cbc5ff95321b800a8a63baa04d1258489b69dae9
15
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 arcattackandClaude Fable 5
|
6ad1e074cc |
Glass lever: port the pod-build GAIT DETENT to the PadRIO throttle slew
Human re-test on the fresh build: zero-stop OK, but the lever could PARK inside
the walk/run hysteresis band (walkStrideLength@0x534 .. reverseSpeedMax@0x538)
where the gait SM has no stable state -- walk<->run hunt, EngineShiftFwd/Rev
retriggering ("mixed walk/run mode + repeated acceleration sample"). The
pod-build keyboard lever already snaps out of the band to the NEARER edge at
key-rest (mech4.cpp GAIT DETENT); the PadRIO slew channel lacked it.
Faithful port, not an invention: mech4.cpp publishes the band in lever units
every driven frame (gBTGaitDetentLo/Hi seam, same guards + the same +0.002
engage margin); L4PADRIO Poll() applies the identical nearer-edge snap to the
Throttle channel when the slew rests (no slew key held, no pad slew axis
deflected). Sweeping through the band while held stays continuous (the
authentic moving lever, one authentic shift).
Verified (DEV.EGG glass, graduated parks each ~0.10 of the sweep,
BT_GAIT_TRACE+BT_MPPR_TRACE): band [0.358026,0.50388) = demand 22.02-30.87;
in-band parks snap both directions (0.4137->lo, 0.465126->hi); no at-rest
sample in the band; exactly ONE walk->run clip (state 10) up and ONE run->walk
(state 14) down across the whole sweep; zero-stop unaffected; pod build
rebuilt + BT_AUTODRIVE smoke clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
||
|
arcattackandClaude Opus 4.8
|
aa004eb7b8 |
Gait: keyboard detent snaps the lever out of the walk/run DEAD BAND + KB
Byte-decoded finding [T1]: the gait SM has no stable state for a demand in
(walkStrideLength@0x534, reverseSpeedMax@0x538). Cap semantics settled from
LoadLocomotionClips @0x4a80d4: 0x538 is the walk->run TRANSITION CLIP's exit
speed (the run ENGAGE threshold -- not a reverse max) and 0x34c (run-cycle avg
root speed) is the mapper's continuous demand multiplier (FUN_004afd10). The
finished-callbacks up-shift when tgt > 0x534 but enter/sustain run only when
tgt >= 0x538, so a demand parked between them (Blackhawk: 22.02-30.87 =
throttle 36-50%) hunts walk -> shift-up -> shift-down forever, firing the
authored EngineShiftFwd/Rev sounds each swing -- the user's repro, faithfully
reproduced by all-authentic logic + data. Likely masked on the pod by
MECHANICAL throttle-quadrant detents ("5 speeds" lore; the software path is
notch-free) [T4] -- added to get-from-Nick.
Accommodation (keyboard = our stand-in lever): at key REST, snap the lever to
the nearer dead-band edge (walk cap, or run engage + margin since the cont
check is >=). Sweeping THROUGH the band while held stays continuous -- an
authentic moving lever; the single shift it fires is the authentic shift.
BT_GAIT_TRACE logs [gaitdetent] snaps.
KB: locomotion.md gait-dead-band section, pod-hardware.md mechanical-notch
hypothesis, open-questions.md Nick item.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
||
|
arcattackandClaude Opus 4.8
|
82e6bb306a |
KB: sweep STALE "unfinished" claims -> match landed commits (context-system audit)
4-agent audit cross-checked every open/pending/stub/deferred claim in the context system against git + code. Fixed ~13 stale spots (currency only, no code change), each contradicted by its own file body + a landed commit: - 0xBD3 SubsystemMessageManager (task #7, |
||
|
arcattackandClaude Fable 5
|
a83995a755 |
Locomotion: wire the AUTHENTIC per-mech turn-rate lerp (task #64b)
The master-perf disasm (0x4aa3d3-0x4aa4ff) computes the yaw rate as lerp(walkingTurnRate@0x574, runningTurnRate@0x578) by ground speed, not a constant. Wired it into the drive, replacing the bring-up constant kDriveTurnRate: - walkingTurnRate/runningTurnRate are now REAL named members. The ctor's `Wword(0x15d)/(0x15e) = model->...TurnRate * DegreesToRadians` writes were NO-OPS -- Wword() returns a shared static scratch cell, so the rates were silently discarded. Now stored + read by name. - The drive (mech4.cpp) computes authTurnRate: base = walkingTurnRate; above reverseSpeedMax(0x538) lerp walk->run across [walkStride 0x534 .. topSpeed 0x34c] with an over-run falloff runningTurnRate/t^2; clamp >=0; zeroed in the death/limbo/airborne leg states (1/2/3) or when deathAnimationLatched(0x650) is set. Falls back to kDriveTurnRate if the model gave no rates. The lerp is a RUN-speed refinement -- below reverseSpeedMax (walk / turn-in-place) the rate is the constant walkTR base; runningTurnRate < walkingTurnRate (a running mech is less maneuverable). Verified headless (DEV mech, BT_TURN_LOG): walkTR 1.309 rad/s (75 deg/s), runTR 0.873 (50 deg/s), turn-in-place + walk yaw at walkTR, no crash. [T1 logic / T2 runtime] Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|
arcattackandClaude Fable 5
|
35c30bda67 |
Locomotion: replace task-#64 trn stand-ins with the AUTHENTIC dispatcher (disasm-decoded)
The turn-in-place dispatcher lives in the master-perf gap 0x4a9b5c-0x4ab188 that Ghidra never lifted. Disassembled it with objdump (scratchpad/masterperf.asm) and hand-decoded the real logic, replacing the two [T3] stand-ins I had invented: - Entry gate was `commandedSpeed < 0.25*standSpeed` (a near-zero guess). AUTHENTIC (0x4aa505): arm trn when TURNING (|angVel|>1e-4, port proxy = turnDemand deadband) AND speed in the FULL [0, standSpeed] sub-walk range AND turnCapable. trn is gated on TURNING, not on being slow. - Case-4 exit had a ×4 fast-forward + SetLegAnimation(5) invention. Restored the VERBATIM decompiled leg-SM exits (part_012.c:12013): standSpeed<spd or reverse -> Standing, else advance the pivot. Added the master-perf turn-STOP exit (0x4aa5c6): turn stopped -> Standing + legResetLatch=1 (folded into case 4 since the port has no separate master-perf frame). - Body case-4 twin mirrors the authentic exits (mj=0, tracks state for lockstep). The fast-forward was invented to dodge a turn->walk stutter that was ACTUALLY the body channel leaking joints (mj=1) -- already root-caused + fixed by mj=0. The trn->walk seam (legFrm7 cycle-end -> swr legFrm1) is the authentic pose-matched boundary and the same seam the fast-forward version hit, just at the correct release speed. Also decoded (documented, not yet wired) the authentic turn-RATE lerp (walkingTurnRate@0x574 -> runningTurnRate@0x578 by speed; angVel = turnDemand*rate); the bring-up drive still yaws at the constant kDriveTurnRate. Headless-verified (BT_AUTODRIVE/BT_FORCE_TURN/BT_WALK_DELAY/BT_GAIT_TRACE): turn-in- place enters state 4, ramp past standSpeed -> 4->stand->swr->walk, both channels in lockstep. Feel re-verification pending. [T1 logic / T2 runtime] Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|
arcattackandClaude Opus 4.8
|
bfdd41bb9d |
Gait: turn-then-walk stutter ROOT-CAUSED + FIXED -- one skeleton writer (task #64)
User repro (100%): turn in place, push forward before the turn stops -> the gait
skips/stutters + visibly reduced bob; standstill starts always clean. Three-layer
fix, user-verified:
1) THE DISPLAY BUG (the actual visible artifact): v5's "body advances first with
mj=1, leg overwrites last, so body drift can't show" was FALSE. Whenever the two
gait channels phase-split, the BODY channel's out-of-phase joint writes leaked
into the rendered skeleton (rhythmic leg skips, averaged-down bob) while every
leg-channel trace read clean -- the leg DATA was fine, the RENDERED pose wasn't
the leg's. v6: AdvanceBodyAnimation(dt, mj=0) -- the body still advances +
projects for replication (records/cycle speeds unchanged) but no longer touches
the skeleton. One writer, structurally; matches the binary's own observable
("body phase drift is locally INVISIBLE in the binary"). BT_BODY_MJ=1 = old A/B.
2) THE SPLIT SEED: the bring-up trn trigger armed only the LEG channel, so a
turn-in-place entry guaranteed the channels re-entered walk frames apart and the
walk cycles ran permanently out of phase. Lockstep: the Standing trn entry arms
BOTH channels the same frame (body case-4 twin added, same rate/keying), gated on
both Standing. The authentic dispatcher (un-decompiled master-perf gap 0x4a9b5c-
0x4ab188, the sole reader of turnDemand/turnCapable) armed both -- the body's
case-4 machinery is dead code otherwise [T1].
3) THE POSE-MATCH INVARIANT: the engine has NO pose blending; transitions avoid
pops purely by authored pose-matched boundaries. The old trn exits cut the pivot
clip MID-STEP (legFrm 7->1 teleport). Now: entry gated on near-zero speed
(turn-IN-place); on forward command the pivot FAST-FORWARDS to completion (x4)
and the authentic finish callback lands Standing at the stand pose -> the normal
pose-matched stand->walk runs. Decompiled reverse abort kept; the standSpeed
mid-clip abort subsumed (it WAS the pose cut). [T3: 0.25*standSpeed threshold +
4x rate stand in for the gap's constants.]
Harness: BT_FORCE_TURN now reaches gBTDrive.turn (was silently inert for the
gait), BT_WALK_DELAY=<s> holds forced throttle then ramps (the turn-first repro),
BT_GAIT_TRACE=1 per-frame gait trace. Regressions: standstill start, turn-entry,
pure pivot loops, run cycle -- all clean, bob full amplitude (1.33/1.32).
KB: locomotion.md v6 section + trn reconstruction + symptom-family closure.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
||
|
arcattackandClaude Fable 5
|
2e9c78d604 |
Gait: the AUTHENTIC walk lean -- interior/exterior clip sets, level cockpit + leaning mechs (task #59)
The MadCat (and every leg-mech) leans -8deg into a walk / -11deg into a run --
an authored jointhip (hingex) pose in the EXTERIOR gait clips that the walk/run
CYCLE clips never rebind, so it HOLDS. The 1995 game keeps a separate INTERIOR
('i'-suffix) clip set for the mech you PILOT: those shake jointshakey (cockpit
rattle) and OMIT jointhip, so your own view stays level while everyone else sees
you lean. The port had this entirely dead -- the local cockpit mech pitched
-8deg into the ground (user-reported "staring at the ground").
ROOT CAUSE (two stacked bugs):
- The authentic ctor clip-set gate (@part_012.c:10308-10320) was reconstructed
as no-op stubs: LoadLowDetailBody/LoadHighDetailBody MISLABELED the
FUN_004a80d4/86c8 GAIT-CLIP loader addresses as a "body LOD" pair, so the gate
did nothing and a separate unconditional LoadLocomotionClips always loaded
EXTERIOR clips.
- LoadLocomotionClipsExt (the interior loader) was a stub aliased to the
exterior loader.
FIX:
- Reconstructed the real 4-char INTERIOR loader (byte-exact vs @004a86c8; the
'i'-suffix table dumped from the exe @0x10d74d: swri/wwri/wwli/...; all 18
base clips confirmed present in BTL4.RES for mad/blh/ava).
- Fixed the ctor gate: getenv("L4VIEWEXT") || (instanceFlags&0xC)==4 -> exterior
(replicant / forced external view); else -> interior (local cockpit master).
Removed the mislabeling no-op stubs.
- PORT ADAPTATION (MaintainViewClipSet, per-frame in PerformAndWatch): the port
never sets the replicant COPY bit (all MP mechs build as local masters --
heat-sim/scoring/torso-watcher-connect all run on both, by design), so the
ctor gate lands everyone on interior. Pick the set by VIEWPOINT instead: the
mech you pilot keeps interior (level cockpit), every other mech flips once to
exterior (the lean). Reloads only on a viewpoint-status change; the sets
share stride data so the swap is seamless. Model pointers stashed at ctor.
VERIFIED live (BT_HIP_LOG probe): your walking mech writes jointhip=0 (level);
the peer's replica in the other pod's view leans at exactly -8.0/-11.1deg --
the authored clip values. A/B control: L4VIEWEXT=1 forces the walking master
back to 74 sustained -8/-11 writes.
KB: locomotion.md ("NO walk lean" audit conclusion CORRECTED -- the lean is
authored + sustained + clip-set-specific), asset-formats.md (.ANI parses
pitch/yaw/roll for all joint types; KeyJointPos translation; cycle clips carry
no pitch), open-questions.md (item closed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
||
|
arcattackandClaude Fable 5
|
02cdfd6576 |
Torso: the TWIST goes LIVE -- electrical watchdog chain, centered crosshair, coherent controls (task #57/#58)
The MadCat torso twists, the view turns with it, and targeting follows. Three reconstruction fronts closed: THE ELECTRICAL WATCHDOG CHAIN (why the torso never powered up): - PowerWatcher::UpdateWatch reconstructed (@004b181c, the REAL registered Performance -- PTR @0050f5fc; Ghidra missed the fn start): the watchdog MIRRORS the watched subsystem's electrical level (+0x278), brownout downgrade when gen output <= minVoltage% x rated. @004b1804 relabeled ResetToInitialState (slot 10) -- the old "Simulation" tag was wrong. - The factory watcher-CONNECT pass reconstructed (vtable slot +0x38, @004aee2c/@004b1a40 byte-identical, recovered from raw exe bytes): watchedLink.Add(roster[watchedSubsystem]) on the master node. Was the SubProxy::Start() no-op -- every watchdog sat at 0 forever. - MinVoltageScale = 0.01 (a 10-byte x87 literal @0x4b1924; was 1.0f = permanent brownout) and PowerWatcher's Derivation chains its REAL base HeatWatcher (the HeatableSubsystem stand-in broke IsDerivedFrom for the whole Torso/Searchlight/ThermalSight family). - KB correction swept: derivation tag 0x50e604 = HEATWATCHER (not "HeatSink"); the btl4gaug heat-widget gate now tests it via the BTIsHeatWatcher bridge. THE CROSSHAIR (task #58 forensics, 6-agent workflow + live probes): - The VIEW is TORSO-MOUNTED: jointtorso -> jointeye -> siteeyepoint in every twist-capable .SKL; the camera + canopy ride the same hinge subtree through HingeRenderable's live matrix-stack compose -- ALREADY WORKING in the port. The crosshair stays screen-centered (center IS the boresight); the twist reads on the tape carets/compass/radar. - The real bug was the port's gBTAimX = tan(twist) slew (the falsified "body-mounted view" model): the camera already carried the twist, so the crosshair counter-slid to hull-forward and the fire ray with it. Deleted; the pick ray inherits the twist from the yawing eye basis. - Two instrumentation traps documented (chase-eye-as-default-camera, BT_FORCE_TORSO clobbering real joints -> the hook now only fills unresolved ones); an over-correcting explicit eye compose was added on those false readings and retired the same day. CONTROLS + REPLICATION: - Q/E spring-center on release (the axis is a twist-RATE demand; the old hold-deflection model drifted forever); X also zeroes the axis and pulses the authentic torso Recenter (@004b6918). M cycles control mode via the real CycleControlMode body. - Torso update-record DIRECTION fixed: engine truth is Write=serialize / Read=apply; @004b6a78 is the READ (was mislabeled Write) and the missing WRITE @004b6a1c recovered from raw disasm (recordLength 0x1C, twist/vel/rate at +0x10/14/18) -- kills the replicant's 0xCDCDCDCD -140-degree ghost twist. - Marching-ghost desync: 4 Standing-case guards zero stale reverse cycleSpeed (negative cadence passed the <= ZeroSpeed stop gate). - Kill credit rerouted to the OBSERVED killer (lastInflictingID -> killer's player link) -- kills count, target K/D populates. KB: subsystems.md (watcher chain), multiplayer.md (record direction), combat-damage.md + gauges-hud.md + cockpit-view.md (torso-mounted view re-correction), decomp-reference.md (new addresses + tag fix), open-questions.md (dead capability-roster loops 2-4, snapshot CD read). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|
arcattackandClaude Fable 5
|
4554ea543a |
KB: locomotion turn-rate note -- authored rates stream correctly post task #4
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|
arcattackandClaude Fable 5
|
3c34ae6de6 |
KB: full staleness audit + sweep -- 35 verified corrections across 16 topics, 15 new glossary terms
Adversarially-verified audit (18 agents) against the task #46-#56 landings: MP/current-state understatements, superseded punch readings, retired env gates, dead paths, line-cite drift; env-gate hub table completed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|
arcattackandClaude Fable 5
|
47aede3094 |
Shadow: tag tshd proxies in the LOADER (V toggle broke the pipeline)
User report: the shadow broke MID-SESSION -- suddenly solid opaque black and buried on any slope. The session log pinned the trigger: pressing V. The view toggle (SetViewInside) RELOADS every segment mesh through the object loader and swaps them into the render tree -- including blh_tshd.bgf -- and the reload path never called SetIsShadow(1) (only the original tree-build call site tagged it). Untagged, the entire shadow pipeline disengages: no translucent TFACTOR state block (solid black), no depth bias (buried), wrong pass. The damage-swap reloads had the same hole. Fix: tag *tshd* filenames in LoadObjectBGF itself, next to the existing all-shadow-material detector (which misses the mech proxy -- plain black material, not shadow_mtl). Every load path -- build, damage swap, view toggle -- now yields a tagged object; the drawOps alphaTest routing reads the flag downstream. User-verified live: the shadow survives repeated V toggles, translucent and slope-correct. This also likely explains the RECURRING "shadow vanishes on inclines" reports: any V toggle silently broke it; every relaunch "fixed" it. KB (locomotion.md): loader-level tag recorded as co-dependent part 5 of the shadow arrangement; the vanish-diagnostic note now lists all three causes (reload untag / gait desync / genuine burial) with check-first guidance. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|
arcattackandClaude Fable 5
|
b6fe686b22 |
Shadow: authentic terrain tilt + decal draw order (feet layer correctly)
User report: the ground shadow painted OVER the mech's feet. Root cause: the
flat quad + big depth-bias (-0.004 ~= 2-4 world units) workaround for slope
burial -- a bias big enough to beat the TERRAIN also beat the FEET in the
depth test, and the shadow drew in the late alpha-blend pass (after the mech),
so z-arbitration was the only layering control.
Fix, four co-dependent parts (user-verified live: feet on top, opaque,
survives inclines/declines):
- TILT (mech.cpp UpdateShadowJoint): apply the SKL's own contract for
jointshadow ("apply terrain angle to pitch and roll") -- quad up aligned to
the surface normal via an orthonormal basis fed to the ENGINE's own
LinearMatrix->EulerAngles conversion. Hand-derived Euler signs are exactly
how the prior tilt attempts "dug into the hillside"; the engine conversion
is convention-proof. ~35 deg cliff-guard cap.
- SAMPLER (mech4.cpp): surface gradient from the collision probes PLUS
BTVisualGroundLift per probe -- the quad hugs the VISIBLE terrain, whose
lift varies across a slope; world->local rotation by the TRUE yaw from
localToWorld, not the drift-prone gDriveHeading mirror (the task-#48 bug
class). BT_SHADOW_LOG traces the normals.
- DRAW ORDER (new PASS_SHADOW, l4d3d.h/L4VIDRND/L4VIDEO): shadows draw
between the STATIC terrain and the DYNAMIC opaque bodies -- the classic
decal order. The mech, drawn after, z-passes over the shadow: the bias
can never paint the shadow over the feet, structurally.
- BIAS pairing (L4D3D.cpp): tilt on (default) -> decal epsilon -0.0008;
BT_SHADOW_TILT=0 flat fallback -> the old -0.004. BT_SHADOW_BIAS overrides.
KB (locomotion.md): the four-part arrangement recorded as co-dependent, with
the gait-desync-symptom diagnostic hint retained.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
||
|
arcattackandClaude Fable 5
|
099cad998d |
KB: incline shadow-vanish co-resolved with gait v5 (user-confirmed live)
The shadow report was a desync symptom, not a bias/tilt bug: the ~2-4 unit depth-bias margin is tight enough that v4's out-of-phase leg pose showing through could push the flat quad under a slope. No shadow code changed; both symptoms cleared together. Noted as a diagnostic hint (check [sync]/BT_SYNC_LOG before touching bias/tilt for this report class). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|
arcattackandClaude Fable 5
|
d4e139a0d9 |
Gait v5: display+travel from the LEG channel (kills the stutter class)
User-reported recurrence of "legs stutter and lose sync". Live [sync] showed
advSum/legSum drifting 5->92 during walk-up churn, then locking. Reproduced
headlessly (BT_FORCE_OSC throttle feathering + BT_FORCE_TURN + the new
per-frame BT_SYNC_LOG probe) and root-caused in two parts:
- SEED: the bring-up turn-in-place trigger arms only the LEG channel, so the
walk re-entry lands one frame apart -> a persistent clip phase offset.
- AMPLIFIER: with offset clips the two end-of-clip callbacks fire on different
frames; a demand change landing between them picks DIFFERENT next states
(one channel winds down, the other keeps walking) -> opposite-phase churn
(observed bs=6 ls=7 -- body on walk-R, leg on walk-L).
Under v4 (display+travel = BODY) the out-of-phase LEG pose showed through on
every frame the body didn't write joints (Standing/wind-down) = the visible
stutter/pop.
DISASM GROUND TRUTH (neither Advance fn has a static decomp caller; found by
byte-scanning the CODE section for e8 calls):
- master perf (0x4a9b5c gap) -> AdvanceLegAnimation @0x4aa399 (air @0x4aa388)
-> -dist/dt into localVelocity (+0x1cc): the LEG drives LOCAL travel and
(writing last) the displayed pose;
- IntegrateMotion (0x4ab1c8, body advance @0x4ab312, caller FUN_004ab430 =
the projected-origin updater) -> -dist/dt into projectedVelocity (+0x2a0):
the BODY is the dead-reckoning PROJECTOR -- locally invisible.
v4's reading of FUN_004ab430 as the travel source was wrong.
FIX: advance body FIRST (projection; its writes get overwritten), leg LAST
(displayed pose); travel = legAdv under the two-channel split. Display ==
travel through the LEG by construction -- body-channel drift can no longer
become visible, structurally (not by input-symmetry luck).
Verified: trn+feathering repro clean; straight drive locked (adv==proj);
solo goto=enemy kill chain intact; MP 2-node drive-to-range cross-pod kill
intact (241 hits, DESTROYED); no crashes. KB corrected (locomotion.md roles
+ P3_LOCOMOTION.md v5 entry). New gated instruments: BT_SYNC_LOG,
BT_FORCE_OSC.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
||
|
arcattackandClaude Opus 4.8
|
f914fc040a |
context-system: complete migration -> CLAUDE.md is now a 160-line router (zero context lost)
Full migration of the 2236-line monolithic CLAUDE.md into the progressive-context knowledge graph (per spark-lesson / expert-seed.md), so the deep RE knowledge loads on-demand instead of every session. ZERO CONTEXT LOST: - docs/PROGRESS_LOG.md = the complete old CLAUDE.md, VERBATIM (byte-identical) -- the lossless safety net + the "full detail" quick-lookup fallback. - 18 context/*.md topic files (1343 lines) digest every section (§1-3 -> project-overview, §4 -> content-archives, §5 -> asset-formats/bgf-format, §5a -> source-completeness, §5b/§8 -> wintesla-port, §7/§10 -> locomotion, §10a -> build-and-run, §10b -> reconstruction-method, §10c -> combat-damage + reconstruction-gotchas, §10d -> subsystems, render notes -> rendering, gauges -> gauges-hud, MP -> multiplayer, §9 -> open-questions). - reference/glossary.yaml (53 terms). decomp-reference.md = the offsets/ClassIDs/addresses hub. CLAUDE.md (160 lines) = router: identity, answer/reason protocols, quick-lookup table, evidence tiers (T0 engine-truth / T1 decompiled+verified / T2 reconstructed+runtime / T3 guarded / T4 hypothesis), conventions + DO-NOT (the systemic bug classes), structure. Retains the load-bearing work directives (build recipe pointer, "keep current" mandate). Knowledge graph validates CLEAN (scratchpad/checkctx.py -- all [[links]] + quick-lookup + docs refs resolve; [[name]] -> topic file or glossary term). docs/*.md ledgers stay as the detailed logs; context/*.md are the curated digests that route into them. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> |