fix(app): stop the parked tone that stood in for the bed while backgrounded

[khagele]

Closes #568

Backgrounding in full stopped the bed and the generative music and started a held tone in their
place. That tone is withdrawn; hidden now starts nothing.

Why

• Nothing masked it. The bed is already stopped by the time it starts, so it was the only
  continuous sound in the mix — and since app: the ambient soundbed is thin and static, give it space and movement #496 dialled it from a C3 sine at 0.025 into a B3
  square with three partials at 0.045, swelling and breathing at 1.2 Hz, the one voice that
  carries no information by design was the most attention-seeking shape in the engine.
• The pocket is not silent without it. cue() and txBlip() have no visibility guard, so
  every reception and every transmission keeps sounding while hidden. "Parked, not dead" was
  already being said by what does carry information.

What changed

• app/src/sound.js: BG, startBgAmbience() and stopBgAmbience() removed, with their call
  sites on the hidden path, the return path, setMode() and destroy(). The hidden branch is now
  if (mode === 'full') { stopBed(); stopMusic() }.
• Removing startBgAmbience() also takes its ensureCtx() off the hidden path, so nothing arms a
  pointerdown listener while the page is not visible.
• docs/2026-07-16-sound-modes.md: an addendum dated 2026-08-31 withdrawing the tone, and a
  pointer to it at the top of the 2026-07-27 addendum that introduced it.
• changelog.json (both copies): one entry, 16 words.

Unchanged: the transition cues in either active mode (falling on backgrounded, rising on resumed),
the hard stop of bed and music on hidden, the restart of both on return, and every other mode.

Tests

soundengine.test.js loses the four tests that pinned the tone's existence, shape and suspended-
context guard, and gains two:

• leaves nothing running while hidden — only the transition cue sounds — counts started
  oscillators after backgrounding in full: two, the cue's own notes. Mutation-checked by putting
  a bare held oscillator back on the hidden path (expected 2, got 3).
• restarts the bed itself on return, not only the music — counts buffer sources rather than
  oscillators, because the bed is looped noise and the music is oscillators, so an oscillator count
  cannot tell a return path that forgot startBed() from one that did not. Mutation-checked by
  dropping startBed() (expected 3 to be greater than 3).

starts nothing while the context is still suspended is kept, retitled: it now covers the whole
hidden path rather than the ambience alone.

Verification

• npx vitest run in app/: 893 passed (48 files).
• npm run build in app/: clean.
• npx vitest run in web/: 509 passed, including parity.test.js on the two changelog copies.