A firefight does not crackle: one output bus with measured headroom, and a limiter on it #211

Five players firing in one room is louder than the output can carry, and what a player hears then is not a firefight but a crackle.

Every voice in src/engine/sound.ts connects straight to context.destination, so what reaches the speaker is the raw sum of whatever happened this tick. The recipes in src/client/audio.ts are not small: a grenade is a 0.7 burst plus a 0.55 tone, a flash 0.55 plus 0.3, each gunshot 0.35 plus 0.3, and a round's worst instant stacks several of them inside a few milliseconds. Nothing in the device knows there is a ceiling at 1.0.

Done looks like

• [ ] Every voice the device starts ends at one master GainNode created with the context — not per voice — and that bus reaches the destination through a DynamicsCompressorNode used as a limiter. Nothing connects to context.destination directly any more.
• [ ] The headroom is arithmetic, not taste: the docstring states the loudest plausible instant summed from the recipes actually in the repository (the numbers, one line of addition), what the bus multiplies it by, and what the limiter is left to catch.
• [ ] The limiter's threshold, knee, ratio, attack and release are named constants with a sentence each. A wrong attack turns every shot into a pumping mess, so the pull says what a single gunshot's peak was before and after.
• [ ] A quiet round does not get quieter than the stated headroom: one gunshot alone is within that of what it is today, and the pull gives both numbers.
• [ ] Muting stays free: the muted early return stays where it is, so a muted page still schedules nothing and pays nothing per shot.
• [ ] A context with no createDynamicsCompressor plays through the gain alone instead of throwing, with a test on that path.
• [ ] Tests would fail without this: the fake contexts assert every source ends at the bus and the bus at the destination, and a budget test sums the recipes' worst case and fails when a future recipe pushes it past the number in the docstring.

Files I expect to change

src/engine/sound.ts, src/engine/sound.test.ts, the fake AudioContexts in src/client/stage.test-helper.ts and src/range/stage.test-helper.ts, src/engine/GUIDE.md, and src/client/audio.ts only if a recipe turns out to be the thing that is wrong.

Risks

• It waits on the panning issue: both rebuild the node path between a voice and the destination, and doing them at once is a conflict for nothing.
• The honest outcome may be smaller than the title. If the measurement says the sum never crosses 1.0 in a real round, ship the master gain and the budget test, say the limiter was not needed, and leave the compressor out — a number that says "no" is the deliverable too.
• The range's device is the same engine device: whatever the bus does, src/range/audio.ts gets it, so the pull should say what the range sounds like after.

Suggested: Torque — this is a measurement with a fix attached, and the docstring has to hold the arithmetic.

Parked by the crew: the check failed and the decider saw no point in another try (pull #273 closed). Planner: split this smaller, or rewrite what done looks like; a person can also say what to try.