, drawn on
the HUD, counted down and played its firing sound for thirty seconds while the
shotgun kept doing 6 a pellet.

Skill and deathmatch never arrived. Only the clock was refreshed in the
combat rules, so skill stayed 1 (easy as dangerous as medium) and
deathmatch stayed false. q2_sim.multiplayer had one writer — the save
loader — so a match started from the menu ran every DM rule in its
single-player form: railgun 100 instead of 150, the armour bias that favours
the wearer, single-player item amounts with no respawn, and the rocket-jump
ceiling applied where the console does not apply it.

Measured

Through the real damage path, 200 armour held, one 100-point hit — the share
absorbed:

	before	after
jacket, normal	31%	31%
combat, normal	31%	61%
body, normal	31%	81%
body, energy	0%	61%
power shield (30-pt hit)	0%	67%, at one cell per two points

A player in body armour was taking 69 points of a 100-point hit. They should
have been taking 19.

Also corrected

Three comments the audit found drifting from their code: weapon.h's refire
note described a 30-tick gate against a table that is deliberately all zeros
(both are right — the write exists, nothing reads it back, the fire clip is the
rate); weapontables.c called the stored muzzle triple (right, up, forward)
where the header and loader either side of it both say down; and the corpse
health floor was spelled out twice under two names.

Testing

35/35 tests pass, including 10 new checks — six in test_combat for the damage
projection, four in test_sim for the session reaching the shot. Client and
tools rebuild clean with no new warnings. Rendering is byte-identical between
the two builds: the same frame comes out of both, which is precisely why these
survived — none of it is visible until something hits you.

For the reviewer

The one figure that had no instruction behind it is now settled. `(damage + 1)

1for the skill-0 halving rounds half up where a MIPS signed/2`
truncates, so it looked like it might be a transcription slip. Read off the
disc with the repository's own MIPS decoder, it is exactly right:

80057D5C  addu  s5, a0            ; attacker
80057D64  addu  s2, a1            ; target
80057D6C  addu  s1, a2            ; damage
80057D94  lw    s0, 12(s2)        ; the target's client block
...
800582C8  beq   s0, zero, ...     ; the target has a client
800582D0  lh    v0, 0x800B334A    ; the skill halfword
800582D8  bne   v0, zero, ...     ; only at skill 0
800582E0  lw    v0, 748(s5)       ; attacker+0x2EC, its entity back-pointer
800582E8  beq   v0, zero, ...
800582F0  lw    v0, 12(s5)        ; the attacker's client block
800582F8  bne   v0, zero, ...     ; only when it has none: a monster
800582FC  addiu v0, s1, 1         ; delay slot: damage + 1
80058300  sra   s1, v0, 1         ; damage = (damage + 1) >> 1

The +1 is unconditional. A compiler's signed divide by two would have emitted
the sign-bit idiom instead (srl v0, x, 31; addu; sra) and there is no such
term, so the source really was (damage + 1) >> 1. 25 points of monster damage
arrive as 13, not 12 - which also gives id's "never rounds to nothing" for
free. Transcribed into combat.c with the four guards either side of it; no
behaviour change, the citation is the point.

That transcription records one thing the port does not reproduce:
0x800582E0 additionally requires attacker+0x2EC - the actor's entity
back-pointer, zeroed at 0x8007F0DC when a body stops being a creature - to be
non-zero, so the console refuses the halving for an attacker that has become a
corpse. attacker != NULL stands in for it, and nothing in this port damages
anyone from a detached actor.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.