{"entries":[{"lastCommit":{"author":{"email":"scythes.lol@gmail.com","name":"Angel D. Munoz","when":"2026-08-11T20:30:59.000000-06:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-08-11T20:30:59.000000-06:00"},"hash":"8e5e3269d1b16fcbc2794688f4403b6d59500c64","message":"Experiments/adaptive architecture (#100)\n\n* feat(adaptive): AdaptiveHeadless runner, AdaptivePong demo, and the Defli tower-defense port\n\nThe adaptive-data architecture experiment (Mibo.Adaptive) plus its two\nreal-scale validations, as an alternative to MVU:\n\n- Mibo.Adaptive: the AdaptiveHeadless runner — pull-based state on\n AdaptiveSlop (State · Projection · Update · Force). State lives in\n changeable roots, derived state in adaptive projections, and the\n runner forces the frame's projections at the end of each Step, so\n draw reads are O(1) and unchanged state recomputes nothing (idle\n Step: 0 recomputes, 0 bytes). Step/StepN/StepUntil/Run/RunAsync and\n observers mirror HeadlessRunner; there is no 'Msg, no Cmd, no Sub —\n handlers write roots and run effects directly.\n- Mibo.Adaptive.Demo: AdaptivePong — PingPong's sim rebuilt on the\n runner, with both a console sim frontend and a raylib frontend.\n- src/Defli: Defli's tower-defense simulation (nine systems, 4,554\n lines) ported from its MVU shell onto AdaptiveHeadless, headless\n first. The World record is the composition root (State); Router.fs\n is the per-frame router, event translation and cold paths (Update);\n Frame.fs packs RenderFrame once per Step (Force). System update/tick\n return only the events they emit — in-place mutation, nothing\n discarded, no recursion. The full Defli test suite (94 tests) passes\n against the port, unchanged in behavior.\n- AdaptiveSlop submodule pinned at eba4b8c (dirty-indicator Version\n fix for journal-and-drain nodes — the suppression chain settles\n bottom-up on read; no pre-read workaround needed).\n- Mibo.slnx, CHANGELOG, and InternalsVisibleTo for the test projects.\n\n* feat(adaptive): windowed raylib host, neutral camera, and the Defli raylib frontend\n\nAdaptiveRaylibGame<'Frame> (Mibo.Raylib) runs an adaptive world in a\nraylib window with the MVU ceremony removed: no Program builder, no\nCmd/Sub machinery, input always registered and polled, services read\nfrom the context in Init/Update. The runner accepts a pre-built\nGameContext, exposes it via AdaptiveContext, and gains Restart\n(dispose world disposables, re-run Init, reset clock, force the first\nframe) consumed by the host after Step.\n\nThe Defli sim camera is now backend-neutral CameraState (target/zoom/\nrotation/worldsize/shake) with pure, unit-tested view math\n(clampToWorld/screenToWorld/viewBounds); the raylib Camera2D is built\nat the view edge. RenderFrame carries the hover/selection/projection\nand Vfx/Map/Diag state so the frontend draws read only the struct.\n\nsrc/Defli.Raylib restores the trimmed view layer from the original\nDefli (Map/Enemies/Towers/Projectiles/Vfx/World/HUD/Diagnostics),\nreading the forced frame; the shell translates input through\nsubscriptions (no Msg/Cmd/Sub), with restart via the world cell +\nRestartRequested. Camera tests were wired into Main.fs (were dropped\nin the milestone-1 port) and extended with the pure math tests.\nRemoved a leftover inline/private FS1113 in Router (startNextWave,\napplyEnemyMsg) that broke the committed tree's build.\n\n* fix(adaptive): per-kind particle buffers in VfxView\n\nThe draw commands are deferred — Renderer2D executes them after the\nview returns, and each command holds the array it was recorded with.\nAll six particle kinds shared one conversion scratch, so every\ncommand drew the LAST kind's particles with its own texture (impact\nsparks stretched over smoke puffs — the square behind the puff).\nOne persistent buffer per kind keeps each command's data isolated;\nsteady state still allocates nothing.\n\n* fix(adaptive): restore the queued-pump ordering in the splash fan-out; fix WASD pan\n\nThe port's direct event handling ran the Killed handling (despawn +\nboss-split spawnAt) INSIDE the splash fan-out's enumeration of\nPositions. The transaction commits mutate the live dictionary\nmid-loop, so the enumerator threw 'Collection was modified' — the\nwave-13 crash (the original queued the ApplyDamage messages and the\npump ran them after the fan-out). The fan-out now applies the damage\nto all targets first and handles the events after the enumeration.\nRegression test: a cannon blast killing a pre-damaged boss + runner\nreproduces the exact Dictionary.Enumerator crash on the old code and\npasses on the fix.\n\nWASD pan: panStep returned the X axis for the vertical keys (Up →\n(1,0), Down → (-1,0) instead of (0,±1)) — the original accumulated\ndy for Up/Down. Vertical panning never happened.\n\n* docs(adaptive): wave 20-35 trace assessment for the Defli.Raylib port\n\nAnalyzes defli-adaptive-sim.speedscope.json (518.3 s, windowed game,\nwaves 20-35 warm end-game) against the original Defli baseline docs:\nbusy 7.2% of wall (1.19 ms/frame), AdaptiveSlop 3.7% of wall\n(0.62 ms/frame) vs the original's 5.7% (0.95 ms/frame), 0 GC frames,\n0 write-dispatch samples. Verdict: the adaptive shell is cheaper per\nframe than the MVU shell at comparable load.\n\nAdds the trace-analysis tooling to tools/ (copied from E:\\Defli\\tools\nplus two new probes: probe-structure.fsx validates the evented\nCPU_TIME trace format, probe-nodes.fsx attributes samples per node\nchain) and registers dotnet-trace in dotnet-tools.json.\n\n* feat(adaptive): convert the Defli projections to AMap.joinOn\n\nThe AdaptiveSlop submodule now pins TheAngryByrd/AdaptiveSlop\nfeat/joinon-groupby-reductions (PR #19): AMap.joinOn, groupBy,\ndifference, and the *A reduction family.\n\nThe mapA + tryFind join idiom is replaced at all four projection sites:\n- Homing: Projectiles.Rows x Enemies.Positions joined on the target\n enemy (the profiled hot spot: the per-projectile subgraph is built\n once and the position input swaps in place, no rebuild on update)\n- Enemies.Views: 3-way joinOn composition (Positions x Healths x\n Motions); the defensive zero row for transient mid-transaction\n reads is preserved\n- Enemies.BossPositions: same-key joinOn into Defs with the\n ValueNone drop (choose semantics)\n- Towers.EffectiveDef: same-key joinOn into Levels\n\nBehavior is unchanged: Defli 104/104, AdaptiveSlop 407/407. The pipe\nform lands in the join's right slot (right |> joinOn key mapping\nleft); record-typed lambda parameters carry annotations because the\ninline lambdas elaborate before the map arguments pin their types.\n\n* feat(adaptive): pin AdaptiveSlop to merged master + post-joinOn trace assessment\n\nThe AdaptiveSlop submodule moves from the PR branch head to the merged\nmaster commit 9e81834 (PR #19, squash). The merged joinOn signature\nreorders the arguments to maps-first — (left) (right) (keyOfLeft)\n(mapping) — so the four Defli projection call sites drop the pipe form\n(the pipe would land in the mapping slot). The maps-first order does\nnot fix the inline-lambda record inference, so Homing and the Views\n3-way join keep their type annotations.\n\nTrace assessment (second wave-20-35 session, README section 6):\nbusy 1.19 -> 0.55 ms/frame, AdaptiveSlop 0.62 -> 0.18 ms/frame,\nHoming join 28.5% -> 0.7% of busy despite a heavier session (more\ntowers, 2.6x more projectile churn). The slowliness is not in the CPU\ndata: the longest busy cluster is 10 ms, the 733 ms CPU_TIME spans\nare 1-sample sampling artifacts, and the 413 idle gaps >= 120 ms\ncluster in the boot minutes (collector pauses). GC measured directly\nfrom the GC events with the new F# tools/gcprobe: 188 GCs, every\npause < 1 ms, worst 0.69 ms. Against the seven original Defli traces\nthe port collects 3.6x less often than the waves-33-35 baseline with\na 3x smaller worst pause (Trace A had 37 ms pauses).\n\nNew tools: tools/probe-spikes.fsx (spans + per-minute census),\ntools/probe-minute.fsx (gap histogram), tools/gcprobe (F# TraceEvent\nGC-lifecycle probe).\n\nTests: Defli 104/104, AdaptiveSlop 407/407.\n\n* chore update probes\n\n* chore: bump ref\n\n* feat(adaptive): mirror the MVU Program API; add MonoGame host; decouple\nthe Defli sim\n\nRewrite the adaptive public surface to mirror Mibo.Elmish.Program by\nname and shape, minus the Msg/Cmd/Sub machinery.\n\nAdaptiveProgram<'Frame> is a single record (Init/Update/Observers +\nConfig/Renderers/ServiceRegistrations/AssetsBasePath/FixedStep/HasInput)\nbuilt with AdaptiveProgram.mkProgram init update |> with*. AdaptiveWorld\nis removed; its slots fold into AdaptiveProgram. AdaptiveInit is kept\nbut Disposables is hidden (ofFrameBuilder defaults it to [],\nwithDisposables/withDisposable augment). AdaptiveFixedStepConfig +\nwithFixedStep: the runner sub-steps Update N times via FixedStep.compute\nand forces once (mirrors TickFrame, no Map/'Msg). AdaptiveRaylibGame\nrewritten to take AdaptiveProgram (reverses\nConfig/Renderers/ServiceRegistrations like RaylibGame) and moved to the\nMibo.Adaptive namespace. New AdaptiveMonoGameGame +\nAdaptiveMonoGameProgram mirror MiboGame/MonoGameProgram (DeviceConfig\nwrapper, Game lifecycle). Input is opt-in via AdaptiveProgram.withInput\non both hosts (mirrors Program.withInput/HasInput).\n\nDefli sim decoupled from raylib: Defli.fsproj no longer references\nRaylib-cs. Particle2D uses Mibo.Color; TileInfo.Rect and\nParticle2D.SourceRect dropped (the raylib view edge builds native\nrectangles via MapView.tileRect); VfxModel.Textures moved to VfxView.\n\nAll consumers migrated (Defli.Raylib, Defli.Tests, Mibo.Adaptive.Demo,\nMibo.Adaptive.Tests). 27 adaptive tests pass (incl. 5 new fixed-step);\nheadless sim still 0 B/frame in the paused phase.\n\n* chore(adaptive): remove Defli sample projects and the adaptive demo from the experiments branch\n\nThe Defli sample (src/Defli, src/Defli.Raylib, src/Defli.Tests) moves to\nMibo.Samples/Defli; the adaptive Pong demo (src/Mibo.Adaptive.Demo) is\ndropped. The adaptive framework pieces (Mibo.Adaptive, hosts, tests) stay.\n\n* feat(adaptive): adopt AdaptiveSlop as Mibo.Adaptive; move the adaptive program into Mibo.Core\n\n- Mibo.Adaptive is now the full adaptive-data library (adopted from\n AdaptiveSlop, namespace renamed AdaptiveSlop.Core -> Mibo.Adaptive),\n dependency-free: no Mibo references, no package dependencies beyond\n the SDK's implicit FSharp.Core. The AdaptiveSlop submodule is removed.\n- AdaptiveProgram/AdaptiveHeadless move into Mibo.Core (the adaptive\n counterpart of the MVU shell); hosts keep namespace Mibo.Adaptive.\n- Mibo.Adaptive.Tests is the adopted xunit + FsCheck suite (417 tests);\n the wrapper tests move to Mibo.Core.Tests. Mibo.Adaptive.Benchmarks\n carries the BenchmarkDotNet suite (FDA comparison).\n- Mibo.Adaptive versions itself from its own CHANGELOG.md (0.1.0);\n release.yml gains a dedicated adaptive release job (tag adaptive-vX.Y.Z)\n and the main pack excludes the adaptive package.\n- Docs, AGENTS.md and the design archive are adopted alongside the code.\n\n* feat(adaptive): mark the AdaptiveProgram module experimental\n\nUsing the adaptive program API now emits an FS0057 warning (with\nRequireQualifiedAccess) — consumers get a compile-time signal that the\nadaptive integration carries no stability guarantees. The changelog\nmoves all adaptive entries to [Unreleased]: nothing adaptive has been\nreleased yet.\n\n* ci: fix the release flow pack step and changelog version resolution\n\n- The main release job packs each library explicitly (a loop — dotnet pack\n accepts a single project), keeping Mibo.Adaptive out of the main release;\n the adaptive job packs it separately with its own version.\n- ChangelogFile now resolves via MSBuildThisFileDirectory instead of\n SolutionDir, so per-project builds/packs derive the package version from\n the changelog even outside a solution context (also silences CNG0002).\n\n* fix(adaptive): address the PR review — runner robustness, publish order, zero-alloc gates\n\nRunner (AdaptiveHeadless):\n- RunAsync: catch game-thread exceptions and rethrow them from MoveNextAsync\n (an unhandled background-thread exception killed the process); reject a\n second concurrent enumerator on the same runner.\n- StepUntil: initialize before the first predicate check (a fresh runner\n evaluated the predicate against a default/null frame).\n- Tests: RunAsync tests build the graph inside Init — the adaptive graph is\n confined to its creating (game) thread and Debug owner-thread checks\n rejected roots built on the test thread, aborting the Debug test host.\n\nCI:\n- release.yml: create-release needs create-adaptive-release so Mibo.Adaptive\n reaches NuGet before Mibo.Core (whose nuspec pins it as a dependency).\n- docs.yml: tests run in Debug, catching the #if DEBUG owner-thread checks\n that Release-only CI skipped.\n\nPerf (invariant 5 — zero allocation on clean reads):\n- Version gates on SetToListNode, ConcatListNode, SortListNode and a new\n SubListNode (subA/takeA/skipA track source + bound versions): measured\n 0 bytes per clean read. PollListSourceNode stays un-gated by design — its\n build closure may read extra adaptive inputs — and is documented as the\n exception.\n\nBackends/docs:\n- MonoGame adaptive host renamed AdaptiveGame -> AdaptiveMonoGameGame\n (matches the documented name and AdaptiveRaylibGame symmetry).\n- Both host headers now state input is opt-in via AdaptiveProgram.withInput\n (the code already gated on HasInput).\n- MonoGame 2D: SpriteState.Rotation is degrees (raylib parity); the backend\n converts to radians for SpriteBatch.Draw and lit tessellation; unit docs\n and the corner-rotation test updated.\n\nChore:\n- Changelog: the experimental marker belongs under Added (unreleased feature).\n- Drop the dev-only tools/ (probe scripts, gcprobe) — moved to Mibo.Samples.\n- Restore dotnet-tools.json (no dev-only tools, no line-ending churn).","tree":"24d51de51d2f1cbff6e664d1cf2e95620c70df36"},"mode":"0100644","name":"2026-08-10-JOIN-DESIGN.md","size":7077},{"lastCommit":{"author":{"email":"scythes.lol@gmail.com","name":"Angel D. Munoz","when":"2026-08-11T20:30:59.000000-06:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-08-11T20:30:59.000000-06:00"},"hash":"8e5e3269d1b16fcbc2794688f4403b6d59500c64","message":"Experiments/adaptive architecture (#100)\n\n* feat(adaptive): AdaptiveHeadless runner, AdaptivePong demo, and the Defli tower-defense port\n\nThe adaptive-data architecture experiment (Mibo.Adaptive) plus its two\nreal-scale validations, as an alternative to MVU:\n\n- Mibo.Adaptive: the AdaptiveHeadless runner — pull-based state on\n AdaptiveSlop (State · Projection · Update · Force). State lives in\n changeable roots, derived state in adaptive projections, and the\n runner forces the frame's projections at the end of each Step, so\n draw reads are O(1) and unchanged state recomputes nothing (idle\n Step: 0 recomputes, 0 bytes). Step/StepN/StepUntil/Run/RunAsync and\n observers mirror HeadlessRunner; there is no 'Msg, no Cmd, no Sub —\n handlers write roots and run effects directly.\n- Mibo.Adaptive.Demo: AdaptivePong — PingPong's sim rebuilt on the\n runner, with both a console sim frontend and a raylib frontend.\n- src/Defli: Defli's tower-defense simulation (nine systems, 4,554\n lines) ported from its MVU shell onto AdaptiveHeadless, headless\n first. The World record is the composition root (State); Router.fs\n is the per-frame router, event translation and cold paths (Update);\n Frame.fs packs RenderFrame once per Step (Force). System update/tick\n return only the events they emit — in-place mutation, nothing\n discarded, no recursion. The full Defli test suite (94 tests) passes\n against the port, unchanged in behavior.\n- AdaptiveSlop submodule pinned at eba4b8c (dirty-indicator Version\n fix for journal-and-drain nodes — the suppression chain settles\n bottom-up on read; no pre-read workaround needed).\n- Mibo.slnx, CHANGELOG, and InternalsVisibleTo for the test projects.\n\n* feat(adaptive): windowed raylib host, neutral camera, and the Defli raylib frontend\n\nAdaptiveRaylibGame<'Frame> (Mibo.Raylib) runs an adaptive world in a\nraylib window with the MVU ceremony removed: no Program builder, no\nCmd/Sub machinery, input always registered and polled, services read\nfrom the context in Init/Update. The runner accepts a pre-built\nGameContext, exposes it via AdaptiveContext, and gains Restart\n(dispose world disposables, re-run Init, reset clock, force the first\nframe) consumed by the host after Step.\n\nThe Defli sim camera is now backend-neutral CameraState (target/zoom/\nrotation/worldsize/shake) with pure, unit-tested view math\n(clampToWorld/screenToWorld/viewBounds); the raylib Camera2D is built\nat the view edge. RenderFrame carries the hover/selection/projection\nand Vfx/Map/Diag state so the frontend draws read only the struct.\n\nsrc/Defli.Raylib restores the trimmed view layer from the original\nDefli (Map/Enemies/Towers/Projectiles/Vfx/World/HUD/Diagnostics),\nreading the forced frame; the shell translates input through\nsubscriptions (no Msg/Cmd/Sub), with restart via the world cell +\nRestartRequested. Camera tests were wired into Main.fs (were dropped\nin the milestone-1 port) and extended with the pure math tests.\nRemoved a leftover inline/private FS1113 in Router (startNextWave,\napplyEnemyMsg) that broke the committed tree's build.\n\n* fix(adaptive): per-kind particle buffers in VfxView\n\nThe draw commands are deferred — Renderer2D executes them after the\nview returns, and each command holds the array it was recorded with.\nAll six particle kinds shared one conversion scratch, so every\ncommand drew the LAST kind's particles with its own texture (impact\nsparks stretched over smoke puffs — the square behind the puff).\nOne persistent buffer per kind keeps each command's data isolated;\nsteady state still allocates nothing.\n\n* fix(adaptive): restore the queued-pump ordering in the splash fan-out; fix WASD pan\n\nThe port's direct event handling ran the Killed handling (despawn +\nboss-split spawnAt) INSIDE the splash fan-out's enumeration of\nPositions. The transaction commits mutate the live dictionary\nmid-loop, so the enumerator threw 'Collection was modified' — the\nwave-13 crash (the original queued the ApplyDamage messages and the\npump ran them after the fan-out). The fan-out now applies the damage\nto all targets first and handles the events after the enumeration.\nRegression test: a cannon blast killing a pre-damaged boss + runner\nreproduces the exact Dictionary.Enumerator crash on the old code and\npasses on the fix.\n\nWASD pan: panStep returned the X axis for the vertical keys (Up →\n(1,0), Down → (-1,0) instead of (0,±1)) — the original accumulated\ndy for Up/Down. Vertical panning never happened.\n\n* docs(adaptive): wave 20-35 trace assessment for the Defli.Raylib port\n\nAnalyzes defli-adaptive-sim.speedscope.json (518.3 s, windowed game,\nwaves 20-35 warm end-game) against the original Defli baseline docs:\nbusy 7.2% of wall (1.19 ms/frame), AdaptiveSlop 3.7% of wall\n(0.62 ms/frame) vs the original's 5.7% (0.95 ms/frame), 0 GC frames,\n0 write-dispatch samples. Verdict: the adaptive shell is cheaper per\nframe than the MVU shell at comparable load.\n\nAdds the trace-analysis tooling to tools/ (copied from E:\\Defli\\tools\nplus two new probes: probe-structure.fsx validates the evented\nCPU_TIME trace format, probe-nodes.fsx attributes samples per node\nchain) and registers dotnet-trace in dotnet-tools.json.\n\n* feat(adaptive): convert the Defli projections to AMap.joinOn\n\nThe AdaptiveSlop submodule now pins TheAngryByrd/AdaptiveSlop\nfeat/joinon-groupby-reductions (PR #19): AMap.joinOn, groupBy,\ndifference, and the *A reduction family.\n\nThe mapA + tryFind join idiom is replaced at all four projection sites:\n- Homing: Projectiles.Rows x Enemies.Positions joined on the target\n enemy (the profiled hot spot: the per-projectile subgraph is built\n once and the position input swaps in place, no rebuild on update)\n- Enemies.Views: 3-way joinOn composition (Positions x Healths x\n Motions); the defensive zero row for transient mid-transaction\n reads is preserved\n- Enemies.BossPositions: same-key joinOn into Defs with the\n ValueNone drop (choose semantics)\n- Towers.EffectiveDef: same-key joinOn into Levels\n\nBehavior is unchanged: Defli 104/104, AdaptiveSlop 407/407. The pipe\nform lands in the join's right slot (right |> joinOn key mapping\nleft); record-typed lambda parameters carry annotations because the\ninline lambdas elaborate before the map arguments pin their types.\n\n* feat(adaptive): pin AdaptiveSlop to merged master + post-joinOn trace assessment\n\nThe AdaptiveSlop submodule moves from the PR branch head to the merged\nmaster commit 9e81834 (PR #19, squash). The merged joinOn signature\nreorders the arguments to maps-first — (left) (right) (keyOfLeft)\n(mapping) — so the four Defli projection call sites drop the pipe form\n(the pipe would land in the mapping slot). The maps-first order does\nnot fix the inline-lambda record inference, so Homing and the Views\n3-way join keep their type annotations.\n\nTrace assessment (second wave-20-35 session, README section 6):\nbusy 1.19 -> 0.55 ms/frame, AdaptiveSlop 0.62 -> 0.18 ms/frame,\nHoming join 28.5% -> 0.7% of busy despite a heavier session (more\ntowers, 2.6x more projectile churn). The slowliness is not in the CPU\ndata: the longest busy cluster is 10 ms, the 733 ms CPU_TIME spans\nare 1-sample sampling artifacts, and the 413 idle gaps >= 120 ms\ncluster in the boot minutes (collector pauses). GC measured directly\nfrom the GC events with the new F# tools/gcprobe: 188 GCs, every\npause < 1 ms, worst 0.69 ms. Against the seven original Defli traces\nthe port collects 3.6x less often than the waves-33-35 baseline with\na 3x smaller worst pause (Trace A had 37 ms pauses).\n\nNew tools: tools/probe-spikes.fsx (spans + per-minute census),\ntools/probe-minute.fsx (gap histogram), tools/gcprobe (F# TraceEvent\nGC-lifecycle probe).\n\nTests: Defli 104/104, AdaptiveSlop 407/407.\n\n* chore update probes\n\n* chore: bump ref\n\n* feat(adaptive): mirror the MVU Program API; add MonoGame host; decouple\nthe Defli sim\n\nRewrite the adaptive public surface to mirror Mibo.Elmish.Program by\nname and shape, minus the Msg/Cmd/Sub machinery.\n\nAdaptiveProgram<'Frame> is a single record (Init/Update/Observers +\nConfig/Renderers/ServiceRegistrations/AssetsBasePath/FixedStep/HasInput)\nbuilt with AdaptiveProgram.mkProgram init update |> with*. AdaptiveWorld\nis removed; its slots fold into AdaptiveProgram. AdaptiveInit is kept\nbut Disposables is hidden (ofFrameBuilder defaults it to [],\nwithDisposables/withDisposable augment). AdaptiveFixedStepConfig +\nwithFixedStep: the runner sub-steps Update N times via FixedStep.compute\nand forces once (mirrors TickFrame, no Map/'Msg). AdaptiveRaylibGame\nrewritten to take AdaptiveProgram (reverses\nConfig/Renderers/ServiceRegistrations like RaylibGame) and moved to the\nMibo.Adaptive namespace. New AdaptiveMonoGameGame +\nAdaptiveMonoGameProgram mirror MiboGame/MonoGameProgram (DeviceConfig\nwrapper, Game lifecycle). Input is opt-in via AdaptiveProgram.withInput\non both hosts (mirrors Program.withInput/HasInput).\n\nDefli sim decoupled from raylib: Defli.fsproj no longer references\nRaylib-cs. Particle2D uses Mibo.Color; TileInfo.Rect and\nParticle2D.SourceRect dropped (the raylib view edge builds native\nrectangles via MapView.tileRect); VfxModel.Textures moved to VfxView.\n\nAll consumers migrated (Defli.Raylib, Defli.Tests, Mibo.Adaptive.Demo,\nMibo.Adaptive.Tests). 27 adaptive tests pass (incl. 5 new fixed-step);\nheadless sim still 0 B/frame in the paused phase.\n\n* chore(adaptive): remove Defli sample projects and the adaptive demo from the experiments branch\n\nThe Defli sample (src/Defli, src/Defli.Raylib, src/Defli.Tests) moves to\nMibo.Samples/Defli; the adaptive Pong demo (src/Mibo.Adaptive.Demo) is\ndropped. The adaptive framework pieces (Mibo.Adaptive, hosts, tests) stay.\n\n* feat(adaptive): adopt AdaptiveSlop as Mibo.Adaptive; move the adaptive program into Mibo.Core\n\n- Mibo.Adaptive is now the full adaptive-data library (adopted from\n AdaptiveSlop, namespace renamed AdaptiveSlop.Core -> Mibo.Adaptive),\n dependency-free: no Mibo references, no package dependencies beyond\n the SDK's implicit FSharp.Core. The AdaptiveSlop submodule is removed.\n- AdaptiveProgram/AdaptiveHeadless move into Mibo.Core (the adaptive\n counterpart of the MVU shell); hosts keep namespace Mibo.Adaptive.\n- Mibo.Adaptive.Tests is the adopted xunit + FsCheck suite (417 tests);\n the wrapper tests move to Mibo.Core.Tests. Mibo.Adaptive.Benchmarks\n carries the BenchmarkDotNet suite (FDA comparison).\n- Mibo.Adaptive versions itself from its own CHANGELOG.md (0.1.0);\n release.yml gains a dedicated adaptive release job (tag adaptive-vX.Y.Z)\n and the main pack excludes the adaptive package.\n- Docs, AGENTS.md and the design archive are adopted alongside the code.\n\n* feat(adaptive): mark the AdaptiveProgram module experimental\n\nUsing the adaptive program API now emits an FS0057 warning (with\nRequireQualifiedAccess) — consumers get a compile-time signal that the\nadaptive integration carries no stability guarantees. The changelog\nmoves all adaptive entries to [Unreleased]: nothing adaptive has been\nreleased yet.\n\n* ci: fix the release flow pack step and changelog version resolution\n\n- The main release job packs each library explicitly (a loop — dotnet pack\n accepts a single project), keeping Mibo.Adaptive out of the main release;\n the adaptive job packs it separately with its own version.\n- ChangelogFile now resolves via MSBuildThisFileDirectory instead of\n SolutionDir, so per-project builds/packs derive the package version from\n the changelog even outside a solution context (also silences CNG0002).\n\n* fix(adaptive): address the PR review — runner robustness, publish order, zero-alloc gates\n\nRunner (AdaptiveHeadless):\n- RunAsync: catch game-thread exceptions and rethrow them from MoveNextAsync\n (an unhandled background-thread exception killed the process); reject a\n second concurrent enumerator on the same runner.\n- StepUntil: initialize before the first predicate check (a fresh runner\n evaluated the predicate against a default/null frame).\n- Tests: RunAsync tests build the graph inside Init — the adaptive graph is\n confined to its creating (game) thread and Debug owner-thread checks\n rejected roots built on the test thread, aborting the Debug test host.\n\nCI:\n- release.yml: create-release needs create-adaptive-release so Mibo.Adaptive\n reaches NuGet before Mibo.Core (whose nuspec pins it as a dependency).\n- docs.yml: tests run in Debug, catching the #if DEBUG owner-thread checks\n that Release-only CI skipped.\n\nPerf (invariant 5 — zero allocation on clean reads):\n- Version gates on SetToListNode, ConcatListNode, SortListNode and a new\n SubListNode (subA/takeA/skipA track source + bound versions): measured\n 0 bytes per clean read. PollListSourceNode stays un-gated by design — its\n build closure may read extra adaptive inputs — and is documented as the\n exception.\n\nBackends/docs:\n- MonoGame adaptive host renamed AdaptiveGame -> AdaptiveMonoGameGame\n (matches the documented name and AdaptiveRaylibGame symmetry).\n- Both host headers now state input is opt-in via AdaptiveProgram.withInput\n (the code already gated on HasInput).\n- MonoGame 2D: SpriteState.Rotation is degrees (raylib parity); the backend\n converts to radians for SpriteBatch.Draw and lit tessellation; unit docs\n and the corner-rotation test updated.\n\nChore:\n- Changelog: the experimental marker belongs under Added (unreleased feature).\n- Drop the dev-only tools/ (probe scripts, gcprobe) — moved to Mibo.Samples.\n- Restore dotnet-tools.json (no dev-only tools, no line-ending churn).","tree":"24d51de51d2f1cbff6e664d1cf2e95620c70df36"},"mode":"0100644","name":"BENCHMARKS.md","size":201751},{"lastCommit":{"author":{"email":"scythes.lol@gmail.com","name":"Angel D. Munoz","when":"2026-08-11T20:30:59.000000-06:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-08-11T20:30:59.000000-06:00"},"hash":"8e5e3269d1b16fcbc2794688f4403b6d59500c64","message":"Experiments/adaptive architecture (#100)\n\n* feat(adaptive): AdaptiveHeadless runner, AdaptivePong demo, and the Defli tower-defense port\n\nThe adaptive-data architecture experiment (Mibo.Adaptive) plus its two\nreal-scale validations, as an alternative to MVU:\n\n- Mibo.Adaptive: the AdaptiveHeadless runner — pull-based state on\n AdaptiveSlop (State · Projection · Update · Force). State lives in\n changeable roots, derived state in adaptive projections, and the\n runner forces the frame's projections at the end of each Step, so\n draw reads are O(1) and unchanged state recomputes nothing (idle\n Step: 0 recomputes, 0 bytes). Step/StepN/StepUntil/Run/RunAsync and\n observers mirror HeadlessRunner; there is no 'Msg, no Cmd, no Sub —\n handlers write roots and run effects directly.\n- Mibo.Adaptive.Demo: AdaptivePong — PingPong's sim rebuilt on the\n runner, with both a console sim frontend and a raylib frontend.\n- src/Defli: Defli's tower-defense simulation (nine systems, 4,554\n lines) ported from its MVU shell onto AdaptiveHeadless, headless\n first. The World record is the composition root (State); Router.fs\n is the per-frame router, event translation and cold paths (Update);\n Frame.fs packs RenderFrame once per Step (Force). System update/tick\n return only the events they emit — in-place mutation, nothing\n discarded, no recursion. The full Defli test suite (94 tests) passes\n against the port, unchanged in behavior.\n- AdaptiveSlop submodule pinned at eba4b8c (dirty-indicator Version\n fix for journal-and-drain nodes — the suppression chain settles\n bottom-up on read; no pre-read workaround needed).\n- Mibo.slnx, CHANGELOG, and InternalsVisibleTo for the test projects.\n\n* feat(adaptive): windowed raylib host, neutral camera, and the Defli raylib frontend\n\nAdaptiveRaylibGame<'Frame> (Mibo.Raylib) runs an adaptive world in a\nraylib window with the MVU ceremony removed: no Program builder, no\nCmd/Sub machinery, input always registered and polled, services read\nfrom the context in Init/Update. The runner accepts a pre-built\nGameContext, exposes it via AdaptiveContext, and gains Restart\n(dispose world disposables, re-run Init, reset clock, force the first\nframe) consumed by the host after Step.\n\nThe Defli sim camera is now backend-neutral CameraState (target/zoom/\nrotation/worldsize/shake) with pure, unit-tested view math\n(clampToWorld/screenToWorld/viewBounds); the raylib Camera2D is built\nat the view edge. RenderFrame carries the hover/selection/projection\nand Vfx/Map/Diag state so the frontend draws read only the struct.\n\nsrc/Defli.Raylib restores the trimmed view layer from the original\nDefli (Map/Enemies/Towers/Projectiles/Vfx/World/HUD/Diagnostics),\nreading the forced frame; the shell translates input through\nsubscriptions (no Msg/Cmd/Sub), with restart via the world cell +\nRestartRequested. Camera tests were wired into Main.fs (were dropped\nin the milestone-1 port) and extended with the pure math tests.\nRemoved a leftover inline/private FS1113 in Router (startNextWave,\napplyEnemyMsg) that broke the committed tree's build.\n\n* fix(adaptive): per-kind particle buffers in VfxView\n\nThe draw commands are deferred — Renderer2D executes them after the\nview returns, and each command holds the array it was recorded with.\nAll six particle kinds shared one conversion scratch, so every\ncommand drew the LAST kind's particles with its own texture (impact\nsparks stretched over smoke puffs — the square behind the puff).\nOne persistent buffer per kind keeps each command's data isolated;\nsteady state still allocates nothing.\n\n* fix(adaptive): restore the queued-pump ordering in the splash fan-out; fix WASD pan\n\nThe port's direct event handling ran the Killed handling (despawn +\nboss-split spawnAt) INSIDE the splash fan-out's enumeration of\nPositions. The transaction commits mutate the live dictionary\nmid-loop, so the enumerator threw 'Collection was modified' — the\nwave-13 crash (the original queued the ApplyDamage messages and the\npump ran them after the fan-out). The fan-out now applies the damage\nto all targets first and handles the events after the enumeration.\nRegression test: a cannon blast killing a pre-damaged boss + runner\nreproduces the exact Dictionary.Enumerator crash on the old code and\npasses on the fix.\n\nWASD pan: panStep returned the X axis for the vertical keys (Up →\n(1,0), Down → (-1,0) instead of (0,±1)) — the original accumulated\ndy for Up/Down. Vertical panning never happened.\n\n* docs(adaptive): wave 20-35 trace assessment for the Defli.Raylib port\n\nAnalyzes defli-adaptive-sim.speedscope.json (518.3 s, windowed game,\nwaves 20-35 warm end-game) against the original Defli baseline docs:\nbusy 7.2% of wall (1.19 ms/frame), AdaptiveSlop 3.7% of wall\n(0.62 ms/frame) vs the original's 5.7% (0.95 ms/frame), 0 GC frames,\n0 write-dispatch samples. Verdict: the adaptive shell is cheaper per\nframe than the MVU shell at comparable load.\n\nAdds the trace-analysis tooling to tools/ (copied from E:\\Defli\\tools\nplus two new probes: probe-structure.fsx validates the evented\nCPU_TIME trace format, probe-nodes.fsx attributes samples per node\nchain) and registers dotnet-trace in dotnet-tools.json.\n\n* feat(adaptive): convert the Defli projections to AMap.joinOn\n\nThe AdaptiveSlop submodule now pins TheAngryByrd/AdaptiveSlop\nfeat/joinon-groupby-reductions (PR #19): AMap.joinOn, groupBy,\ndifference, and the *A reduction family.\n\nThe mapA + tryFind join idiom is replaced at all four projection sites:\n- Homing: Projectiles.Rows x Enemies.Positions joined on the target\n enemy (the profiled hot spot: the per-projectile subgraph is built\n once and the position input swaps in place, no rebuild on update)\n- Enemies.Views: 3-way joinOn composition (Positions x Healths x\n Motions); the defensive zero row for transient mid-transaction\n reads is preserved\n- Enemies.BossPositions: same-key joinOn into Defs with the\n ValueNone drop (choose semantics)\n- Towers.EffectiveDef: same-key joinOn into Levels\n\nBehavior is unchanged: Defli 104/104, AdaptiveSlop 407/407. The pipe\nform lands in the join's right slot (right |> joinOn key mapping\nleft); record-typed lambda parameters carry annotations because the\ninline lambdas elaborate before the map arguments pin their types.\n\n* feat(adaptive): pin AdaptiveSlop to merged master + post-joinOn trace assessment\n\nThe AdaptiveSlop submodule moves from the PR branch head to the merged\nmaster commit 9e81834 (PR #19, squash). The merged joinOn signature\nreorders the arguments to maps-first — (left) (right) (keyOfLeft)\n(mapping) — so the four Defli projection call sites drop the pipe form\n(the pipe would land in the mapping slot). The maps-first order does\nnot fix the inline-lambda record inference, so Homing and the Views\n3-way join keep their type annotations.\n\nTrace assessment (second wave-20-35 session, README section 6):\nbusy 1.19 -> 0.55 ms/frame, AdaptiveSlop 0.62 -> 0.18 ms/frame,\nHoming join 28.5% -> 0.7% of busy despite a heavier session (more\ntowers, 2.6x more projectile churn). The slowliness is not in the CPU\ndata: the longest busy cluster is 10 ms, the 733 ms CPU_TIME spans\nare 1-sample sampling artifacts, and the 413 idle gaps >= 120 ms\ncluster in the boot minutes (collector pauses). GC measured directly\nfrom the GC events with the new F# tools/gcprobe: 188 GCs, every\npause < 1 ms, worst 0.69 ms. Against the seven original Defli traces\nthe port collects 3.6x less often than the waves-33-35 baseline with\na 3x smaller worst pause (Trace A had 37 ms pauses).\n\nNew tools: tools/probe-spikes.fsx (spans + per-minute census),\ntools/probe-minute.fsx (gap histogram), tools/gcprobe (F# TraceEvent\nGC-lifecycle probe).\n\nTests: Defli 104/104, AdaptiveSlop 407/407.\n\n* chore update probes\n\n* chore: bump ref\n\n* feat(adaptive): mirror the MVU Program API; add MonoGame host; decouple\nthe Defli sim\n\nRewrite the adaptive public surface to mirror Mibo.Elmish.Program by\nname and shape, minus the Msg/Cmd/Sub machinery.\n\nAdaptiveProgram<'Frame> is a single record (Init/Update/Observers +\nConfig/Renderers/ServiceRegistrations/AssetsBasePath/FixedStep/HasInput)\nbuilt with AdaptiveProgram.mkProgram init update |> with*. AdaptiveWorld\nis removed; its slots fold into AdaptiveProgram. AdaptiveInit is kept\nbut Disposables is hidden (ofFrameBuilder defaults it to [],\nwithDisposables/withDisposable augment). AdaptiveFixedStepConfig +\nwithFixedStep: the runner sub-steps Update N times via FixedStep.compute\nand forces once (mirrors TickFrame, no Map/'Msg). AdaptiveRaylibGame\nrewritten to take AdaptiveProgram (reverses\nConfig/Renderers/ServiceRegistrations like RaylibGame) and moved to the\nMibo.Adaptive namespace. New AdaptiveMonoGameGame +\nAdaptiveMonoGameProgram mirror MiboGame/MonoGameProgram (DeviceConfig\nwrapper, Game lifecycle). Input is opt-in via AdaptiveProgram.withInput\non both hosts (mirrors Program.withInput/HasInput).\n\nDefli sim decoupled from raylib: Defli.fsproj no longer references\nRaylib-cs. Particle2D uses Mibo.Color; TileInfo.Rect and\nParticle2D.SourceRect dropped (the raylib view edge builds native\nrectangles via MapView.tileRect); VfxModel.Textures moved to VfxView.\n\nAll consumers migrated (Defli.Raylib, Defli.Tests, Mibo.Adaptive.Demo,\nMibo.Adaptive.Tests). 27 adaptive tests pass (incl. 5 new fixed-step);\nheadless sim still 0 B/frame in the paused phase.\n\n* chore(adaptive): remove Defli sample projects and the adaptive demo from the experiments branch\n\nThe Defli sample (src/Defli, src/Defli.Raylib, src/Defli.Tests) moves to\nMibo.Samples/Defli; the adaptive Pong demo (src/Mibo.Adaptive.Demo) is\ndropped. The adaptive framework pieces (Mibo.Adaptive, hosts, tests) stay.\n\n* feat(adaptive): adopt AdaptiveSlop as Mibo.Adaptive; move the adaptive program into Mibo.Core\n\n- Mibo.Adaptive is now the full adaptive-data library (adopted from\n AdaptiveSlop, namespace renamed AdaptiveSlop.Core -> Mibo.Adaptive),\n dependency-free: no Mibo references, no package dependencies beyond\n the SDK's implicit FSharp.Core. The AdaptiveSlop submodule is removed.\n- AdaptiveProgram/AdaptiveHeadless move into Mibo.Core (the adaptive\n counterpart of the MVU shell); hosts keep namespace Mibo.Adaptive.\n- Mibo.Adaptive.Tests is the adopted xunit + FsCheck suite (417 tests);\n the wrapper tests move to Mibo.Core.Tests. Mibo.Adaptive.Benchmarks\n carries the BenchmarkDotNet suite (FDA comparison).\n- Mibo.Adaptive versions itself from its own CHANGELOG.md (0.1.0);\n release.yml gains a dedicated adaptive release job (tag adaptive-vX.Y.Z)\n and the main pack excludes the adaptive package.\n- Docs, AGENTS.md and the design archive are adopted alongside the code.\n\n* feat(adaptive): mark the AdaptiveProgram module experimental\n\nUsing the adaptive program API now emits an FS0057 warning (with\nRequireQualifiedAccess) — consumers get a compile-time signal that the\nadaptive integration carries no stability guarantees. The changelog\nmoves all adaptive entries to [Unreleased]: nothing adaptive has been\nreleased yet.\n\n* ci: fix the release flow pack step and changelog version resolution\n\n- The main release job packs each library explicitly (a loop — dotnet pack\n accepts a single project), keeping Mibo.Adaptive out of the main release;\n the adaptive job packs it separately with its own version.\n- ChangelogFile now resolves via MSBuildThisFileDirectory instead of\n SolutionDir, so per-project builds/packs derive the package version from\n the changelog even outside a solution context (also silences CNG0002).\n\n* fix(adaptive): address the PR review — runner robustness, publish order, zero-alloc gates\n\nRunner (AdaptiveHeadless):\n- RunAsync: catch game-thread exceptions and rethrow them from MoveNextAsync\n (an unhandled background-thread exception killed the process); reject a\n second concurrent enumerator on the same runner.\n- StepUntil: initialize before the first predicate check (a fresh runner\n evaluated the predicate against a default/null frame).\n- Tests: RunAsync tests build the graph inside Init — the adaptive graph is\n confined to its creating (game) thread and Debug owner-thread checks\n rejected roots built on the test thread, aborting the Debug test host.\n\nCI:\n- release.yml: create-release needs create-adaptive-release so Mibo.Adaptive\n reaches NuGet before Mibo.Core (whose nuspec pins it as a dependency).\n- docs.yml: tests run in Debug, catching the #if DEBUG owner-thread checks\n that Release-only CI skipped.\n\nPerf (invariant 5 — zero allocation on clean reads):\n- Version gates on SetToListNode, ConcatListNode, SortListNode and a new\n SubListNode (subA/takeA/skipA track source + bound versions): measured\n 0 bytes per clean read. PollListSourceNode stays un-gated by design — its\n build closure may read extra adaptive inputs — and is documented as the\n exception.\n\nBackends/docs:\n- MonoGame adaptive host renamed AdaptiveGame -> AdaptiveMonoGameGame\n (matches the documented name and AdaptiveRaylibGame symmetry).\n- Both host headers now state input is opt-in via AdaptiveProgram.withInput\n (the code already gated on HasInput).\n- MonoGame 2D: SpriteState.Rotation is degrees (raylib parity); the backend\n converts to radians for SpriteBatch.Draw and lit tessellation; unit docs\n and the corner-rotation test updated.\n\nChore:\n- Changelog: the experimental marker belongs under Added (unreleased feature).\n- Drop the dev-only tools/ (probe scripts, gcprobe) — moved to Mibo.Samples.\n- Restore dotnet-tools.json (no dev-only tools, no line-ending churn).","tree":"24d51de51d2f1cbff6e664d1cf2e95620c70df36"},"mode":"0040000","name":"archive","size":0}],"lastCommit":{"author":{"email":"scythes.lol@gmail.com","name":"Angel D. Munoz","when":"2026-08-11T20:30:59.000000-06:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-08-11T20:30:59.000000-06:00"},"hash":"8e5e3269d1b16fcbc2794688f4403b6d59500c64","message":"Experiments/adaptive architecture (#100)\n\n* feat(adaptive): AdaptiveHeadless runner, AdaptivePong demo, and the Defli tower-defense port\n\nThe adaptive-data architecture experiment (Mibo.Adaptive) plus its two\nreal-scale validations, as an alternative to MVU:\n\n- Mibo.Adaptive: the AdaptiveHeadless runner — pull-based state on\n AdaptiveSlop (State · Projection · Update · Force). State lives in\n changeable roots, derived state in adaptive projections, and the\n runner forces the frame's projections at the end of each Step, so\n draw reads are O(1) and unchanged state recomputes nothing (idle\n Step: 0 recomputes, 0 bytes). Step/StepN/StepUntil/Run/RunAsync and\n observers mirror HeadlessRunner; there is no 'Msg, no Cmd, no Sub —\n handlers write roots and run effects directly.\n- Mibo.Adaptive.Demo: AdaptivePong — PingPong's sim rebuilt on the\n runner, with both a console sim frontend and a raylib frontend.\n- src/Defli: Defli's tower-defense simulation (nine systems, 4,554\n lines) ported from its MVU shell onto AdaptiveHeadless, headless\n first. The World record is the composition root (State); Router.fs\n is the per-frame router, event translation and cold paths (Update);\n Frame.fs packs RenderFrame once per Step (Force). System update/tick\n return only the events they emit — in-place mutation, nothing\n discarded, no recursion. The full Defli test suite (94 tests) passes\n against the port, unchanged in behavior.\n- AdaptiveSlop submodule pinned at eba4b8c (dirty-indicator Version\n fix for journal-and-drain nodes — the suppression chain settles\n bottom-up on read; no pre-read workaround needed).\n- Mibo.slnx, CHANGELOG, and InternalsVisibleTo for the test projects.\n\n* feat(adaptive): windowed raylib host, neutral camera, and the Defli raylib frontend\n\nAdaptiveRaylibGame<'Frame> (Mibo.Raylib) runs an adaptive world in a\nraylib window with the MVU ceremony removed: no Program builder, no\nCmd/Sub machinery, input always registered and polled, services read\nfrom the context in Init/Update. The runner accepts a pre-built\nGameContext, exposes it via AdaptiveContext, and gains Restart\n(dispose world disposables, re-run Init, reset clock, force the first\nframe) consumed by the host after Step.\n\nThe Defli sim camera is now backend-neutral CameraState (target/zoom/\nrotation/worldsize/shake) with pure, unit-tested view math\n(clampToWorld/screenToWorld/viewBounds); the raylib Camera2D is built\nat the view edge. RenderFrame carries the hover/selection/projection\nand Vfx/Map/Diag state so the frontend draws read only the struct.\n\nsrc/Defli.Raylib restores the trimmed view layer from the original\nDefli (Map/Enemies/Towers/Projectiles/Vfx/World/HUD/Diagnostics),\nreading the forced frame; the shell translates input through\nsubscriptions (no Msg/Cmd/Sub), with restart via the world cell +\nRestartRequested. Camera tests were wired into Main.fs (were dropped\nin the milestone-1 port) and extended with the pure math tests.\nRemoved a leftover inline/private FS1113 in Router (startNextWave,\napplyEnemyMsg) that broke the committed tree's build.\n\n* fix(adaptive): per-kind particle buffers in VfxView\n\nThe draw commands are deferred — Renderer2D executes them after the\nview returns, and each command holds the array it was recorded with.\nAll six particle kinds shared one conversion scratch, so every\ncommand drew the LAST kind's particles with its own texture (impact\nsparks stretched over smoke puffs — the square behind the puff).\nOne persistent buffer per kind keeps each command's data isolated;\nsteady state still allocates nothing.\n\n* fix(adaptive): restore the queued-pump ordering in the splash fan-out; fix WASD pan\n\nThe port's direct event handling ran the Killed handling (despawn +\nboss-split spawnAt) INSIDE the splash fan-out's enumeration of\nPositions. The transaction commits mutate the live dictionary\nmid-loop, so the enumerator threw 'Collection was modified' — the\nwave-13 crash (the original queued the ApplyDamage messages and the\npump ran them after the fan-out). The fan-out now applies the damage\nto all targets first and handles the events after the enumeration.\nRegression test: a cannon blast killing a pre-damaged boss + runner\nreproduces the exact Dictionary.Enumerator crash on the old code and\npasses on the fix.\n\nWASD pan: panStep returned the X axis for the vertical keys (Up →\n(1,0), Down → (-1,0) instead of (0,±1)) — the original accumulated\ndy for Up/Down. Vertical panning never happened.\n\n* docs(adaptive): wave 20-35 trace assessment for the Defli.Raylib port\n\nAnalyzes defli-adaptive-sim.speedscope.json (518.3 s, windowed game,\nwaves 20-35 warm end-game) against the original Defli baseline docs:\nbusy 7.2% of wall (1.19 ms/frame), AdaptiveSlop 3.7% of wall\n(0.62 ms/frame) vs the original's 5.7% (0.95 ms/frame), 0 GC frames,\n0 write-dispatch samples. Verdict: the adaptive shell is cheaper per\nframe than the MVU shell at comparable load.\n\nAdds the trace-analysis tooling to tools/ (copied from E:\\Defli\\tools\nplus two new probes: probe-structure.fsx validates the evented\nCPU_TIME trace format, probe-nodes.fsx attributes samples per node\nchain) and registers dotnet-trace in dotnet-tools.json.\n\n* feat(adaptive): convert the Defli projections to AMap.joinOn\n\nThe AdaptiveSlop submodule now pins TheAngryByrd/AdaptiveSlop\nfeat/joinon-groupby-reductions (PR #19): AMap.joinOn, groupBy,\ndifference, and the *A reduction family.\n\nThe mapA + tryFind join idiom is replaced at all four projection sites:\n- Homing: Projectiles.Rows x Enemies.Positions joined on the target\n enemy (the profiled hot spot: the per-projectile subgraph is built\n once and the position input swaps in place, no rebuild on update)\n- Enemies.Views: 3-way joinOn composition (Positions x Healths x\n Motions); the defensive zero row for transient mid-transaction\n reads is preserved\n- Enemies.BossPositions: same-key joinOn into Defs with the\n ValueNone drop (choose semantics)\n- Towers.EffectiveDef: same-key joinOn into Levels\n\nBehavior is unchanged: Defli 104/104, AdaptiveSlop 407/407. The pipe\nform lands in the join's right slot (right |> joinOn key mapping\nleft); record-typed lambda parameters carry annotations because the\ninline lambdas elaborate before the map arguments pin their types.\n\n* feat(adaptive): pin AdaptiveSlop to merged master + post-joinOn trace assessment\n\nThe AdaptiveSlop submodule moves from the PR branch head to the merged\nmaster commit 9e81834 (PR #19, squash). The merged joinOn signature\nreorders the arguments to maps-first — (left) (right) (keyOfLeft)\n(mapping) — so the four Defli projection call sites drop the pipe form\n(the pipe would land in the mapping slot). The maps-first order does\nnot fix the inline-lambda record inference, so Homing and the Views\n3-way join keep their type annotations.\n\nTrace assessment (second wave-20-35 session, README section 6):\nbusy 1.19 -> 0.55 ms/frame, AdaptiveSlop 0.62 -> 0.18 ms/frame,\nHoming join 28.5% -> 0.7% of busy despite a heavier session (more\ntowers, 2.6x more projectile churn). The slowliness is not in the CPU\ndata: the longest busy cluster is 10 ms, the 733 ms CPU_TIME spans\nare 1-sample sampling artifacts, and the 413 idle gaps >= 120 ms\ncluster in the boot minutes (collector pauses). GC measured directly\nfrom the GC events with the new F# tools/gcprobe: 188 GCs, every\npause < 1 ms, worst 0.69 ms. Against the seven original Defli traces\nthe port collects 3.6x less often than the waves-33-35 baseline with\na 3x smaller worst pause (Trace A had 37 ms pauses).\n\nNew tools: tools/probe-spikes.fsx (spans + per-minute census),\ntools/probe-minute.fsx (gap histogram), tools/gcprobe (F# TraceEvent\nGC-lifecycle probe).\n\nTests: Defli 104/104, AdaptiveSlop 407/407.\n\n* chore update probes\n\n* chore: bump ref\n\n* feat(adaptive): mirror the MVU Program API; add MonoGame host; decouple\nthe Defli sim\n\nRewrite the adaptive public surface to mirror Mibo.Elmish.Program by\nname and shape, minus the Msg/Cmd/Sub machinery.\n\nAdaptiveProgram<'Frame> is a single record (Init/Update/Observers +\nConfig/Renderers/ServiceRegistrations/AssetsBasePath/FixedStep/HasInput)\nbuilt with AdaptiveProgram.mkProgram init update |> with*. AdaptiveWorld\nis removed; its slots fold into AdaptiveProgram. AdaptiveInit is kept\nbut Disposables is hidden (ofFrameBuilder defaults it to [],\nwithDisposables/withDisposable augment). AdaptiveFixedStepConfig +\nwithFixedStep: the runner sub-steps Update N times via FixedStep.compute\nand forces once (mirrors TickFrame, no Map/'Msg). AdaptiveRaylibGame\nrewritten to take AdaptiveProgram (reverses\nConfig/Renderers/ServiceRegistrations like RaylibGame) and moved to the\nMibo.Adaptive namespace. New AdaptiveMonoGameGame +\nAdaptiveMonoGameProgram mirror MiboGame/MonoGameProgram (DeviceConfig\nwrapper, Game lifecycle). Input is opt-in via AdaptiveProgram.withInput\non both hosts (mirrors Program.withInput/HasInput).\n\nDefli sim decoupled from raylib: Defli.fsproj no longer references\nRaylib-cs. Particle2D uses Mibo.Color; TileInfo.Rect and\nParticle2D.SourceRect dropped (the raylib view edge builds native\nrectangles via MapView.tileRect); VfxModel.Textures moved to VfxView.\n\nAll consumers migrated (Defli.Raylib, Defli.Tests, Mibo.Adaptive.Demo,\nMibo.Adaptive.Tests). 27 adaptive tests pass (incl. 5 new fixed-step);\nheadless sim still 0 B/frame in the paused phase.\n\n* chore(adaptive): remove Defli sample projects and the adaptive demo from the experiments branch\n\nThe Defli sample (src/Defli, src/Defli.Raylib, src/Defli.Tests) moves to\nMibo.Samples/Defli; the adaptive Pong demo (src/Mibo.Adaptive.Demo) is\ndropped. The adaptive framework pieces (Mibo.Adaptive, hosts, tests) stay.\n\n* feat(adaptive): adopt AdaptiveSlop as Mibo.Adaptive; move the adaptive program into Mibo.Core\n\n- Mibo.Adaptive is now the full adaptive-data library (adopted from\n AdaptiveSlop, namespace renamed AdaptiveSlop.Core -> Mibo.Adaptive),\n dependency-free: no Mibo references, no package dependencies beyond\n the SDK's implicit FSharp.Core. The AdaptiveSlop submodule is removed.\n- AdaptiveProgram/AdaptiveHeadless move into Mibo.Core (the adaptive\n counterpart of the MVU shell); hosts keep namespace Mibo.Adaptive.\n- Mibo.Adaptive.Tests is the adopted xunit + FsCheck suite (417 tests);\n the wrapper tests move to Mibo.Core.Tests. Mibo.Adaptive.Benchmarks\n carries the BenchmarkDotNet suite (FDA comparison).\n- Mibo.Adaptive versions itself from its own CHANGELOG.md (0.1.0);\n release.yml gains a dedicated adaptive release job (tag adaptive-vX.Y.Z)\n and the main pack excludes the adaptive package.\n- Docs, AGENTS.md and the design archive are adopted alongside the code.\n\n* feat(adaptive): mark the AdaptiveProgram module experimental\n\nUsing the adaptive program API now emits an FS0057 warning (with\nRequireQualifiedAccess) — consumers get a compile-time signal that the\nadaptive integration carries no stability guarantees. The changelog\nmoves all adaptive entries to [Unreleased]: nothing adaptive has been\nreleased yet.\n\n* ci: fix the release flow pack step and changelog version resolution\n\n- The main release job packs each library explicitly (a loop — dotnet pack\n accepts a single project), keeping Mibo.Adaptive out of the main release;\n the adaptive job packs it separately with its own version.\n- ChangelogFile now resolves via MSBuildThisFileDirectory instead of\n SolutionDir, so per-project builds/packs derive the package version from\n the changelog even outside a solution context (also silences CNG0002).\n\n* fix(adaptive): address the PR review — runner robustness, publish order, zero-alloc gates\n\nRunner (AdaptiveHeadless):\n- RunAsync: catch game-thread exceptions and rethrow them from MoveNextAsync\n (an unhandled background-thread exception killed the process); reject a\n second concurrent enumerator on the same runner.\n- StepUntil: initialize before the first predicate check (a fresh runner\n evaluated the predicate against a default/null frame).\n- Tests: RunAsync tests build the graph inside Init — the adaptive graph is\n confined to its creating (game) thread and Debug owner-thread checks\n rejected roots built on the test thread, aborting the Debug test host.\n\nCI:\n- release.yml: create-release needs create-adaptive-release so Mibo.Adaptive\n reaches NuGet before Mibo.Core (whose nuspec pins it as a dependency).\n- docs.yml: tests run in Debug, catching the #if DEBUG owner-thread checks\n that Release-only CI skipped.\n\nPerf (invariant 5 — zero allocation on clean reads):\n- Version gates on SetToListNode, ConcatListNode, SortListNode and a new\n SubListNode (subA/takeA/skipA track source + bound versions): measured\n 0 bytes per clean read. PollListSourceNode stays un-gated by design — its\n build closure may read extra adaptive inputs — and is documented as the\n exception.\n\nBackends/docs:\n- MonoGame adaptive host renamed AdaptiveGame -> AdaptiveMonoGameGame\n (matches the documented name and AdaptiveRaylibGame symmetry).\n- Both host headers now state input is opt-in via AdaptiveProgram.withInput\n (the code already gated on HasInput).\n- MonoGame 2D: SpriteState.Rotation is degrees (raylib parity); the backend\n converts to radians for SpriteBatch.Draw and lit tessellation; unit docs\n and the corner-rotation test updated.\n\nChore:\n- Changelog: the experimental marker belongs under Added (unreleased feature).\n- Drop the dev-only tools/ (probe scripts, gcprobe) — moved to Mibo.Samples.\n- Restore dotnet-tools.json (no dev-only tools, no line-ending churn).","tree":"24d51de51d2f1cbff6e664d1cf2e95620c70df36"}}