Close the pre-registered question ledger

[jadoon200]

Three of the ten registered questions were unaccounted for. For a project whose central claim
is "pre-registered questions, answered whichever way they fell", that's the gap most worth
closing.

Q9 — never measured, and the ROADMAP implied otherwise

generate_multi_resource existed and was unit-tested, but no experiment used it, while
docs/ROADMAP.md claimed the synthetic regime "carries the joint-provisioning test". It
didn't. Corrected rather than quietly deleted.

The registered direction is refuted. Independent per-resource sizing under-provisions,
and the sign follows from max(a,b) >= a without needing an experiment — the joint quantile is
never below either marginal, so combining independently sized resources can only buy less.

What's actually measured is the coverage it costs (target q = 0.90):

offset	correlation	joint cov.	indep. cov.
0.000	+0.778	0.9688	0.8442
0.250	+0.010	0.9688	0.8229
0.500	−0.770	0.9688	0.8016

The shortfall widens monotonically as the resources decorrelate — the more the peaks avoid
each other, the more often the resource you aren't watching is the binding one. You cannot
reach a joint SLO from one dashboard per resource.

Three nulls first, all kept in the record

A null from an instrument that cannot detect the effect is not evidence of absence, so the
sequence is documented: peaks narrower than the tail a quantile discards; peaks landing in
only one half of the train/test split; and a median-based replica scale that let the
wider-swinging resource bind at every step. The tell each time was a null that was too
clean — identical to four decimals across conditions that should have differed. The generator
gained peak-width and diurnal-phase parameters, both defaulting to the original fixture.

Q2 — the comparator was a strawman

It measured conformal against an uncalibrated forecast. It now carries the fixed safety
margin it registered, both tuned and at the conventional +15% VPA ships:

Method	p95 coverage	Mean p95	Recovery
fixed_margin_conventional_+15%	0.917	113.563	46
aci_gamma_0.05	0.924	101.976	28

ACI holds better coverage on 11% less capacity and recovers in 28 steps where a fixed
margin never recovers — it can't, it doesn't respond to anything. The tuned margin is sharper
still: raw p95 already over-covers on stationary validation, so the honest tuned margin is
+0%, offering no protection whatever when the level shifts.

Q6 — measured, never labelled

The GPU structural claim is the answer to Q6. It was carried out under its milestone number
and never tied back — a bookkeeping failure in a project whose discipline is that ledger.
Labelled and shipped.

Result

Served findings go from 6 to 9. Makefile targets added for all four new experiments.

Gate

ruff, format, mypy, pytest, validate-simulator — all verified by exit code. 201 tests, M4
gate GREEN.