ANIM-b1 3b part 1: the runtime animator as a translated root, with its animator suite

[thejackshelton]

T065a ANIM-b1, PR 3b part 1 of 2: the runtime animator as one translated implementation (binding spec docs/goals/milestone-2-proof/notes/T065-anim-b-spec.md §2 R2, R4, R7, R9, R10, R16, §6.1 b1-3). Based on master; landable alone. It changes no row, no native output and no static case. Part 2 (the device runtime: tables in the state program sources, the display and virtual drivers, the state.ts hunks, the device-anim lane and the native animation rows) follows once this is on master (no-stacks rule).

Why a translated root

The spec's b1-3 lists "Animator (TS, Swift, Kotlin)" in emit/runtime/anim.ts. This PR writes the animator once, in the translator subset, as a new engine root (packages/layout/src/rt-animator.ts), so:

• the Swift and Kotlin animators are generated, not hand-written twice;
• the frame lanes against Chrome judge the very code the devices run, because the parity reference now wraps this root;
• a new p1 suite checks the translations bit for bit against it, on the host (native:swift, native:kotlin) and on every device through device-vectors.

The device glue in part 2 becomes thin: it converts the compiled tables to literals, applies colour writes to views and lays out the patched input.

What changed

• packages/layout/src/rt-animator.ts (new, translated root, subset-clean):
  • the tables as plain data: EasingCode, ValueCode, slots, animation lists, @keyframes, rendered, bases and closure;
  • animatorStart / animatorEvent / animatorAdvance: R4, R7 and R2, on rt-transition and rt-animations;
  • animatorFrame, closureFrame (R9) and frameColors: colours as Chrome draws them;
  • patchInput / patchRoot: lengths back into the engine input as px, % or a calc() sum (R16);
  • lengthBase: endpoints from each assignment's resolved input (R14);
  • animatorBusy: the display driver runs only while true (R3);
  • the four runtime plants as AnimatorFaults;
  • a node missing from the rendered table now throws, where the old reference treated it as hidden.
• rt-interpolate.ts: an additive colorRgba8, built from the same rounding helpers as serializeColor.
• lower/anim-program.ts: animTablesOf, the compiled tables as the root's plain data (layout types only, so the core boundary holds).
• Parity reference (anim-cases.ts): Animator, closureFrame and applyFrame now wrap the root. Their API is unchanged.
• Animator suite:
  • pnpm run parity:anim-vectors writes packages/layout/rt-vectors/animator/cases.json: per frame case, its tables, every assignment's resolved engine input and its frame script as animator steps;
  • harness op rt-animator, with exact-key decoders;
  • corpus suite animator after hit, whose expected results are the TypeScript harness's answers; a line it throws on or refuses fails the build;
  • generate.ts root list; targets.ts device-vectors declaration;
  • regen step anim-vectors and its .gitattributes line.

Changed tests and pins, and why

• translate.test.ts: the corpus size pin gains animator: 18; the other suites keep their sizes.
• device-vectors.test.ts: the animator suite is declared with one case per frame case.
• p1-floor.json: animator appended with DRAGON_FLOOR_WRITE=1; additions only.
• New: packages/layout/test/rt-animator.test.ts (4 tests) and packages/translate/test/rt-animator.test.ts (3 tests).
• No tolerance, check or test is loosened, skipped or removed.

Outside the spec

• packages/layout/src/rt-animator.ts: the spec put the animator in emit/runtime/anim.ts, hand-written per language. Here it is one translated root.
• packages/translate/harness/harness.ts, src/corpus.ts, src/generate.ts and test/rt-animator.test.ts: the animator suite.
• packages/parity/src/cli/anim-vectors.ts and the parity:anim-vectors script.

North-star delta

None in this part: no row changes. Part 2's device-anim rows bring the −8 on ios and android.

Deferred (device step)

The device run of p1/animator through device-vectors is the landing driver's.

What passed

• pnpm typecheck passes, and pnpm run layout:subset reports 0 violations.
• pnpm run parity:anim-report against the committed Chrome captures, now through the translated root: 18 cases, 2328 samples (4880 values, 11704 boxes, 582 chrome-dual samples, 18 settles), failed 0.
• anim-frames.test.ts: 21/21, and every runtime and rt plant still fails its own fixture.
• /tmp/heavy-lease.sh pnpm run native:swift and native:kotlin: status pass, animator 18/18 on both.
• Targeted runs pass: rt-animator (layout and translate), rt-vectors, rt-harness, translate, anim-frames, device-vectors, regen, ua, anim-program and chrome-ports (122 passed, 2 skipped).
• Step-3 audit:
  • the rendered-table lookup now throws on a missing node, with a test that fails on the old silent false;
  • colorRgba8 agrees with serializeColor on rounding and alpha edge cases;
  • the advance step refuses negative, NaN and infinite steps;
  • an event to an assignment the tables lack is refused, and the suite treats a refusal as a build failure, not a computed result.
• /tmp/heavy-lease.sh pnpm regen: fixed point, outputs in their own commit (Regenerate: pnpm regen).
• /tmp/heavy-lease.sh pnpm test: Tests 15 failed | 4760 passed | 2 skipped (4777). Every failure is device step pending (lanes, lanes-records, device-failures, p6a-promotion, lanes-concurrent, land.test evidence:stamp), except planted-swift.test.ts, which timed out under machine load and passes when rerun alone.

🤖 Generated with Claude Code