port(main): apply PR #251 node updates onto public main

[QuQingY]

Port of ComfyResearch-dev PR #251 onto public main

This branch applies the change set of PR #251 ("Merge self-driving-complete node updates into main") onto this repository's main, without carrying over other features from the dev repo's history.

What is included

• New moving_statistics node (backend definition, generated contracts, frontend component, wiring)
• Node definition fixes (metric_compare generalization, representationId runtime resolution with fallback, fourier/information-plane/weight-product warnings + NaN, vocabCap removal from toy-language datasets)
• Engine fixes (multi-slot CE loss, representation collector fallback, information-plane per-feature MI, CPU eigh retry, NDJSON error events)
• Frontend canvas/visualization updates (1D curve transform wiring, tensor selector hub, compare handles, sidebar connectability grouping)
• Tests and refreshed golden snapshots

Conflicts resolved against this repo's refactored baseline

• test_frontend_channel_ledger.py / test_frontend_channel_endstate.py: kept this repo's naming (REGISTERED_FRONTEND_TYPES), bumped counts 51 -> 52 (moving_statistics)
• recorder.py: kept this repo's comment, added the new observable_warnings state
• generate-connection-golden.ts / connectionRules.ts: adopted the shared CONNECTION_PROBE_HANDLE_PAIRS export
• GenericObservableNode.tsx: adopted the representation dropdown / auto-fix logic
• selfDrivingGraph.ts: this repo's lottery path already had no vocabCap; no change needed

Verified on this branch

• comfy_research/nodes/generate.py --check passes (node manifest / generated contracts current)
• validate_defs() registry validation passes
• Backend tests: 16 passed (frontend-channel ledger/endstate, representation resolve smoke)
• Frontend: full npm run build (vite + all codegen verifiers) passes; 87 tests in the PR-related suites pass

[JinxinWonderWorld]

Review result: request changes

I found two merge-blocking regressions and four additional correctness or contract gaps:

1. [P1] moving_statistics changes its React Hook count on the normal empty to computed transition.
2. [P1] saxe_fixed_width_0_07 information-plane binning emits negative state ids, crashes np.bincount, and the exception is silently converted to an empty frame.
3. [P2] generalized Metric compare discards available xTicks and relabels real steps as 0..N-1.
4. [P2] observable warnings are set but never cleared by a later clean run.
5. [P2] sidebar socket metadata omits the new legal tensor_selector.tensor_list to tensor_viz_general or tensor_viz_2d tensor_list path.
6. [P2] vocabCap removal is incomplete: ToyLanguageLmDatasetNode.tsx still reads and writes the removed key at lines 112-115, 225-235, and 319-325. A read-only fallback for old graphs is reasonable, but new edits should write only vocabSize.

Verification: GitHub CI is green; locally, all 320 frontend tests and npm run build plus codegen gates pass, and the targeted backend set passes. A minimal negative-activation model reproduces the Saxe failure as ValueError: list argument must have no negative elements. Current tests do not exercise the moving-statistics render transition or this Saxe option.

Also, git diff --check reports trailing whitespace at the EOF of five newly added files.

[JinxinWonderWorld]

Re-review of 3372334 (incremental from 455ec118).

Verified that this commit fixes the six previously reported issues: React hook ordering, negative Saxe state IDs plus warning propagation, Metric Compare xTicks, stale warning clearing, socket metadata, and vocabCap writeback. Local verification passed: 54 frontend test files / 325 tests, production build plus codegen/manifest gates, 26 targeted backend tests, and git diff --check; GitHub CI is also green.

One curve-correctness issue remains, detailed inline: Moving Statistics does not preserve the selected output step coordinates when stride or minPeriods shortens the series.

[JinxinWonderWorld]

Final re-review of 7fa231f (incremental from 3372334).

The output-anchor math now correctly maps non-uniform ticks and the previous thread is addressed. I cannot approve this head yet because the fix introduces one main-path runtime failure and still drops coordinates for valid chained transforms:

1. Tensor selector → Moving Statistics calls a runtime helper whose import was removed, causing ReferenceError during render.
2. Moving Statistics still reads coordinates from a bespoke graph walker instead of the resolved tensor, so Moving Statistics → Moving Statistics and other valid 1D-series inputs lose xTicks.

Verification: 54 frontend test files / 330 tests passed; production build and codegen/manifest gates passed; 26 targeted backend tests passed; both incremental and full git diff --check passed; GitHub CI is green. A targeted TypeScript check reports five errors in MovingStatisticsNode.tsx, including the missing runtime helper. Details are inline.

[JinxinWonderWorld]

Re-review of 20abac1 (incremental from 7fa231f).

Both prior Moving Statistics blockers are fixed: the unresolved selector helper path was removed, compute now consumes hydrated.xTicks, and all earlier review threads are resolved. I cannot approve this head yet because two correctness gaps remain:

1. Non-finite eigen results raise ValueError outside the retry handling, so the remaining jitter attempts and CPU-double fallback are skipped.
2. MultiSlot CE checks ranks but not that target [B, K] matches logits first two dimensions, allowing a transposed or wrong-slot target with the same element count to compute a silently mispaired loss.

Verification: 54 frontend files / 331 tests passed; production build plus codegen/manifest gates passed; the 6 new backend tests and Ruff undefined-name gate passed; incremental and full git diff --check passed; GitHub CI is green. The full local backend run reached 791 passed / 6 skipped, with 4 unrelated failures because this local environment lacks the repository-pinned antlr4 runtime. Both reported gaps were independently reproduced; details are inline.

[JinxinWonderWorld]

Reviewed head 87885e6, including the incremental fix over 20abac1. The two prior blockers are resolved: MultiSlot CE now enforces target [B,K] against pred [B,K,V] at both trainer and criterion boundaries, and non-finite eigh outputs remain inside the jitter/CPU fallback flow. Local targeted tests: 9 passed; Ruff and diff check passed. Current-head CI passed backend, frontend build/drift guard, and Playwright; all review threads are resolved. No remaining blocking findings on the Standards or Spec axes.