fix(studio): kill false-positive shadow GSAP fidelity mismatches

[vanceingalls]

What

Kills two false-positive classes in the SDK shadow GSAP value-fidelity diff (sdkShadowGsapFidelity.ts).

1. Float precision — numericEqual compared exactly, so SDK-computed 3.0999999999999996 vs server 3.1 flagged as drift. Now a relative epsilon (abs(a-b) <= 1e-6 * max(1,|a|,|b|)); real 2 vs 1 still flags.
2. Selector-form divergence — [data-hf-id="X"] (SDK writer) vs .class/#id (server writer) for the same element produced phantom present/absent pairs. makeSelectorResolver now keys tweens by resolved element (incl. nodes with no data-hf-id), unifying the forms.

Why

Surfaced by production SDK-shadow parity telemetry — gsap_fidelity was the noisiest real-traffic op; both are diff-harness artifacts, not SDK drift.

Tests

Epsilon (clean + real-drift) + selector-unification for #id/.class/[data-hf-id].

🤖 Generated with Claude Code

[vanceingalls]

• refactor(studio): dedup shadow numeric-equal + GSAP script extraction #1516
• fix(studio): serialize per-animationId GSAP meta commits (shadow ease race) #1512
• fix(sdk): agree removeElement/getElement on duplicate bare ids #1511
• fix(sdk): setStyle removes hyphenated properties (was kebab/camel key mismatch) #1510
• feat(studio): shadow telemetry for GSAP keyframe ops (gsap_keyframe) #1509
• fix(studio): suppress shadow-parity false positives in timing + text #1508
• fix(studio): kill false-positive shadow GSAP fidelity mismatches #1507 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[james-russo-rames-d-jusso]

Cluster context — Reviewing this alongside #1508 (timing+text FP suppression) and #1516 (dedup of numericEqual + extractGsapScript). All three target SDK shadow parity noise reduction.

Looks good to me at comment quality — no blockers.

Verified

• Tolerance shape is correct: 1e-6 * max(1, |a|, |b|) is a relative epsilon with a floor of 1e-6, so 3.1 vs 3.0999999999999996 (delta ~4e-16) absorbs, 2 vs 1 (delta 1, tolerance 2e-6) still flags. Tight enough that 0.5 vs 0.49 would also flag (delta 0.01 ≫ tolerance 5e-7). Math checks out.
• True-positive guardrail test (STILL flags a real integer duration drift (2 vs 1)) is present alongside the FP-suppression test. This is the pattern I want to see on every false-positive suppression PR.
• makeSelectorResolver change correctly handles the no-data-hf-id case via a per-node WeakMap-keyed node:<n> token, namespaced with hfid:/node: prefixes so canonical keys can't collide with raw-selector fallbacks. Three tests cover the path (#id/.class/[data-hf-id] collapse; SDK-vs-server selector unification; attribute-less-element unification). Resolver export is needed for the test, fine.

Concerns (non-blocking)

• Ambiguous-selector ponytail — querySelector first-match could still misroute if a class is shared across multiple elements. The comment acknowledges this; safe for current studio templates (one tween per element). Worth a console.warn or telemetry tick if querySelectorAll(sel).length > 1, so when it bites you'll see it in production rather than as a phantom mismatch. Not a blocker.
• nodeKeys capture — the closure-captured nextNode counter and WeakMap are reused across resolver invocations within the same call, so an element gets a stable key for the lifetime of the resolver. Good. Just confirming this is intentional (it is — re-querying the same selector twice must return the same key).

Cross-PR note

• #1516 hoists this PR's numericEqual body into a shared relEqual in sdkShadowNumeric.ts and re-imports it here. Tolerance formula is preserved verbatim (1e-6 * Math.max(1, |a|, |b|)). The extractGsapScript export in #1516 also preserves the marker set (gsap / __timelines / ScrollTrigger) verbatim from this PR. Dedup is semantically clean.

Stamp recommendation — fine to land once HF-runtime/SDK reviewer (Miga) signs off on the selector-unification semantics. I'm not stamping (review at --comment quality).

— Rames D Jusso

[miga-heygen]

Approve. The relative-epsilon float comparison and upgraded selector resolver cleanly kill the two false-positive classes.

One note for follow-up: canonicalProps still uses JSON.stringify for property value comparison — if the same float-formatting noise that affected duration also appears in tween property values (opacity, x, y), those fields could still produce false positives. Consider adding epsilon-aware comparison there too (or round numeric values to fixed precision before stringifying).

The node:<n> WeakMap-backed resolver is sound — identity-based stability ensures consistency within a single gsapFidelityMismatches invocation regardless of call order.

Tests are thorough: both positive (noise suppressed) and negative (real drift caught) cases present.

— Miga