perf: faster first frames, playback and loading, with less memory

.agentx/experiments/2026-10-27-performance-improvements/2026-10-27-01-baseline-discovery.json

@@ -0,0 +1,88 @@
+{
+  "schemaVersion": 1,
+  "id": "f0287285-7a36-4cb5-8eef-94cdd163e5a7",
+  "loopId": "9e7a8b6d-0104-4e72-a31e-3b168c0b3881",
+  "index": 1,
+  "type": "baseline",
+  "name": "Baseline discovery",
+  "slug": "baseline-discovery",
+  "path": "2026-10-27-01-baseline-discovery.json",
+  "createdAt": "2026-10-27T20:14:29Z",
+  "startedAt": "2026-10-27T20:14:29Z",
+  "finishedAt": "2026-10-27T20:16:00Z",
+  "state": "done",
+  "status": "success",
+  "checkpointCommit": "a9eba8e8fb24625468974d341186c970ecfdad3c",
+  "resultCommit": "a9eba8e8fb24625468974d341186c970ecfdad3c",
+  "rollbackCommand": "git -C \"/Users/dinko/work-private/projects/dXn/dXn.VectorAnimation\" reset --hard a9eba8e8fb24625468974d341186c970ecfdad3c",
+  "repositories": [
+    {
+      "name": "main",
+      "path": ".",
+      "role": "main",
+      "checkpointCommit": "a9eba8e8fb24625468974d341186c970ecfdad3c",
+      "resultCommit": "a9eba8e8fb24625468974d341186c970ecfdad3c",
+      "rollbackCommand": "git -C \"/Users/dinko/work-private/projects/dXn/dXn.VectorAnimation\" reset --hard a9eba8e8fb24625468974d341186c970ecfdad3c",
+      "status": "success"
+    }
+  ],
+  "repositoryCommits": [
+    {
+      "name": "main",
+      "path": ".",
+      "role": "main",
+      "commit": "a9eba8e8fb24625468974d341186c970ecfdad3c"
+    }
+  ],
+  "atomicDiscipline": {
+    "oneHypothesisOnly": true,
+    "singleImplementationAttempt": true,
+    "noUnplannedFixesInsideExperiment": true,
+    "failedAttemptRequiresImmediateRevert": true,
+    "successfulAttemptRequiresCommitBeforeNextExperiment": true,
+    "rule": "Run exactly the printed hypothesis. If the attempted change fails, degrades, or suggests a different fix, revert to the checkpoint and finish this experiment as failure. Start a new experiment for the next idea."
+  },
+  "primaryTarget": "native release build on macOS over the 53-file corpus (artifacts-temp/corpus)",
+  "hypothesisMarkdown": "## Baseline Discovery\n\nRead-only discovery of where playback time, allocations, memory and time to first frame go in the scene model (`Scene::frame_at`), the importers, the reference renderer and the NSH player (`source::read`, `plan::compile`, rasters, the component's load path). No source edits.",
+  "measurementPlanMarkdown": "### Measurement Plan\n\n- `cargo build --release -p dxn_vector_animation_check` then `vector-animation-check profile artifacts-temp/corpus` (53 corpus files, 200 points, 3 pixels a point): plan compile time, graph nodes and breakpoints, canvases, layers, ops a frame, engine-thread ms a frame.\n- Source reading of the hot paths.\n- Gap: no allocation counter, no import or time-to-first-frame timing, no `frame_at` timing exists in the repository: the next experiment adds a corpus bench (measurement enablement only).",
+  "expectedOutcomeMarkdown": "### Expected Outcome\n\nStarting numbers and a ranked backlog of single-mechanism candidates.",
+  "diagnostics": {
+    "playwright": [],
+    "rayx": [],
+    "counters": [],
+    "artifacts": []
+  },
+  "measurements": [
+    {
+      "metric": "plan compile, sum over 53 files",
+      "before": "~70 ms (max 21.7 ms repeater, 18.2 ms merge paths)",
+      "after": "",
+      "delta": "",
+      "source": "vector-animation-check profile artifacts-temp/corpus"
+    },
+    {
+      "metric": "graph nodes, sum",
+      "before": "~6,100",
+      "after": "",
+      "delta": "",
+      "source": "profile"
+    },
+    {
+      "metric": "graph breakpoints, sum",
+      "before": "~124,000 (thunder.svg 29,142; blend-28 22,483; repeater 17,107; merge paths 12,634)",
+      "after": "",
+      "delta": "",
+      "source": "profile"
+    },
+    {
+      "metric": "engine-thread work a frame",
+      "before": "0 except tristate 0.03 ms, onboarding 0.22 ms (moving blur rasters), merge paths 0.019 ms (sampled)",
+      "after": "",
+      "delta": "",
+      "source": "profile"
+    }
+  ],
+  "reportMarkdown": "## Baseline Discovery Finished\n\n### Where the work is\n\n- **Time to first frame (player)**: `component.rs` reads the source off thread, then compiles the plan **twice** (`plan::compile` at identity only to learn native/mixed, then `plan::compile_for` at the fitted base), clones the whole scene for `played` (`Rc::new(scene.clone())`), and per raster unit runs `plan::in_view` (deep clone of the scene root), `plan::still` (17 `frame_at` calls compared) and the first `plan::raster` (render + PNG encode + base64 data URI).\n- **Playback (UI thread)**: NSH evaluates the graph (nodes, interpolation breakpoints) and replays canvases whose values move; fewer nodes, breakpoints and ops are less work each frame and less memory. Breakpoints dominate the graph's size (124k over the corpus).\n- **Playback (engine thread / reference)**: `Scene::frame_at` allocates per shape: a `Vec<Path>` of its paths, a merged `Path`, then clones that path for the fill and again for the stroke; `Animated::Static` paths are cloned on every sample; gradient stops are collected per sample. `from_frame` + painter run each frame for sampled parts. Moving rasters render, convert premultiplied pixels into a new Vec and PNG-encode each frame.\n- **Import**: SVG (roxmltree + CSS + SMIL) and Lottie (serde_json into a Value tree, then conversion) are not timed or allocation-counted anywhere.\n\n### Ranked candidate backlog\n\n1. Measurement: a corpus bench with an allocation counter (read, compile, time to first frame, retained bytes, frame_at and engine work a frame, graph size). Mechanism: measurement only.\n2. Player: compile the plan once on load instead of twice (TTFF, allocations).\n3. Core `frame_at`: move the shape's path into the last item that takes it instead of cloning for each (allocations a frame).\n4. Core `frame_at`: a shape of one path without trim or path transform takes it without the intermediate `Vec<Path>` and merged copy.\n5. Plan: `still()` short-circuits when nothing in the scene is animated (TTFF for still rasters).\n6. Plan/graph: interpolation breakpoints thinned where a straight line stands for the samples (memory, UI-thread work) — needs pixel-equivalence via `tests/plans.rs`.\n7. Graph: dedupe identical interpolation nodes / constant folding (nodes).\n8. Rasters: PNG encode without the intermediate RGBA Vec; cheaper filter (engine work a frame).\n9. Importers: avoid intermediate `serde_json::Value` copies / string allocations (TTFF).\n10. Track sampling: `Animated::Static` sample by reference where callers only read.\n\nUncertain: tiny-skia render cost (external), NSH graph evaluation (external, not measurable here).\n\nAlready simple, leave: `Track::sample` binary search, `fit`, `Playback`.",
+  "decisionMarkdown": "### Decision\n\nBaseline recorded; no source changed. Next: the measurement harness experiment.",
+  "notes": "Raw profile output: artifacts-temp/agentx-experiments/performance-improvements/01-baseline/profile.txt (transient)."
+}

.agentx/experiments/2026-10-27-performance-improvements/2026-10-27-02-corpus-bench-harness.json

@@ -0,0 +1,116 @@
+{
+  "schemaVersion": 1,
+  "id": "08798e18-46c1-499e-bcbb-9ac3dd3b826f",
+  "loopId": "9e7a8b6d-0104-4e72-a31e-3b168c0b3881",
+  "index": 2,
+  "type": "hypothesis",
+  "name": "Corpus bench harness",
+  "slug": "corpus-bench-harness",
+  "path": "2026-10-27-02-corpus-bench-harness.json",
+  "createdAt": "2026-10-27T20:16:05Z",
+  "startedAt": "2026-10-27T20:16:05Z",
+  "finishedAt": "2026-10-27T20:18:20Z",
+  "state": "done",
+  "status": "success",
+  "checkpointCommit": "a9eba8e8fb24625468974d341186c970ecfdad3c",
+  "resultCommit": "8b0bf9378dbd560fbdfaef63e91cd49811441164",
+  "rollbackCommand": "git -C \"/Users/dinko/work-private/projects/dXn/dXn.VectorAnimation\" reset --hard a9eba8e8fb24625468974d341186c970ecfdad3c",
+  "repositories": [
+    {
+      "name": "main",
+      "path": ".",
+      "role": "main",
+      "checkpointCommit": "a9eba8e8fb24625468974d341186c970ecfdad3c",
+      "resultCommit": "8b0bf9378dbd560fbdfaef63e91cd49811441164",
+      "rollbackCommand": "git -C \"/Users/dinko/work-private/projects/dXn/dXn.VectorAnimation\" reset --hard a9eba8e8fb24625468974d341186c970ecfdad3c",
+      "status": "success"
+    }
+  ],
+  "repositoryCommits": [
+    {
+      "name": "main",
+      "path": ".",
+      "role": "main",
+      "commit": "8b0bf9378dbd560fbdfaef63e91cd49811441164"
+    }
+  ],
+  "atomicDiscipline": {
+    "oneHypothesisOnly": true,
+    "singleImplementationAttempt": true,
+    "noUnplannedFixesInsideExperiment": true,
+    "failedAttemptRequiresImmediateRevert": true,
+    "successfulAttemptRequiresCommitBeforeNextExperiment": true,
+    "rule": "Run exactly the printed hypothesis. If the attempted change fails, degrades, or suggests a different fix, revert to the checkpoint and finish this experiment as failure. Start a new experiment for the next idea."
+  },
+  "primaryTarget": "wasm/web with headed Playwright and WebGPU enabled",
+  "hypothesisMarkdown": "## Corpus Bench Harness Hypothesis\n\nA `bench` subcommand of `vector-animation-check`, with a counting global allocator (`stats_alloc`), measures reproducibly over the corpus: time to first frame (read, compile as the component does it twice, scene kept, first rasters and sampled frames), its allocations and bytes, bytes held after, the engine thread's work a frame, the reference `frame_at` a frame, and the graph's size.\n\n**Tested mechanism:** measurement enablement only.\n**Allowed changes:** `crates/dxn_vector_animation_check` (new `bench.rs`, `stats_alloc` dependency, dispatch), README line.\n**Forbidden:** any optimization.",
+  "measurementPlanMarkdown": "### Measurement Plan\n\n`vector-animation-check bench artifacts-temp/corpus` three times (release); totals agree within 10%.",
+  "expectedOutcomeMarkdown": "### Expected Outcome\n\nRepeatable totals; allocation counts identical run to run.",
+  "diagnostics": {
+    "playwright": [],
+    "rayx": [],
+    "counters": [],
+    "artifacts": []
+  },
+  "measurements": [
+    {
+      "metric": "read ms (sum)",
+      "before": "n/a",
+      "after": "73.9",
+      "delta": "",
+      "source": "bench median of 3"
+    },
+    {
+      "metric": "compile ms (sum)",
+      "before": "n/a",
+      "after": "179.6",
+      "delta": "",
+      "source": "bench"
+    },
+    {
+      "metric": "time to first frame ms (sum)",
+      "before": "n/a",
+      "after": "253.5",
+      "delta": "",
+      "source": "bench"
+    },
+    {
+      "metric": "ttff allocations / MB allocated",
+      "before": "n/a",
+      "after": "2,867,296 / 3061 MB",
+      "delta": "",
+      "source": "bench"
+    },
+    {
+      "metric": "held MB",
+      "before": "n/a",
+      "after": "979.7",
+      "delta": "",
+      "source": "bench"
+    },
+    {
+      "metric": "frame_at us/frame (sum) / allocs",
+      "before": "n/a",
+      "after": "354 / 7789",
+      "delta": "",
+      "source": "bench"
+    },
+    {
+      "metric": "engine us/frame (sum) / allocs",
+      "before": "n/a",
+      "after": "833 / 370",
+      "delta": "",
+      "source": "bench"
+    },
+    {
+      "metric": "graph nodes / breakpoints / ops",
+      "before": "n/a",
+      "after": "5026 / 115,936 / 45,331",
+      "delta": "",
+      "source": "bench"
+    }
+  ],
+  "reportMarkdown": "## Corpus Bench Harness Finished\n\n**Passed Status**: true: three runs within 2% on time, allocations identical.\n\n**Attempt Count:** 1. **Out-of-scope changes:** none.\n\nHot spots it shows: merge paths (read 40 ms, 2 GB allocated, 331 MB held), the repeater (compile 43 ms, 372 MB allocated, 222 MB held), onboarding (compile 60 ms; 0.78 ms a frame of moving blur rasters), thunder.svg (29k breakpoints), timeline.svg (frame_at 57 us, 1804 allocs a frame).\n\n| Repo | Path | Checkpoint | Result Commit | Status |\n| --- | --- | --- | --- | --- |\n| main | . | a9eba8e | 8b0bf93 | committed |",
+  "decisionMarkdown": "### Decision\n\nCommitted: 8b0bf93 feat(check): bench a folder's load, first frame, allocations and frames.",
+  "notes": ""
+}

.agentx/experiments/2026-10-27-performance-improvements/2026-10-27-03-switch-detection-without-sampled-paths.json

@@ -0,0 +1,88 @@
+{
+  "schemaVersion": 1,
+  "id": "11eedca8-3f9f-46b5-a2dc-83044d989326",
+  "loopId": "9e7a8b6d-0104-4e72-a31e-3b168c0b3881",
+  "index": 3,
+  "type": "hypothesis",
+  "name": "Switch detection without sampled paths",
+  "slug": "switch-detection-without-sampled-paths",
+  "path": "2026-10-27-03-switch-detection-without-sampled-paths.json",
+  "createdAt": "2026-10-27T20:19:04Z",
+  "startedAt": "2026-10-27T20:19:04Z",
+  "finishedAt": "2026-10-27T20:20:24Z",
+  "state": "done",
+  "status": "success",
+  "checkpointCommit": "8b0bf9378dbd560fbdfaef63e91cd49811441164",
+  "resultCommit": "050fa8f45d1d2f58fd95b45beb3001a88cc9a5b6",
+  "rollbackCommand": "git -C \"/Users/dinko/work-private/projects/dXn/dXn.VectorAnimation\" reset --hard 8b0bf9378dbd560fbdfaef63e91cd49811441164",
+  "repositories": [
+    {
+      "name": "main",
+      "path": ".",
+      "role": "main",
+      "checkpointCommit": "8b0bf9378dbd560fbdfaef63e91cd49811441164",
+      "resultCommit": "050fa8f45d1d2f58fd95b45beb3001a88cc9a5b6",
+      "rollbackCommand": "git -C \"/Users/dinko/work-private/projects/dXn/dXn.VectorAnimation\" reset --hard 8b0bf9378dbd560fbdfaef63e91cd49811441164",
+      "status": "success"
+    }
+  ],
+  "repositoryCommits": [
+    {
+      "name": "main",
+      "path": ".",
+      "role": "main",
+      "commit": "050fa8f45d1d2f58fd95b45beb3001a88cc9a5b6"
+    }
+  ],
+  "atomicDiscipline": {
+    "oneHypothesisOnly": true,
+    "singleImplementationAttempt": true,
+    "noUnplannedFixesInsideExperiment": true,
+    "failedAttemptRequiresImmediateRevert": true,
+    "successfulAttemptRequiresCommitBeforeNextExperiment": true,
+    "rule": "Run exactly the printed hypothesis. If the attempted change fails, degrades, or suggests a different fix, revert to the checkpoint and finish this experiment as failure. Start a new experiment for the next idea."
+  },
+  "primaryTarget": "wasm/web with headed Playwright and WebGPU enabled",
+  "hypothesisMarkdown": "## Switch detection without sampled paths Hypothesis\n\nBisecting where a morphing outline switches commands sampled the whole path each step (60 steps a switch) only to compare its commands. `commands_at` picks the keyframe whose commands `Track::sample` would give, without making the path: the same switch times, far less work.\\n\\n**Tested mechanism:** switch detection by keyframe choice. **Allowed:** `plan.rs` switches. **Forbidden:** other compile changes.",
+  "measurementPlanMarkdown": "### Measurement Plan\n\nRelease bench over the 53 corpus files (median of 3 runs, best of 5 each), plus the focused tests of the touched area.",
+  "expectedOutcomeMarkdown": "### Expected Outcome\n\nThe targeted metric improves with identical graph output and passing focused tests.",
+  "diagnostics": {
+    "playwright": [],
+    "rayx": [],
+    "counters": [],
+    "artifacts": []
+  },
+  "measurements": [
+    {
+      "metric": "compile ms",
+      "before": "179.6",
+      "after": "158.0",
+      "delta": "",
+      "source": "scratchpad bench.sh: median of 3 `vector-animation-check bench artifacts-temp/corpus`"
+    },
+    {
+      "metric": "ttff MB allocated",
+      "before": "3061",
+      "after": "1344",
+      "delta": "",
+      "source": "scratchpad bench.sh: median of 3 `vector-animation-check bench artifacts-temp/corpus`"
+    },
+    {
+      "metric": "ttff allocations",
+      "before": "2,867,296",
+      "after": "2,617,700",
+      "delta": "",
+      "source": "scratchpad bench.sh: median of 3 `vector-animation-check bench artifacts-temp/corpus`"
+    },
+    {
+      "metric": "graph nodes/breaks/ops",
+      "before": "5026/115936/45331",
+      "after": "5026/115936/45331 (identical)",
+      "delta": "",
+      "source": "scratchpad bench.sh: median of 3 `vector-animation-check bench artifacts-temp/corpus`"
+    }
+  ],
+  "reportMarkdown": "## Switch detection without sampled paths Finished\n\n**Passed Status**: True\n\n**Attempt Count:** 1. **Out-of-scope changes:** none.\n\nCompile -12%, bytes allocated at load -56% (merge paths). Tests: plans, moving_outlines, partial_sampling pass; clippy clean.\n\n| Repo | Path | Checkpoint | Result Commit | Status |\n| --- | --- | --- | --- | --- |\n| main | . | 8b0bf93 | 050fa8f | committed |",
+  "decisionMarkdown": "### Decision\n\nCommitted 050fa8f: perf(nsh): find where outlines switch commands without making the paths",
+  "notes": ""
+}