feat: remake native enemy and boss attacks

[hoainho]

Scope

Closes #131.

Remakes 93 enemy/boss attacks on 25 original native bodies, 97 physical attack/body pairs across eight regions. Includes Koro Branchbreaker, Asha, Ruun/Ghor aliases, Solkar, Orun human/Crowned Beast, six scripted aspects and both actual possessed-companion bodies. No simulation/campaign authority source changes; no paid generation or replacement of approved native geometry.

• Bake dedicated continuous native attack tracks from immutable source rigs; retain original mesh/skin/bind/material/texture bytes and original clips. Signed hinge conditioning is opt-in; existing gait authoring stays on its prior path.
• Cut runtime consumers over from generic attack plans to one explicit attack/body/motion/socket inventory and authoritative phase clock.
• Queue fixed-step cues, evaluate actual native poses, then emit/update bounded body-attached effects before the final upload. Preserve original cue birth times and current frame shader time.
• Keep per-instance anatomical/equipped-weapon anchors, current authoritative attack roots/yaw, and offscreen/pause/hit-stop clock alignment. Scope cancellation cannot resurrect an owner; model rebinding preserves other live scopes.
• Add explicit action-specific late-boss/hunter effects without renderer-created hits. Candidate required actions are actual GLB load requirements, not production verification.
• Update all affected consumers, producer/catalog contracts, meaningful physical/lifetime boundary tests and four existing feature documents. Temporary servers, browser and qualification scaffolds removed.

Runtime evidence

Acceptance and main-session review · joined inventory · 93-action runtime

• All 93 actions traverse windup/commit/active/recovery: 12,138 attack samples, 445,879 bone-scale checks, 412,358 constant-joint-position checks, 15,353 native-tip ancestry checks. Zero native/root-XZ/effect-birth clock error, zero relative bone-scale error, no foreign tip or missing phase.
• Actual GPU pose audit: 1,164 poses across all 97 pairs. Original 25 GLBs' geometry, skins, bind transforms, materials, textures and original BIN/accessor/animation prefixes retained.
• Actual scripted six-aspect integration and two Fall possession cases; Sky simultaneously runs Orun and three independent native Sky Echo owners. Controlled entry/placement, real phase scripts/cues/damage.
• Real duplicate-body scope isolation, offscreen attack evaluation, pause/hit-stop, actual Perfect Dodge slow policy, reset/unload, 20-step coalesced first-frame presentation and actual airborne nonlethal trial subdual.
• Normal GameApplication keyboard/AI combat and actual built-production UI/candidate native-load closure exercised separately from controlled activation. No hand-cleared campaign claim.
• Final review caught and repaired renderer-owned gate feedback inheriting the previous frame's clock: actual GameApplication smoke 16.67 ms stale before → 0 after, with a deliberately injected renderer feedback event, not injected game progression.

Watchable native media

All are real 1280×720 H.264/yuv420p, 24 fps encoded artifacts; encoding is not game frame-rate evidence.

• 97 native attack/body pairs — 3m43s: labeled in-place textured/skinned pose review, not campaign gameplay.
• Eight-region native-body/VFX inspection — 23s: real-engine controlled attacks; CameraDirector-owned close inspection framing.
• Normal keyboard/AI trial combat — 11s: ordinary gameplay camera, not a trial-clear assertion.
• Actual airborne subdual — 6s: actual Light hits, hit-stop, pose settling and trial-stage transition.

Verification

Observed local gates; counts overlap and must not be summed:

• Integrated app/effects/rendering plus named combat/trial/facing suites: 960 pass, 0 fail, 76 files, 484,504 assertions, 140.59 s. Followed by final focused validation after the recipe-regexp hoist and gate-clock fix.
• Final bun test tests/app tests/effects: 422 pass, 0 fail, 33 files, 286,869 assertions, 68.33 s.
• Native authoring/physical offline Node gate: 312 pass, 0 fail, 55 suites, 28.607 s.
• Final Node 24.22.0 node --test tests/release/*.test.mjs: 236 pass, 0 fail, 4.59 s. This is release-contract coverage, not strict release qualification.
• Final tsc -b and Vite production build passed; 332 modules, 8.85 s Vite. Existing large-chunk warning retained.
• Actual 25-native rebake, native manifest sync check and five existing regional catalog CLI rebuilds passed; built 25-GLB SHA/magic and seven-catalog closure retained.

Cost / boundaries

Identical before/after workload receipt: canonical 2d86748 vs this implementation, 1280×720 DPR1, 120 warmup + 360 measured frames, one actual boar activation, fixed step/frame. Mostly late attack/recovery/idle; not worst-case boss load.

• CPU median 2.7→2.6 ms, GPU median 21.62→21.27 ms, draw median 135→135. Single runs; no optimization/locked-60-fps claim.
• Independent forced-GC whole-page JS heap 60.98→78.42 MB; document/node counts differ. Scene-geometry backings +368,256 B, cached-clip backings +955,900 B (separate/non-additive); renderer geometry/texture/program counts 81/76/60→82/76/60. Memory cost is explicit; no leak-free claim.
• Actual authoritative combat snapshots match exactly. Failed baseline QA setup attempts are excluded and documented, not benchmark samples.
• Base-Koro fruit-burst alternate mapping is native-pose-tested, not an independently exercised AI encounter. Actual slow-policy trace includes player damage and is not an invulnerability assertion.
• All candidate evidence stays blocked/unverified; production manifest/strict fail-closed release gates unchanged. Existing Later producer Orun knockdown/get-up block remains independent.
• No supplied/decompiled Kena or ZZZ commercial binary or copied proprietary implementation/assets. Actual prior REA/observable choreography lessons only.
• Actual model request/token counters unavailable/null; no quota or quality inference. Immersive title, introduction, storytelling and loading with measured runtime performance #130 worktree preserved; broader [A103] Complete Koro boss and recruitment #66/[A202] Complete Ruun boss, leadership choices, and Sable #71/[A204] Complete Ghor boss and river restoration #73 criteria are not closed here.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
wild-crown	Ready	Preview	Oct 18, 2026 3:47pm UTC

[hoainho]

Main-session review — 06a6be4

Coordinator-owned spec/code/lifetime/security review, not an independent external approval. Implementation accepted for the issue-131 native presentation/candidate scope; observed GitHub CI is still required before merge.

1. Contracts / authority / persistence: inspected producer→seven catalogs→candidate gate→actual GLB loader→animation driver→queued presentation→native socket effects. No combat/campaign authority or save-schema source changes. Required native clips are load contracts, not evidence promotion; blocked/unverified production boundaries retained. Original 25 native mesh/skin/bind/material/texture/BIN prefixes retained; 97 new clips physically checked.
2. Behavior / geometry: actual all-93 four-phase run passed with 12,138 attacking samples; independent original-source bone-scale/constant-joint-position and actual skinned GPU checks. Six actual scripted aspects and both possessed companion bodies passed; Sky has four independent native owners. Actual normal keyboard/AI UI, built native-load closure, 20 coalesced steps, hit-stop/slow, offscreen pause and airborne subdual exercised. Video scopes and base-Koro fruit-burst pose-only alternate are explicitly qualified.
3. Ownership / lifetime / security: per-instance bones/weapon tips and bounded preallocated native trail slots. Repaired cancelled-pending snapshot resurrection, offscreen/pause LOD debt, active-cue root rebind dropping another live scope, and renderer-owned gate birth using the previous frame clock. Scoped cancellation, rebind, repeats, pause and unload checked; latest root-rebind regression failed before and passed after. No purchases, new external service integration, proprietary reference implementation or release-gate weakening.

Defect-catching question: can a cancelled/rebound/offscreen actor emit from an old, foreign or unowned cast on the first presented frame? Final actual runtime receipts and focused ownership regressions show no such failure in the exercised cases. Native/body clocks and root XZ errors are zero across the 93-action run.

Observed gates: integrated 960/0; final application/effects 422/0 after final fixes; native physical/authoring 312/0; release-contract 236/0; final typecheck/build passed. Counts overlap; not a full-repository-suite claim. Large bundle warning, positive retained-memory cost, imperfect-dodge trace damage, blocked asset/campaign qualification and existing Later producer block are disclosed, not suppressed.

No unresolved implementation blocker found within this scope. #130 remains preserved and unpublished; #66/#71/#73 remain independent. See committed asset-source/playtests/evidence/enemy-attack-remake/acceptance.json, inventory, raw runtime receipts and actual H.264 media.