Correct the speculative-decoding crossover in inference-optimization

[theFong]

The skill shipped a "speculative decoding has a concurrency crossover" finding —
a win below C=4, a 3.1x loss at C=32 — and recommended shipping two serving
profiles. That was measured on a faulty GPU and is wrong.

What happened

The original numbers came from a DGX Station whose GB300 was power throttled:
HW Slowdown asserted, module power limit capped at 950 W against a 1600 W
default, drawing 190-240 W where a healthy peer drew 600 W. The skill's own
hardware-validation.md documents that machine — but the crossover had already
been measured on it and written up before the fault was found.

Re-measured on healthy hardware

Single GB300, k=8, max_num_seqs=32, OSL=256, aggregate decode tok/s:

ISL	C	Spec ON	Spec OFF	Winner
256	1	278	158	ON +76%
256	8	1198	866	ON +38%
256	16	1875	1431	ON +31%
8K	8	591	409	ON +44%
8K	16	654	456	ON +43%
32K	8	225	92	ON +145%
32K	16	218	112	ON +95%
128K	8	46	35	ON +29%
128K	16	54	37	ON +45%

No crossover up to C=16. Speculation wins every case by 29-145%. Its only
cost is KV pool: freeing the draft cache gives +23% (6.0M -> 7.4M tokens), which
does not come close to paying for the throughput.

On a starved part the k+1 verify batch is extra work the GPU cannot absorb;
with headroom it is nearly free and the accepted tokens are pure gain.

Changes

• reference/vllm-tuning.md — replaces the crossover section with the healthy
  hardware table, and keeps the old result as a cautionary tale rather than
  deleting it, since the failure mode is the instructive part.
• SKILL.md — the Phase 3 tuning bullet no longer asserts a crossover.
• reference/methodology.md — adds the generalized lesson: a hardware fault
  invalidates conclusions drawn before you found it, so re-validate everything
  measured on that machine, not just the experiment in flight.

The plausible mechanism ("verify batch exceeds the tuned kernel range") is what
made a hardware artifact look like a software law. That shape warning is real,
but on healthy hardware it was never large enough to flip the decision.

Validation

./test.sh passes (including the frontmatter check added with the skill).