feat(06-02): honest ETA — per-axis EMA + independent display deadbands + arrival clock (GREEN)

- updateEta(prev|null, remainingS, remainingM): numbers-only signature
  (D-07 structural — no Fix can enter); first update seeds directly
- EMA alpha 0.25/fix per axis; display refreshes only past its axis
  deadband (+-1 min time / +-5% remaining, axes independent)
- composeEtaClock: 24-hour HH:MM instrument format (local wall clock)
- suite E1-E8 green: seed, jitter-holds-display, real-drop-updates,
  independent axes, deadband symmetry, pinned constants, purity gate

web/test/guidance-eta.test.mjs

@@ -78,7 +78,12 @@ test("E4 a genuine 5-minute drop updates within 2 fixes and converges to the tar
   // alpha 0.25 convergence: (1 - 0.25)^n < 0.1 at n = 9 — 12 fixes lands inside 10%
   for (let i = 0; i < 12; i++) s = updateEta(s, 300, 10000);
   assert.ok(Math.abs(s.etaS - 300) <= 30, `smoothed converges to the drop (got ${s.etaS})`);
-  assert.ok(Math.abs(s.displayedEtaS - 300) <= 30, `display converges too (got ${s.displayedEtaS})`);
+  // the display trails the converged value by at most the deadband — its
+  // whole contract (and why the instrument never whiplashes)
+  assert.ok(
+    Math.abs(s.displayedEtaS - s.etaS) <= DEADBAND_MIN_S,
+    `display within one deadband of the smoothed truth (got ${s.displayedEtaS} vs ${s.etaS})`,
+  );
 });
 
 /* ————— E5 — independent axes ————— */