BT411/scratchpad/SESSION_HANDOFF_20260803.md

# Session handoff — 2026-08-03 (armor visibility + eject arm chain)

## Committed this session (all on master, exe deployed to content/)
- `fe48acc` **#87 root cause**: zone-level replication send was missing — the binary's
  effect watcher (FUN_0042aa2c) ForceUpdates on BAND-THRESHOLD crossings (master-gated);
  the port only had the gstate branch. Observers saw 0.0 forever. Fixed + 2-node verified
  (replicant peaks == master finals to 4 decimals).
- `440cee1` #87 dig closure: host response is LINEAR (raw re-read @004573e4/@00457784,
  0.1 literal at call site); the pixel curve was i860 DIV-CARD firmware (flush_material
  @0048d4d4 just marshals) — logged in open-questions. Field claim reconciled: one 25-pt
  hit snaps a 77-pt arm 0→0.325 (measured; snap is visible, 2-pt ERS creep is not).
  Bench: BT_WALK_ZONES filter (legwalk.sh), BindArmourDamage ORPHAN inventory (0 orphans).
- `8165cee` **#118 eject alive**: the evaluator's per-frame caller found by E8-scan —
  FUN_004a9b5c (Mech MASTER PERF, export hole)+0x10 calls @0049fa1c every frame. Restored
  at top of Mech::PerformAndWatch (master-gated). Mislabel swept: mech+0x414 =
  ejectPermitted (NOT missionReviewMode; real review mode = GLOBAL DAT_004fd550).
  Verified: gen-kill → panic-arm mode ON → miniconsole Panic button LIGHTS (pixel proof
  panicbtn.png) → punch-out → respawn disarms. BT_EJECT_LOG=1 logs edges.
  Also: msg id 0x1a = DuckRequest@0049fa00 latches mech+0x398 — NO reader in image (vestigial).
- `0089b32` **G = Flush binding** (was: action existed, NO key bound — desktop players
  could never flush!) + BT_FLUSH_HOLD=<frames> scalpel.

## IN FLIGHT — the user's firsthand eject test (they asked, then had to shut down)
Dry run result (content/fl.log): full flush drains the TANK to 0/14 in ~30 s but does
NOT arm eject — InjectCoolant moves tank coolant INTO the loops (bank RISES); idle heat
draw won't burn the loops below the 5% arm threshold in minutes.
**Firsthand recipe to give the user next session** (needs build with 0089b32 — exe in
content/ is current, but the DIST zip is not):
1. Solo (play_solo), get in.
2. HOLD **G** until the COOLANT RES gauge hits 0 (~30 s; flush cloud + audio confirm).
3. Then run HOT: hold fire (lasers) continuously / keep moving. With the tank dry the
   loops burn coolant with no refill. Watch coolant bars fall.
4. When the bank crosses 5%: **the Panic button (bottom-left panel) lights bright** —
   press it (or Backspace = key Back action Eject) → punch-out.
   Faster alternates that also arm: lose all 4 generators; or (novice sim) leg-gimp.
OPEN: how long step 3 takes firsthand (not yet measured — the 3-min idle bench never
armed; sustained fire should be much faster). If it's tediously long, consider whether
weapon heat actually draws LOOP coolant in our reconstruction (#100/#120 territory —
check ConsumeCoolant path under fire before telling testers "minutes").

## Next-session queue
1. Give the user the firsthand recipe above (they explicitly asked to test it).
2. Time-to-arm under sustained fire — bench it (fire scalpel + BT_COOL_LOG) BEFORE
   promising the recipe works in reasonable time.
3. Cut dist 717: carries score fix, 28 Hz, eject core+arm chain, #119, #124 triangle
   pick, #87 replication send, G-flush binding. Testers verify: missile snap darkening
   visible in MP now; eject lamp + punch-out.
4. Gitea: #87 and #118 both commented + awaiting-verification (labels set). #118 note
   says flash-vs-solid lamp is answerable if pod veterans remember blinking (GaugeAlarm
   SetAlertState overlay — not authored for panic in shipped streams as far as checked).
5. Carried: #55, #100 (reservoir drain — RELATED to step 2 above), #120 (flush floor —
   note: our tank DID reach 0/14, so the floor report may be the LOOP levels, not the
   tank), #121-#123, #125-#130; mover-gravity dt-less lead in open-questions.

## Bench inventory added today
- scratchpad/night10/zonewalk.sh (A now has BT_ARMOR_LOG + BT_SHOT_EVERY=zwA) — the
  standing zone-replication receipt bench; digest_walk.py reads it.
- scratchpad/night10/legwalk.sh — leg/foot concentrated audit (BT_WALK_ZONES=leg,foot).
- scratchpad/night10/armorpx.sh — BT_ARMOR_FORCE 0/1 pixel A/B.
- BT_EJECT_LOG, BT_FLUSH_HOLD, BT_WALK_ZONES, armor ORPHAN dump (BT_ARMOR_LOG).
- Proof artifacts in content/: panicbtn.png, forceladder.png, snapdiff.png, legdiff.png,
  dollcrop_*.png (all disposable).

## KB updated (all committed)
combat-damage §Zone-LEVEL replication + snap-vs-creep fidelity boundary;
decomp-reference §4 (effect watcher cluster) + EJECT ARM CHAIN entry;
multiplayer (band-granularity observer model); test-harness (zone-walk armor receipts);
open-questions (i860 DIV-CARD firmware ask).

---
## EVENING SESSION ADDENDUM (the eject sprint + the MFD corruption arc)

### Eject system — COMPLETE and firsthand-verified (commits 8165cee..7872254)
- Arm chain per-frame (FUN_004a9b5c+0x10, E8-scan) [T1]; mislabel swept (mech+0x414
  = ejectPermitted; review mode = GLOBAL DAT_004fd550).
- EjectPilot @0049f854 confirmed via the image message table ('EjectPilot' id 0x19);
  DuckRequest = id 0x1a @0049fa00 (mech+0x398 latch, NO reader — vestigial).
- ⚠ CORRECTED (same night): an earlier "eject was UNFINISHED/dead code in 4.10"
  claim was WRONG — the absolute-pointer scan missed the E8-relative call; the
  master perf (FUN_004a9b5c+0x10) evaluates per frame, so the ARCADE's eject was
  LIVE. Only the PORT was missing the call (our PerformAndWatch recon predated
  knowing the export-hole fn's contents). Swept from code comments same night.
- THE PRESS MODEL [settled]: pilot eject = the PILOT KEYPAD bank while armed
  (+ the PANIC key, which reports through that matrix). The MFD soft keys NEVER
  pilot-eject in the binary — every page routes them to authored functions
  (weapon eng: EjectAmmo; myomer eng: seek toggle) — the flashing engEject cell
  is the INVITE LAMP. Our page-gated eng-key eject is a MARKED PORT CONVENIENCE
  (deviation), operator-requested; drop it if strictness wins.
- Inputs, all live: Panic button (0x3D→pilot-keypad wire), NUMPAD = pilot keypad
  (CONTROLS.MAP `keypad pilot` bindings), Backspace/pad action, and the eng-page
  EJECT key (0x0B/0x23/0x03) page-gated in the RIO drain [T3 intent-completion].
  MFD soft keys otherwise keep authored functions (streamed .CTL: no 0x200000
  records — the 109-record dump is the proof).
- Generator destruction → stateAlarm 4 (GeneratorOut, renamed from placeholder)
  [T3 inferred link] → the AUTHORED invite trail flashes (gotoEngineering 0x80 +
  engEject 0x85, alarm conditions 4/5).
- KeypressMessageID pinned to binary 0x19 (chain arithmetic gave 0x17 — the armed
  press was silently swallowed; two mapper ids 0x17/0x18 remain unreconstructed).

### The MFD corruption arc (gotcha #27) — RESOLVED
Root cause = MY two new draws, not eject logic: (1) the eject-wipe gauge scribbling
a shared ENG/weapon composite plane during eject cycles (re-stubbed, class kept in
btl4gaug for a proper FUN_004c5fb8 draw-op RE); (2) the full-face flash overlays
(L4VB16 + L4GLASSWIN, both REMOVED — the side buttons are big under-surface rects;
the protruding edge IS the lamp; DrawDevSurface's bound texture turned un-reset
fills into gauge-atlas stripes). Pixel-verified clean under full stress; edge lamp
0x27 square-waves 34.0↔28.9 across captures = the flash works at authentic scale.

### Next-session queue (no release went out tonight)
1. Cut dist 717 after a fresh operator pass on 0535a7a-era build.
2. Eject-wipe draw-op RE (FUN_004c5fb8 view vtbl+0x58 arg roles) → un-stub the slot
   graphic.
3. The eng-page label-vs-lamp position question: operator saw the flash beside
   GEN-MODE, not EJECT, on one page — cross-check the DRAWN eng-page layout
   (streamed gauge pages) against kBTEngBankTop math per bank.
4. Authentic generator death (combat crits / overload, not the scalpel) reaching
   state 4/5 — then the [T3] MarkGeneratorOut bridge can be checked against the
   real transition site (MechSubsystem::TakeDamage/OnAlarmChanged cluster RE).
5. Flash presentation: the edge-lamp scale is the design; if the operator wants
   more presence it is a DELIBERATE glass-layer decision (discuss, don't impose).

---
## DAY-2 ADDENDUM (2026-08-04): eject epilogue + the stale-tracker sweep + #108/K-D

### Settled today (all committed, tip 0bd9bb6+)
- Eject FINAL: strict binary model shipped (soft keys never eject; Panic/numpad/
  Backspace do; invite lamps decoded -- conds 4/5 = WEAPON ammo/jam purge invite,
  cond 6 = bus invite via !HasVoltage; hysteresis tried+REMOVED, flap is authentic;
  MarkGeneratorOut bridge removed -- state 4 = the thermal breaker, destruction
  touches nothing electrical @0049c9a8). TechStatus bit table in decomp-reference.
- #119 leak-voice stutter: AUTHENTIC (wooHoo latch authored all-zeros + no armer
  in the image; the hunt at drained tanks clips the phrase-sequenced voice; our
  only real bug was 2x cadence, already fixed). Deterministic leak scalpel added:
  BT_KILL_SUBSYS partial form "Condenser4=0.5".
- Tracker STALE sweep: #80 (crits -- landed a5fb96a 07-29, re-verified live:
  [critroll] chance=0.01+0.7*lvl^2 byte-extracted) CLOSED; #2 (experience --
  wired 07-20, [exp] sentinel live) CLOSED; #22 marked awaiting-verification
  (respawn ammo-refill receipts in tonight's logs).

### IN FLIGHT: #108 ghost recurrence + the K/D accuracy report (SAME ACTION)
- #108 hypothesis: per-entity replication dropout on one node (miss death mode +
  wreck gstate + respawn teleport together -> frozen ghost at the old spot while
  hits register elsewhere). Lead mechanic: the wreck SINK (-8u underground) may
  make the entity interest-culled for distant nodes; if re-interest never fires
  after the respawn teleport, that node's replicant freezes forever. UNPROVEN --
  693 field logs lack entity ids and every discriminating marker is env-gated.
- K/D inaccuracy (night-11 report, build 716): the tally/credit receipts
  (PLAYER_DEAD matchlog rows, [score] panel DRAW probe) are ALL env-gated off in
  field builds -- logs show only roster rebuilds. Cannot confirm or deny from
  the field data. NOTE: K/D columns are replicated state; a #108-style stream
  dropout would corrupt both ghosts AND tallies on the same node -- one cause.
- THE ACTION (next block, before cutting 717): make field logs self-diagnosing:
  1. players/play_steam.bat (+join*.bat): set BT_MATCHLOG=1 (or a curated
     death/score subset if per-shot DMG volume is too hot -- check matchlog.hpp
     categories) + BT_SCORE_LOG=1.
  2. Entity IDs on the ungated render lines (wreck swap / MakeMechRenderables /
     RemakeEntity -- btl4vid).
  3. NEW ungated [ghost] staleness detector: a replicant whose last-applied
     update record is > N s old while not buried -> one line with entity id +
     age + last-known position (the #108 indictment line).
  4. Un-gate (or default-on) the respawn un-wreck line.
  Then 717's night-11 logs convict #108's mechanism and settle the K/D claim.

### Next-session queue (post-717-cut)
1. The instrumentation block above -> cut 717 -> field night.
2. #35 Owens laser crash; #126 invisible peer fire; #89 splash; #92 Loki feet;
   #73 (check against the #124 triangle-pick fix -- likely closable).
3. #118 last tail: MP console notice delivery audit.

### #108 CONFOUND (operator recall, 2026-08-04): Conn Man ran a WRONG BUILD for
much of night 9 -- raw-struct replication + version skew can desync one
connection's stream (one node loses one entity: the exact ghost signature), and
same-build benches can't reproduce (59 clean cycles again today). K/D bookkeeping
verified EXACT both columns (4-node deaths + 2-node cross-player kills,
matchcheck-reconciled). NEW 717 item: BUILD HANDSHAKE at join/relay (version
already in matchlog HDR) -- refuse mismatched builds loudly. Instrumentation
block unchanged (it discriminates skew vs real bug).