test: reference board + measured baseline for the failure knee (#74 step 0)

[pshenok]

Step 0 of the work on #74 ("bored in 10 minutes"), and deliberately no product code.

Why this comes first

The pre-implementation critique of the knee design produced five different "reference boards" and five different collapse points (10–20 rps). That is another way of saying none of the design's claims were falsifiable. A board described in prose is not a measurement.

tests/helpers/reference-board.mjs pins one: exact wiring and tiers, a fixed traffic mix (kept out of CONFIG so a survival balance tweak cannot silently move every baseline), mulberry32 seeding, and a frame loop mirroring animate()'s order and its game-scaled dt. sweepAt() reports only the quantities the design makes claims about — failures, min reputation, band residency, mean dwell, time from first drop to a reputation floor — so a claim can be falsified without writing new instrumentation. Harness guards throw on a board that did not build or a sweep with no RPS, because a silently dead sweep otherwise reports a clean board and a passing objective.

The measured baseline on main (60 s per point)

rps	failures	min reputation
4	0	100
5	3	97.7
6	7	95.3
7	54	59.3
8	165	−43.5
20	691	−556.8

Two properties, asserted — both true today, both meant to become false

1. The readable middle is a knife edge. This measurement corrected the design's own wording: a point where the board drops requests and still survives does exist (5–6 rps). What does not exist is width — 12.5 % of the collapse load, against a 20 % target. A player adjusting an architecture, rather than turning an rps dial, cannot find it. Width is the metric, not existence.

2. The band is a strobe. Mean dwell of the bottleneck's instantaneous utilization inside (0.90, 1.20) never exceeds 0.245 s at any load. No curve placed on this axis can produce a readable middle, because the state it describes does not persist long enough to perceive. This is why the design changes the axis (an EWMA), not the curve.

Also lands the design doc at rev 2

Rev 1's central mechanic — congestion latency — was killed by the critique with measurements, before a line of product code: multiplying processingTime feeds back into the very quantity totalLoad measures, and at the specced k=3 a board healthy at 8 rps went from 463 completed / 0 failures to 53 completed / 416 failures (stability ceiling ≈ 0.25–0.4). Its payoff was imaginary too — recordServiceSuccess is called only from finishRequest, so a congested intermediate node never records a latency sample at all.

Rev 1's errors are kept in the doc on purpose, including the gate-vs-switch ambiguity that would have made 16 shipped campaign levels unlosable (level 15's design comment says outright that its pre-built Compute "sits ~10 % over its own throughput").

826/826, eslint clean. Steps 1–5 (ramp interpolation → smoothedLoad → rings/alert → the knee switch → ASG) land as their own PRs, each verified against this baseline.