Files
Joe DiPrimaandClaude Opus 5
277337c3a5
tracker: date undated tester entries from gitea evidence, and record the self-reference trap. 12 entries pinned to real dates by finding where their text is quoted in a ticket body or an earlier comment -- if a report is quoted in a body, it existed by that ticket's creation date, which is how #170's three crash lines pin to 2026-08-12 (its body says FIELD 2026-08-12 and lists those exact times, settling that they are the ORIGINAL occurrences and not a post-fix recurrence). 27 entries that nothing in gitea pins stay marked (undated) rather than guessed. THE TRAP, hit once during this pass: dating an entry from a comment written while harvesting the sheet -- those quote the tester text verbatim, so a naive search dates the report to your own housekeeping instead of the event (a #170 line briefly carried the harvest date). Exclude same-day comments and take the earliest match.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
#125 FIXED + 2-node benched: peers now see a killed arm WRECK + its gun pod DROP OFF on the FIRST hit. Root pair: (1) the MechDeathHandler Tick transcription carried an INVENTED write -- ApplyDamageGraphicState(descriptor->graphicState) -- where the raw @0042aa2c (part_003.c:10963-93) only PICKS an effect resource (level flag & 4 -> DescriptorForLevel; gstate flag & 8 -> DescriptorForGraphicState(the zone's CURRENT gstate), overwriting the level pick); with the port's pick hardcoded to Destroyed(1) the invention STOMPED every destruction-cascade child's authored Gone(2) back to 1 in the same tick, master-side, BEFORE the record shipped -- the wire itself carried the corruption, so no peer could ever render the pod dropping (receipts: legacy A remade jointrgun to gstate 2 then re-remade it to 1; legacy B received 'dz_rgun gstate 0->1'). Gotcha 36: an added write is worse than a stub -- it actively fights the authentic writers on every node. (2) the destroyed-mesh swap trigger was descriptor-gated (DescriptorForLevel returns NULL past the last threshold; cascade children have gstate edges with NO level edge) -- the binary needed no trigger at all (its renderer re-picked every segment's mesh from zone gstate PER FRAME); the port's built-once render tree now gets the event equivalent: per-zone gstate caches in Tick fire ONE coalesced RemakeEntityRenderables on any edge, which is also the PEER's only trigger for replicated gstate-only beats. Restored authored two-stage story: a directly-killed gun zone hangs its wrecked variant mesh (thrdrgun-family, exterior skeleton; every zone's descriptor table carries lvl=-1 gs=1/gs=2 terminals -- unreachable by level, existing exactly for the gstate picks); a killed ARM swaps to its wrecked variant AND the pod segment vanishes. Effects suppressed on ->Exists edges (the binary's heal writes the cell raw, no flag -> no respawn puff; regression benched 0 heal puffs across a full death/respawn cycle). 2-node benched both ways (night18/mp_zone125.sh + zonedump.sh): fix = peer receives dz_rarm gstate 0->1 + dz_rgun 0->2 on ONE beat, swaps thrdrarm.bgf and HIDES the pod, first-beat remake receipt; legacy = the stomp conviction above. KB: combat-damage.md MechDeathHandler section rewritten to the corrected reading; gotcha 36.
#102 FIXED + benched: the Owens run->reverse 'hang' was the port missing the machine's whole HARD-BRAKE system -- master perf @0x4aa158-0x4aa365 (raw part_013.c:3353-95): a forward gait meeting a negative demand latches +0x3f4 (the recon's 'airborneSelect' -- another jump-jet-family misread: it is the BRAKE flag) and swaps the gait slew to the model's SuperStopAcceleration (+0x5bc, ex-'airborneCycleRate'), kicks the gyro with a random-sign Z lurch, judders the cab 0.2 every 0.4s while braking above walk speed (gyroRumbleTimer@0x5c4), ships the flag on the type-8 record, restores the ground rate on release. Content: heavies brake harder than they drive (10->20); the sprinters brake WEAKER -- owens 50->25, blkhawk 30->10 -- the Owens' long coast into reverse is AUTHORED, and the pod communicated it through the seat; the un-ported port coasted with zero feedback = the felt hang. Ported to the mech4 drive block ([brake] receipts under BT_GAIT_TRACE) + the replicant rate select restored in the peer motion path (the binary's IntegrateMotion:298 line -- dead TU in the port, Simulate bypassed). Plus gotcha 35: the binary's SM switch re-reads the state cell after the wind-down-gate/death-latch preamble (one cell: @0x3b0/@0x728 == the alarm level); the recon's split alarm/int pair dispatched the STALE pre-gate state for a frame -- re-synced, both channels. The 'leg in the air until you twist' statue is the MACHINE's own trn turn-stop MID-PIVOT park, authentically INVISIBLE on the pod (no external self-view existed; the chase cam is a port addition) -- bench-reproduced at trn legFrm=10 + twist rescue, documented not-a-bug (Owens reads worst: deepest mid-swing of the 8 chassis, thigh -43deg vs madcat -27). Reverse-input truth [T1]: reverse is the red button 0x3F/LALT held with the throttle OPEN (manual p7); demand = -runSpeed*throttle (NO fwdScale term in reverse); the throttle channel is [0,1] like the pod lever, so brake keys can NEVER produce reverse -> handout education line. Bench rig: BT_KEYSIM/_REV/_TURN drive the REAL lever+detent+SM chain headlessly (scratchpad/night18/gaithang_bench.sh). Benched (owens/grass): LALT flip at full run -> [brake] ENGAGE rate=25 (ground 50) -> 12->15->6->9->0->16->18<->19 back cycles at the authored -14.1 cadence -> clean recovery on release; held brake key -> authored stand park, ZERO brake engagements (authentic: no button, no brake). Side-find filed: hardware-RIO cab reverse is DEAD (AddOrErase stubs; open-questions row). 674 forensic: that build boots GLASS by default -> reverse was exclusively LALT in the field.
Oracle-return verification pass, both CONTRADICTS run to ground against the binary same night: #176 the seek-gear square law is the Emitter ctor's OWN calibration (@004bb120 EC = E/(0.5 seekV[rec]^2) [T1]) and the authored table (LLaser {6000,7000,8000,9900}/10000, rec=8000) yields the REAL ladder 56/77/100/153 -- seek 3 IS the published calibration point exactly as Oracle said, seek 4 is an unclamped OVERCHARGE gear (logged 5.80 > authored 5.0 proves it), and the guide's 'half-charge = quarter' example was the wrong part (master guide corrected with the real ladder + the downshift-for-rate / seek-4-alpha doctrine). #177 the BTPlayer ctor stamp table read verbatim (@004c0bc8: +0x260 = level>=2 -> heat at veteran AND expert in 4.10 [T1]; +0x274 = the raw level; +0x26c = standard|veteran) -- the port is faithful; Oracle + the 4.0 manual describe the earlier machine (a genuine 4.0->4.10 balance change); ticket premise corrected (the KB already carried the stamp values -- 'writer never read' was another check-own-history strike). experience-levels.md upgraded to the writer address; #176 open residue: per-weapon gear-fraction variance.
radar audit ledger: docs/RADAR_MAP_AUDIT.md distills the 3-agent audit (architecture, authenticity verdicts, the #49 root+fix, deliberate port deviations, stale-note sweep, follow-ups incl. the arena-edge era question and the 45deg-wedge-vs-manual-60 flag); gauges-hud gains the pointer section; the mech.cpp 'SetTargetRange is stubbed' stale comment swept (live since the zoom work).
#102 FIXED + benched: the Owens run->reverse 'hang' was the port missing the machine's whole HARD-BRAKE system -- master perf @0x4aa158-0x4aa365 (raw part_013.c:3353-95): a forward gait meeting a negative demand latches +0x3f4 (the recon's 'airborneSelect' -- another jump-jet-family misread: it is the BRAKE flag) and swaps the gait slew to the model's SuperStopAcceleration (+0x5bc, ex-'airborneCycleRate'), kicks the gyro with a random-sign Z lurch, judders the cab 0.2 every 0.4s while braking above walk speed (gyroRumbleTimer@0x5c4), ships the flag on the type-8 record, restores the ground rate on release. Content: heavies brake harder than they drive (10->20); the sprinters brake WEAKER -- owens 50->25, blkhawk 30->10 -- the Owens' long coast into reverse is AUTHORED, and the pod communicated it through the seat; the un-ported port coasted with zero feedback = the felt hang. Ported to the mech4 drive block ([brake] receipts under BT_GAIT_TRACE) + the replicant rate select restored in the peer motion path (the binary's IntegrateMotion:298 line -- dead TU in the port, Simulate bypassed). Plus gotcha 35: the binary's SM switch re-reads the state cell after the wind-down-gate/death-latch preamble (one cell: @0x3b0/@0x728 == the alarm level); the recon's split alarm/int pair dispatched the STALE pre-gate state for a frame -- re-synced, both channels. The 'leg in the air until you twist' statue is the MACHINE's own trn turn-stop MID-PIVOT park, authentically INVISIBLE on the pod (no external self-view existed; the chase cam is a port addition) -- bench-reproduced at trn legFrm=10 + twist rescue, documented not-a-bug (Owens reads worst: deepest mid-swing of the 8 chassis, thigh -43deg vs madcat -27). Reverse-input truth [T1]: reverse is the red button 0x3F/LALT held with the throttle OPEN (manual p7); demand = -runSpeed*throttle (NO fwdScale term in reverse); the throttle channel is [0,1] like the pod lever, so brake keys can NEVER produce reverse -> handout education line. Bench rig: BT_KEYSIM/_REV/_TURN drive the REAL lever+detent+SM chain headlessly (scratchpad/night18/gaithang_bench.sh). Benched (owens/grass): LALT flip at full run -> [brake] ENGAGE rate=25 (ground 50) -> 12->15->6->9->0->16->18<->19 back cycles at the authored -14.1 cadence -> clean recovery on release; held brake key -> authored stand park, ZERO brake engagements (authentic: no button, no brake). Side-find filed: hardware-RIO cab reverse is DEAD (AddOrErase stubs; open-questions row). 674 forensic: that build boots GLASS by default -> reverse was exclusively LALT in the field.
#102 FIXED + benched: the Owens run->reverse 'hang' was the port missing the machine's whole HARD-BRAKE system -- master perf @0x4aa158-0x4aa365 (raw part_013.c:3353-95): a forward gait meeting a negative demand latches +0x3f4 (the recon's 'airborneSelect' -- another jump-jet-family misread: it is the BRAKE flag) and swaps the gait slew to the model's SuperStopAcceleration (+0x5bc, ex-'airborneCycleRate'), kicks the gyro with a random-sign Z lurch, judders the cab 0.2 every 0.4s while braking above walk speed (gyroRumbleTimer@0x5c4), ships the flag on the type-8 record, restores the ground rate on release. Content: heavies brake harder than they drive (10->20); the sprinters brake WEAKER -- owens 50->25, blkhawk 30->10 -- the Owens' long coast into reverse is AUTHORED, and the pod communicated it through the seat; the un-ported port coasted with zero feedback = the felt hang. Ported to the mech4 drive block ([brake] receipts under BT_GAIT_TRACE) + the replicant rate select restored in the peer motion path (the binary's IntegrateMotion:298 line -- dead TU in the port, Simulate bypassed). Plus gotcha 35: the binary's SM switch re-reads the state cell after the wind-down-gate/death-latch preamble (one cell: @0x3b0/@0x728 == the alarm level); the recon's split alarm/int pair dispatched the STALE pre-gate state for a frame -- re-synced, both channels. The 'leg in the air until you twist' statue is the MACHINE's own trn turn-stop MID-PIVOT park, authentically INVISIBLE on the pod (no external self-view existed; the chase cam is a port addition) -- bench-reproduced at trn legFrm=10 + twist rescue, documented not-a-bug (Owens reads worst: deepest mid-swing of the 8 chassis, thigh -43deg vs madcat -27). Reverse-input truth [T1]: reverse is the red button 0x3F/LALT held with the throttle OPEN (manual p7); demand = -runSpeed*throttle (NO fwdScale term in reverse); the throttle channel is [0,1] like the pod lever, so brake keys can NEVER produce reverse -> handout education line. Bench rig: BT_KEYSIM/_REV/_TURN drive the REAL lever+detent+SM chain headlessly (scratchpad/night18/gaithang_bench.sh). Benched (owens/grass): LALT flip at full run -> [brake] ENGAGE rate=25 (ground 50) -> 12->15->6->9->0->16->18<->19 back cycles at the authored -14.1 cadence -> clean recovery on release; held brake key -> authored stand park, ZERO brake engagements (authentic: no button, no brake). Side-find filed: hardware-RIO cab reverse is DEAD (AddOrErase stubs; open-questions row). 674 forensic: that build boots GLASS by default -> reverse was exclusively LALT in the field.
KB: the playtest tracker sheet is now a first-class topic (context/playtest-tracker.md) -- the testers moved issue tracking off the .txt handout onto a shared Google Sheet on 2026-08-14, making it the primary feedback channel with gitea still the system of record. Documents the two-tab shape (OPEN priority-ordered / CLOSED), the dev-status vocabulary with the colour EMBEDDED IN THE LABEL, the priority tiers, and the Dev-Comments contract (what is wrong / STATUS / NEED FROM YOU, written for a player in under ~55 words with a concrete ask). Captures the working rules that were learned the hard way this session: read the gitea COMMENT THREAD never the body (re-triaging 53 tickets from threads moved a third of the statuses and dropped P1 from 5 to 2); undated tester rows are HISTORICAL and reading them as fresh produced a false 'still crashing after the fix' claim on #170 that had to be retracted across seven tickets; never write DONE on the OPEN tab; merge duplicates to one row per ticket; never use cell fills for status (fills live on cells, labels on text -- every re-sort desyncs them, re-fixed three times before embedding); a leftover FILTER hides rows after a rewrite (skipping row numbers in the gutter = hidden rows, not lost data); pin row heights or autofit makes wrapped rows unreadably tall. Access documented both ways: read via the public CSV export (the Claude Drive connector CANNOT see it -- scoped to files it created, and it has no cell-write capability at all, so sharing does not help), write via an Apps Script bridge whose URL+token stay in local operator memory and NEVER in this repo. Quick Lookup row added; handout demoted to secondary; operator-console cross-linked. checkctx CLEAN.
tracker: date undated tester entries from gitea evidence, and record the self-reference trap. 12 entries pinned to real dates by finding where their text is quoted in a ticket body or an earlier comment -- if a report is quoted in a body, it existed by that ticket's creation date, which is how #170's three crash lines pin to 2026-08-12 (its body says FIELD 2026-08-12 and lists those exact times, settling that they are the ORIGINAL occurrences and not a post-fix recurrence). 27 entries that nothing in gitea pins stay marked (undated) rather than guessed. THE TRAP, hit once during this pass: dating an entry from a comment written while harvesting the sheet -- those quote the tester text verbatim, so a naive search dates the report to your own housekeeping instead of the event (a #170 line briefly carried the harvest date). Exclude same-day comments and take the earliest match.
#160 CLOSED -- the cab's RIO axis dropouts + frozen boots were the FTDI adapter's BUFFER SIZE set too high (Nick, on the cab). An oversized receive buffer batches incoming serial instead of delivering it promptly: analog samples arrive late and in clumps (the field 'encoder dropouts') and the boot handshake can miss its window entirely (the wedge). NO port code was ever at fault -- every software hypothesis in the pod-hardware RIO section (power-rail sag under mission load, board-side wedge, print deadlock) is superseded; what survives from those notes is the diagnostic procedure, not the theories. Recorded in context/pod-hardware.md 9a with the setting path (Device Manager -> COM port -> Port Settings -> Advanced -> lower Receive (Bytes); Latency Timer flagged [T4] as the adjacent standard knob, not confirmed part of the fix) and as an operator-facing troubleshooting row in docs/OPERATOR_GUIDE.md, because a fresh FTDI install on any future cab build comes up with defaults that reproduce it. Ledger lesson: a fault that reproduces on ONE machine and survives every code change is a DRIVER/CONFIG suspect before it is a code suspect.
#125 FIXED + 2-node benched: peers now see a killed arm WRECK + its gun pod DROP OFF on the FIRST hit. Root pair: (1) the MechDeathHandler Tick transcription carried an INVENTED write -- ApplyDamageGraphicState(descriptor->graphicState) -- where the raw @0042aa2c (part_003.c:10963-93) only PICKS an effect resource (level flag & 4 -> DescriptorForLevel; gstate flag & 8 -> DescriptorForGraphicState(the zone's CURRENT gstate), overwriting the level pick); with the port's pick hardcoded to Destroyed(1) the invention STOMPED every destruction-cascade child's authored Gone(2) back to 1 in the same tick, master-side, BEFORE the record shipped -- the wire itself carried the corruption, so no peer could ever render the pod dropping (receipts: legacy A remade jointrgun to gstate 2 then re-remade it to 1; legacy B received 'dz_rgun gstate 0->1'). Gotcha 36: an added write is worse than a stub -- it actively fights the authentic writers on every node. (2) the destroyed-mesh swap trigger was descriptor-gated (DescriptorForLevel returns NULL past the last threshold; cascade children have gstate edges with NO level edge) -- the binary needed no trigger at all (its renderer re-picked every segment's mesh from zone gstate PER FRAME); the port's built-once render tree now gets the event equivalent: per-zone gstate caches in Tick fire ONE coalesced RemakeEntityRenderables on any edge, which is also the PEER's only trigger for replicated gstate-only beats. Restored authored two-stage story: a directly-killed gun zone hangs its wrecked variant mesh (thrdrgun-family, exterior skeleton; every zone's descriptor table carries lvl=-1 gs=1/gs=2 terminals -- unreachable by level, existing exactly for the gstate picks); a killed ARM swaps to its wrecked variant AND the pod segment vanishes. Effects suppressed on ->Exists edges (the binary's heal writes the cell raw, no flag -> no respawn puff; regression benched 0 heal puffs across a full death/respawn cycle). 2-node benched both ways (night18/mp_zone125.sh + zonedump.sh): fix = peer receives dz_rarm gstate 0->1 + dz_rgun 0->2 on ONE beat, swaps thrdrarm.bgf and HIDES the pod, first-beat remake receipt; legacy = the stomp conviction above. KB: combat-damage.md MechDeathHandler section rewritten to the corrected reading; gotcha 36.
#170 root-caused + fixed: the #163 SteamAPI_Shutdown was the crash -- at round end every peer's close events are ALWAYS in flight, and Shutdown tears client state down while steamclient's own service thread dispatches them (the field stack: a steamclient worker calling literal NULL, zero btl4 frames, 7 hits / 3 machines / one module-relative stack). BTSteamNet_ShutdownAll is now a QUIESCE: steamActive gated FIRST, linger closes ('round over' not timeouts), one SNS drain, 60ms grace, NO SteamAPI_Shutdown -- and every exit path dies via TerminateProcess (menu-quit no longer returns through the CRT; it uses BTOrderlyDie like everything else, so the 883 detach deadlock stays dead too). Status callback guarded for the dying window. Crash filter now names foreign DLL frames (module+offset via VirtualQuery/GetModuleFileName) so the next foreign-thread crash arrives pre-symbolized. Exit contract: steam-networking.md
#173 FINAL: port cadence defect FIXED -- the velocity ring now samples at the pod's 28Hz by default (BT_MYO_RING_HZ overrides). Measured [T2, six curves m173v2]: at ~59fps the |a| operand carried 64-66% gait-aliasing noise = drive heat 2x the machine's; at 28Hz + the authored loop-5 boost (detent 50 receipt-proven, flow 0.909) a vulture holds FULL supercharge 54-57 u/s indefinitely at stable Tm~1500 -- the machine's valve lever worked, the manual's Super Charged print was honest, witnesses 7-0 (Draco/Oracle/Elengil + the user's eyes beat the solver AND the first bench). Balanced seek-4 still cooks authentically (~18s); wrap trap stands; display pixel-verified in sync. BT_VALVE5 bench hook added (receipted detent presses). Post-promotion smoke: boosted vulture stable ~1500 dipping under. Gotcha 32 (sampled-history operands at render cadence; a validated model EXTRAPOLATES, only the bench INTERPOLATES); cadence census -> BTSimClock migration checklist in open-questions (gyro bounce timescale next, footstep smoothers, doppler; heatLoad was already pod-clocked #119 = the reference pattern). Guide 2B rewritten with the confession + the measured fix
weapons batch VERIFIED on the final tree + last two gaps closed: BYSTANDER SWEEP landed (guided damage rounds test the flight segment vs every mech, detonate on interposers like the binary's world sweep @0042291c -- 11 legitimate strikes in the sweep bench, zero self/phantom/fuze-race hits per forensics) + COAST GRAVITY (-6.5 after burnout). Comprehensive 2-node weapons_sweep bench PASS, all families: ENERGY 184 named FIRED / 79 hits; AC 17 type=1 hits ALL burst=1 + 84 ballistic pushes ALL dmg=0 (hitscan live-proven multi-node on final code); MISSILE 593 type=2 hits, 576 salvo bursts (4.0 fuze + drag speeds work); #174 guard worst-repeat=1 across 13 lives; detector SELF-TESTED against synthetic pre-fix shapes before any verdict trusted. Rig lessons -> test-harness.md: BT_FIRE_AT_ICON designates BUILDINGS (use BT_GOTO=enemy + throttled BT_AUTOFIRE/BT_AF_MISSILE, the kd pattern); mad2 'Zanin Neko' carries AFC50+ERPPC+LRM15+SRM6x2+ERMLaser -- all families on one engaging chassis; -net 1501 -> pilot :1502 mapping
KB: #111 paperwork -- translocation-warp topic updated (peer sphere retired: port extension vs the POV-only authentic effect; the DropZoneLocation-anchor nuance and the slot-contention skip both moot; observers see the un-wreck swap). Release notes gain the no-bubble-on-others line crediting Lynx's night-9 call. Tracker comment posted, field-verify Friday.