refactor(events): remove event.selections and read state.selections

[jvdd]

Closes #86.

Change

• InteractionEvent no longer has selections. /dashboard/update passes spec.state.selections to FlexEngine.process, which takes them as a new selections parameter (next to viewports_by_figure, the same order as build_cubes). Both entry points type it Sequence[SelectionState].
• The engine semantics do not change: _active_selections still returns [] for init and deselect. The server cache and the layer policy treat deselect as unfiltered.
• JS: no event carries selections any more. postDashboardUpdate captures DASHBOARD_SPEC.state.selections in the same synchronous step that serializes the spec, and applies the response with that list (isUnfilteredBaseForFigure). _fvFigureCacheEligible and fvEnsureOverlayBackground read the state.
• JS: postDashboardUpdate is the one place that picks selection or deselect. Callers that report a selection change send selection; it goes out as deselect when the state it serializes has no selections. Before, five callers computed list.length ? 'selection' : 'deselect' themselves. fvOnDeselect still sends deselect.
• Spec version 0.6 → 0.7 (spec.py, and the example in the packaged flexviz-explore skill). One bump for the breaking changes of this release: unknown color_scale names raise (fix(heatmap): resolve color_scale to the plotly.js built-in names so the default draws viridis #111), two heatmaps and a corr_heatmap color_norm are refused (refactor(figure): allow at most one heatmap trace per figure #112, fix(server): validate decoded specs at /view, /h and /share #114), a stored log viewport is in data units (fix(adapters): keep log axis viewports in data units #113), and the event format (refactor(events): remove event.selections and read state.selections #115).
• Architecture.md: the event fields and the engine signature.

Impact

• The selections travel once per request. Before, they travelled in both state.selections and event.selections.
• Every caller sent the same list in both places, except in one race: the second request of a restore and a cross-filter mode switch captured the list before an await. If the user changed the selections during that await, the old code filtered by the stale list while the spec carried the new one. Now the server filters by the state it receives.
• An old client that still sends event.selections gets no error. The server ignores the field (the model does not forbid extra fields). A spec of version 0.6 is refused, as for every other version change.

Tests

• New test_state_selections_filter_the_request (tests/test_integration.py): a request without event.selections is filtered by state.selections, and a stale event.selections is ignored. Both cases fail on main.
• New test_remove_selection_posts_type_matching_state[plotly|echarts] (tests/test_browser.py): removing one of two selections posts selection with one left; removing the last posts deselect with state.selections == []. It passes before and after the refactor.
• test_bar_deselect_clears_filter now sends the deselect with a selection still in the state, so the deselect guard is tested.
• Engine, server, integration, cube and browser tests pass the selections through the state instead of the event. The 4 removed assertions have equal replacements.
• Mutation check: with the server passing [] instead of state.selections, 25 tests fail.
• pytest tests flexviz_polars/tests: 2427 passed, 2 skipped.
• pytest benchmarks: 31 passed.
• pytest -m browser tests: 341 passed, 15 skipped (before the last commit, which splits one if line and removes a source-text test). Local browser runs afterwards were unstable with page-load timeouts on main too; CI is the reference.

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing refactor/event-selections-from-state (e3a5194) with main (d03f887)¹}

Footnotes

1. No successful run was found on main (e3a5194) during the generation of this report, so d03f887 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩