feat(studio): interactive dilution equation, curated backgrounds, custom bimodal modes (schema 0.5.0)

[aliakherati]

Interactive dilution, full-profile hover, curated backgrounds, custom modes

Four review requests in one message; two stacked PRs deliver them. This branch carries both commits
(the first is also PR #98's base): schema 0.4.0 → 0.5.0.

The dilution tab is now the equation (first commit)

The model's two-piece form is displayed with the selected regime's constant beside it:

V/V₀ = t^0.8              (t ≤ 10⁴ s)
V/V₀ = 1585 · exp(k · (t − 10⁴)^1.5)        k (s⁻¹·⁵) = 8.890e−9

• Regimes are chips on the graph, each carrying its k — introspected from
  coupled/dilution.py's own segment tuples, never re-typed (a test compares against the same
  tuples). Clicking a chip sets dilution.regime for the run — choosing on the graph is
  choosing.
• "try k" draws any k in [1e-10, 1e-6] as a dashed gold curve, computed by the model's own
  _two_piece/_eval_segment machinery. At a named regime's k the custom curve equals that
  regime's bit for bit — the test that proves it is not a lookalike formula. Clearly marked a
  picture: the model runs named regimes or a constant rate.
• The ERA5 profile hover reads the whole level — p · z · T · H₂O (p 50 hPa · z 20.8 km · T 211.4 K · H₂O 4.20 ppmv), tracking along the pressure axis. Also retires a latent defect:
  the x-tracking hover assumed sorted xs, which a temperature profile does not have.
• Sub-decade gridlines on log axes (2–9 per decade), faint and dashed.

Curated backgrounds — hidden is not removed (second commit)

Picker now offers: sabr_330, sabr_310, sabr_220, aer_geo (the spec's GEOENG reference — Pierce
et al. AER 2D), custom. Hidden: redcircles, cesm_g6, cesm_g6_amb — still valid, because Tier
B's archived cases use cesm_g6, and a schema that refused them would disconnect the archive from its
own configurations. SciField.hidden_choices is what the generator filters on; an archival value
still displays, marked "(archival)".

Custom bimodal distribution

(N₁, Dp₁, σ₁) + (N₂, Dp₂, σ₂) with an explicit STP/ambient basis — required because the
bridge refuses a custom mode list without one (SCIENCE-3's per-dataset basis trap, made a field).
Defaults are SABR-220's mode with N₂ = 0: custom starts citable and unimodal, and with untouched
defaults it seeds the same bins as sabr_220, asserted bin-for-bin. The panel redraws live as you
type, through the same to_scenario → initial_tomas_state path the run uses.

Found by test at the seam: the model refuses zero-N modes (N must be > 0), so the seam filters
them — N₂ = 0 is a one-mode list, not a zero entry — and refuses the all-zero case with a message
naming the fields.

Also

max_wall_time_s no longer cites "BLOCKING-4" bare — raised in review as confusing, which it was.
It now says what the cap does and where the hour came from.

371 Python Tier-A, 60 vitest, both lanes green. Verified live in the browser: chip click changed
the run's regime; typing N₂ = 8 drew the bimodal distribution on the spot.

Stacked on #97 → #96. Merge bottom-up; rebase on request.