Follow-up to 4fa7eee, driven by an adversarial review pass and three rig cycles.
Each item below is a real defect that pass found, not polish.
DEATHS now uses the binary's own column. PilotList::Execute @0x4cabd0 draws
`fild [edi+0x27c]` (KILLS) and `fild [edi+0x280]` (DEATHS); +0x280 has exactly
two runtime writers image-wide plus the ctor zero. Our port had declared it
`pad_0x280`, never written, and displayed the ENGINE's Player::deathCount
(+0x200) instead -- which is the respawn-handshake identity, seeded -2, and is
why it needed a clamp to pass as a count. It is now BTPlayer::deathTally (our
offset 0x274, offsetof-locked), incremented beside ++deathCount, read by the
gauge, and replicated. deathCount is left to the engine's handshake. So this
half moves TOWARD the binary; it also closes the plan's Headline-1.
(`deathTally`, NOT `deaths`/`deathCount` -- those would shadow the base.)
The mirror was not self-healing: update records are UNRELIABLE by construction
(Entity::UpdateMessage clears ReliableFlag, ENTITY3.h:112; the relay's UDP path
also drops stale/reordered datagrams), and the pair was dirtied only on an
event. One lost datagram left every peer stale until that pilot's next kill or
death -- forever for the last kill of a round. Added a 2s heartbeat that
re-dirties the record, which also bounds how long the binary's phantom
partner-increment stays visible. The timer is a function static deliberately: a
data member would change sizeof(BTPlayer) and break the offset locks.
Also fixed: the SBMIRROR row AND its change-detect both still read deathCount (a
constant -2 here), so the "log only the edge" guard could never be false and
every row printed deaths=-2. That is what made the first rig runs look like
DEATHS was broken when a WRITE/READ trace proved the transport correct. Row
count per node fell from ~20 to 3, one per real change.
Guarded the record against per-bit layouts: update_model is a BIT INDEX and
Entity::WriteUpdateRecord switches on it (ENTITY.cpp:329-352) -- the DamageZone
bit emits a variable-length packed stream whose length it computes itself, so
appending two ints and re-stamping recordLength over that would corrupt it.
Deleted BTPlayerCountObservedDeath -- definition, call site, extern and friend
together, since /FORCE hides stragglers. It never executed (its call site sat
inside the once-per-death transition, which a replicant never enters) and could
not have worked (0 of 8800 corpus DMG rows target a replicant, which is why every
DEATH inst=R row reads killer=0:0). Under replication it would have been a
second writer of a replicated counter.
New forensics: NOCREDIT names the failing link when a kill credit is skipped (it
used to be completely silent -- the counter simply never moved), and PLAYER_LINK
records whether the one-shot link resolved. Both retire the NULL-playerLink
theory: every rig shows `PLAYER_LINK inst=R resolved=1` and no NOCREDIT rows.
PLAYER_DEAD now logs both counters (deaths=handshake, tally=scoreboard).
KB sweep of the claims that hid this bug for so long:
* context/gauges-hud.md's "RESOLVED -- deaths tally per node from locally
observed events" was FALSE; corrected with the measured evidence.
* docs/GAUGE_COMPOSITE.md + btl4gau3.cpp "the dead pad_0x280" -- never dead,
merely unwritten.
* btplayer.hpp attributed VehicleDeadMessageHandler to @004c012c; it is
@004c05c4 (absent from the decomp export -- the #60 gap).
* docs/RESPAWN_REARM_PLAN.md's "#45 SUBSUMED" -- the PLAYER_DEAD symptom was
subsumed, the scoreboard defect was not.
* The corpus figure in the record banner now states its method so it is
reproducible.
Rig-verified (2-node loopback, several cycles): owner 3:1 kills 2/tally 1 read
`kills=2 deaths=1` on the peer; owner 2:1 kills 1/tally 2 read `kills=1 deaths=2`
on the peer. Respawn unaffected, no crash, no GLITCH rows. Still awaiting live
multi-pod verification by a human; all pods must run the same build.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
