Delegate Coords and Spool task construction to the node layer

[d-chambers]

Description

Continues the canvas/headless parity work from #47: Coords and Spool were the last two widgets that re-implemented node logic instead of delegating to it, so the task that ran on the canvas could differ from the task exported into a saved workflow, and a headless run could diverge from the canvas. After this PR the node layer is the single owner of task construction and validation for both.

Coords

• CoordsTask now validates eagerly and executes lazily: preflight(patch) runs exactly the validation run() would, raising a structured CoordsValidationError (kind/label/detail). The widget builds its task only via NODE_SPEC.build_task, preflights it, and maps failures onto its existing error banners — its four duplicated validators (mapping, selection, transpose, set-coords) are deleted, along with the eight-branch hand-built task constructor. Canvas-run and exported tasks are now always identical.
• resolve_set_coord is a shared free function used by run() and the widget's draft validation, so both accept and reject exactly the same set-coords inputs.
• coords_task_from_params promotes the set_coords draft fields the way canvas rehydration does (drafts win; the *_applied_* mirror is used only when no draft is present). A headless author filling only set_coords_dim/start/stop/step previously got a silent no-op; now they get the coordinate update.
• One deliberate UX change: data-flipping a non-dimension coordinate now reports through the specific "Invalid flip selection" banner at preflight instead of a generic operation failure at runtime (one test updated accordingly).

Spool

• New apply_spool_transforms + parse_spool_scalar in nodes/spool.py are the single definition of the select → chunk pipeline, used by SpoolTask.run, SpoolTransformTask.run, the widget's snapshot executor (both settings and input-spool modes, previously two near-identical 30-line blocks), and the display recompute.
• Parity fix: a chunk value that parses to None (e.g. the literal text None) now disables chunking headlessly — previously it chunked with a None value while the canvas no-opped. Ellipsis (...) parsing is preserved (ast.literal_eval handles it).
• Deleted widget copies: _apply_chunk_transform, _apply_select_transform, _parse_chunk_scalar, _iter_active_select_filters, and the test-only _snapshot_loader/_load_from_example; their tests now exercise load_spool_from_settings and the shared transforms at the node layer.

Net −180 lines in src/. CHANGELOG updated under Unreleased → Fixed.

Counterpart review: the Codex CLI was unavailable (usage limit until Aug 8), so no counterpart review was run; an adversarial self-review record covering the same risk areas (closure late-binding, validation ordering, draft-promotion precedence, Ellipsis parsing, settings-sync freshness) is in the untracked .scratch/review_node_delegation.md. That self-review caught and fixed one real issue: the draft-promotion precedence initially preferred the applied mirror over drafts, which contradicts canvas rehydration semantics.

Checklist

I have (if applicable):

• referenced the GitHub issue this PR closes.
• documented new behavior with docstrings and a CHANGELOG entry.
• included tests (new tests/test_nodes/test_coords.py; node-level source-routing/chunk-parsing tests in test_spool_node.py; widget tests repointed at the delegated paths).
• run prek run --all-files and pytest tests (2106 passed, 44 skipped; one unrelated canvas test flaked once in test_orange_qt.py and passes in isolation, with its file, and on a full-suite rerun).
• for hot-path changes, run python scripts/bench_compare.py --baseline main (table below) and added the benchmark label.

Benchmarks

Full suite (core + qt) vs origin/main — threshold 20%: 0 slower, 0 faster, 33 compared, control −6.1% (ok):

benchmark                                                baseline       head    median       min status
-------------------------------------------------------------------------------------------------------
...estWaterfallBenchmarks::test_compute_default_levels     9.19ms     9.93ms    +8.0%    +1.9%   --
...aterfallBenchmarks::test_render_alternating_patches    79.20ms    84.52ms    +6.7%    -3.9% noisy   (baseline rel stdev 13%)
...s.py::TestTaskBenchmarks::test_callable_source_hash    980.5us     1.01ms    +2.7%    +1.8%   --
...marks.py::TestTaskBenchmarks::test_task_fingerprint     6.01ms     6.17ms    +2.5%    +0.4%   --
...marks.py::TestTaskBenchmarks::test_port_spec_cached    948.3us    965.7us    +1.8%    +0.1% noisy   (baseline rel stdev 11%)
...py::TestWaterfallBenchmarks::test_render_same_patch    86.63ms    88.17ms    +1.8%    +8.6%   --
...stTaskBenchmarks::test_function_task_class_creation    322.1us    327.6us    +1.7%    -1.6%   --
...rks.py::TestGraphBenchmarks::test_validate_uncached    311.3us    315.9us    +1.5%    -0.5%   --
...py::TestGraphBenchmarks::test_deserialize_from_dict    113.7us    114.7us    +0.9%    +0.7%   --
...TestPlotAxesBenchmarks::test_nearest_value_unsorted     1.68ms     1.69ms    +0.5%    +1.1% noisy   (baseline rel stdev 11%)
...rks.py::TestGraphBenchmarks::test_serialize_to_dict     96.5us     96.8us    +0.4%    -0.9%   --
...chmarks.py::TestTaskBenchmarks::test_port_spec_cold    158.0us    157.9us    -0.1%    -1.5%   --
...estWaterfallBenchmarks::test_should_reset_identical     1.56ms     1.55ms    -0.6%    +0.7% noisy   (baseline rel stdev 12%)
...::TestGraphBenchmarks::test_ensure_validated_cached    677.5us    672.3us    -0.8%    -2.1%   --
...s.py::TestWiggleBenchmarks::test_render_small_patch   135.73ms   134.03ms    -1.3%    +0.8%   --
...s.py::TestTaskBenchmarks::test_extract_return_names     1.19ms     1.18ms    -1.4%    -0.2%   --
...arks.py::TestGraphBenchmarks::test_build_chain_pipe    738.4us    724.7us    -1.9%    -0.4%   --
...TestPlotAxesBenchmarks::test_nearest_axis_ascending     1.77ms     1.74ms    -1.9%    -2.6%   --
...ks.py::TestExecutionBenchmarks::test_get_provenance     65.7us     64.3us    -2.0%    -1.1%   --
...arks.py::TestGraphBenchmarks::test_topological_sort    255.8us    249.4us    -2.5%    -3.8%   --
...s.py::TestImportBenchmarks::test_import_spool_utils     5.70ms     5.52ms    -3.3%    -1.6% noisy   (baseline rel stdev 17%)
...chmarks.py::TestExecutionBenchmarks::test_run_chain    308.0us    297.8us    -3.3%    +0.2%   --
...marks.py::TestExecutionBenchmarks::test_run_diamond    202.9us    195.8us    -3.5%    -0.9%   --
...chmarks.py::TestExecutionBenchmarks::test_map_chain     1.48ms     1.40ms    -5.1%    -0.9% noisy   (baseline rel stdev 16%)
...mplingBenchmarks::test_percentile_levels_subsampled    925.0us    877.9us    -5.1%    -0.3%   --
...estPlotAxesBenchmarks::test_nearest_axis_descending     2.01ms     1.91ms    -5.2%    -5.0%   --
...::TestExecutionBenchmarks::test_build_runtime_specs    854.9us    810.4us    -5.2%    -0.1% noisy   (baseline rel stdev 16%)
...nnotationMaskBenchmarks::test_box_mask_object_dtype    10.79ms    10.02ms    -7.2%    -4.1%   --
...s.py::TestExecutionBenchmarks::test_stream_pipeline    106.2us     98.2us    -7.6%    -1.4% noisy   (baseline rel stdev 23%)
...ks.py::TestAnnotationMaskBenchmarks::test_span_mask    395.9us    365.7us    -7.6%    -5.7%   --
...TestAnnotationMaskBenchmarks::test_box_mask_numeric     3.03ms     2.75ms    -9.2%    -1.2% noisy   (baseline rel stdev 11%)
...rks.py::TestSamplingBenchmarks::test_strided_sample    157.8us    143.2us    -9.2%    -6.0% noisy   (baseline rel stdev 14%)
...arks.py::TestImportBenchmarks::test_import_workflow    11.02ms     9.19ms   -16.6%    -6.2% noisy   (baseline rel stdev 12%)

The worst median (+8.0%, waterfall default-levels) is in code this PR does not touch and its min-delta is +1.9% — machine noise, not a regression.