dsh-orquestrator 0.8.3

Three fixes to the model area (asked right after 0.8.2): the score follows the effort select, the effort
select is legible in both themes, and Cancel/✕/Esc abort the send.

• The score follows the effort. The badge re-resolves whenever the effort select changes. What the data
  supports (verified across all 469 models of OpenRouter's public catalog): the intelligence index is ONE
  scalar per model and does not vary with effort — so an Intelligence badge is constant across levels, which
  is exactly the "show the intelligence at max" fallback asked for. Terminal-Bench 4 IS per effort: a model
  whose leaderboard rows have the chosen level shows that accuracy (GPT-6 Astra: low 50.6 → medium 54.2 →
  high/xhigh 57.9 → max 58.2), otherwise the value at the model's best level (GLM 5.3 has a single max row:
  41.8% at every level). The snapshot keeps the per-effort rows now (scripts/gen-bench.mjs regenerates it).
• The select is legible. The effort select no longer renders white-on-white: explicit background, text,
  border and option-list colors from DSH's tokens and a color-scheme that follows the app theme — contrast
  15.0:1 dark / 18.9:1 light (5.8:1 disabled), focused and unfocused alike.
• Cancel means cancel. Cancel, the ✕, Escape and a mask click now ABORT the send: nothing goes out, no
  bubble appears (the optimistic echo is abandoned), the typed text stays in the composer as a draft, and
  nothing is stored. Only "Send with these options" sends. (Those paths used to send the message with stock
  behavior — the 0.6.0 rule.)
• Tests: 900 on the Mac mini (39 data-layer, 108 dialog/gate). In the browser: the skill script 202/203
  (the single miss is S10 in a reused home — a test-state artifact, not the product) and every legacy phase
  green (cancel 29/29, effort 25/25, light 7/7, small 10/10, confirm 14/14, workflow 11/11). The three
  behaviors were re-verified independently against the DOM, the POSTed configs, the session logs and
  screenshots — including a live badge that moved 58.2/50.6/54.2/57.9/57.9/58.2 through the picker on a
  temporary multi-effort model.

Compatibility: none on the wire — 0.8.0 through 0.8.3 halves mix fine.