…on centering

Three changes wrapped in one PR because they share the same files
and shipped in the same iteration.

1. Group-by selector
   - New <wpd-select> on the toolbar next to the post-type chips.
     Options: No grouping / Category / Tag / Author / Year / Year-month.
     Per-facet self-describing option text ("Group by category", …)
     so the control reads as a single-line peer of the filter chips
     instead of a stacked label + select.
   - Session-local state — selector resets to None on each window
     open (no per-user persistence in v1; that's a follow-up if the
     ask surfaces).
   - REST /nodes payload extended with per-node author_id, year,
     year_month, category_ids, tag_ids and a top-level groups
     catalog of { authors, categories, tags } scoped to entries
     referenced by visible nodes. Bulk wp_term_relationships JOIN
     populates the per-post arrays in a single query.
   - New cluster-attractor force in ForceSim: emergent centroids
     (running average of member positions per group key); each
     non-pinned node receives a 1/N-weighted pull per membership,
     so multi-term posts settle at the force balance between
     centroids instead of committing to one. Date facets pin
     centroid X to a chronological lattice so year / year-month
     clusters lay out left-to-right oldest-to-newest.
   - New groupLabelLayer in GraphScene paints a translucent
     per-facet-tinted pill at each non-empty cluster centroid:
     Category=blue, Tag=green, Author=purple, Year/Year-month=
     orange. White text + slight transparency so node titles
     behind the pill remain visible and cluster labels can't be
     confused with node labels (dark text on white).
   - Picking a facet pre-builds per-node target positions on a
     polar seed (unordered facets) or chronological lattice
     (date facets), then lerps each node to its target over
     ~450ms with an ease-out cubic. The sim is paused during the
     tween so the layout mechanisms don't fight, then a small
     reheat lets the cluster force polish the final positions.
     Camera fit-to-view runs against the target bounds so the
     zoom animates in parallel with the layout flow. No more
     "everything jammed in the middle until you drag a node".
   - Cache invalidation extended with set_object_terms so
     retagging a post outside the post-edit path doesn't leave
     stale clusters behind the 6h TTL.

2. Click-to-deselect
   - Second click on the focused node now clears the focus, hides
     the panel, and lets the layout relax. Mirrors the gesture
     used to open it instead of forcing a trip to the close button
     or empty canvas.

3. Focused-post icon centering
   - The admin-post dashicon's bbox-centred anchor placed its visible
     pushpin glyph in the upper-left of the focused-node disc. New
     ICON_NUDGE table applies a per-icon fractional offset
     (admin-post = +6% x, +6% y of fontSize) so the glyph reads as
     centred at every node radius.

Tests
   - Vitest: 6 cases in content-graph-sim.test.ts (convergence,
     separation, multi-membership force balance, null disables,
     pinned exempt, chronological lattice). Suite green: 1283/1283.
   - PHPUnit: 6 cases in contentGraphGroupBy.php (per-node fields,
     empty-arrays, multi-term, catalog scoping, author names,
     retag cache-bust). Not run locally — port 8890 was held by
     a sibling worktree's wp-env — CI will catch any regression.

Polish + a real bug fix on top of the group-by feature.

Bug: Pixi v8 multi-Application destroy race
  Repro: open Content Graph → open Posts → click the Categories tab
  (loads the categories mindmap, a second Pixi.Application on the page)
  → close Posts. Result: CG canvas froze + the console looped forever
  with "Cannot read properties of null (reading 'clear')" thrown from
  `Batcher.break()` inside Pixi every frame.

  Diagnosis: tearing down one `Pixi.Application` while another live one
  is on the page corrupts the surviving app's batcher pipe map. Tried
  and rejected: `destroy()` without `texture: true`, deferred
  `setTimeout` destroy, `sharedTicker: false` on the CG app — the race
  lives below all of them.

  Fix: don't call `app.destroy()` at all on the secondary apps. Stop
  their ticker, detach the canvas, let GC reclaim. Applied to
  categories-mindmap, tags-cloud, and the heartbeat widget. Leaks
  ~hundreds of KB per open/close cycle — acceptable.

Belt-and-braces on Content Graph
  - `sharedTicker: false` so other bundles' Pixi apps can't drive CG's
    renderer.
  - `webglcontextlost` listener on the CG canvas → stops the ticker if
    Chrome ever drops the GL context.
  - Wrap `app.renderer.render` in try/catch → if a future race slips
    past the above, we stop the ticker and log one warn instead of
    looping rAF forever.
  - Zero-dimension guard on the resize observer so `renderer.resize(0,
    0)` (which happens during window layout shuffles) can't put the
    renderer into the broken state.
  - Cluster labels rendered as DOM elements over the canvas, not as
    Pixi children — sidesteps the batched-renderer fragility for our
    own per-frame label painting too.

Polish
  - Satellite icons: split the `term` kind so categories render with
    the folder icon and tags with the tag icon (both were showing the
    tag glyph and reading as the same kind).
  - Year-month grouping: alternate clusters above/below the
    chronological axis (`groupOrderStaggerY`) so ~60 month buckets
    don't crowd into each other on a single line.
  - Chronological lattice for date facets: `groupOrder` array on the
    sim pins centroid X to a left-to-right oldest-to-newest layout.
  - Various smaller hardening touches on `setGrouping` / `destroy` /
    `clearGroupViews`.