fix(docs): point the landing page at the real 8192 reference regime (Fixes #248)

[luciferlive112116]

Fixes #248.

PR kind

• fix
• feat

Summary

index/index.html still advertised the pre-8192 reference regime. The engines
and every other surface say 8192, so the project's front door disagreed with the
code it describes:

surface	says
strategy/cli.py / matmul/cli.py	--n default 8192
README.md:33, BENCHMARKS.md:3	"Reference setup: 8192 × 8192"
README.md:37, BENCHMARKS.md:58	python -m eval --n 8192 --pairs 3
index/index.html (before)	N = 12,000 — CCO's reference regime, tag N=12000, python -m eval --n 12000 --pairs 3

Three stale claims, all re-derived from the shipped default:

spot	before	after
KPI card	3.5 TFLOP (2·12000³ = 3.456e12)	1.1 TFLOP (2·8192³ = 1.0995e12)
KPI sub	N = 12,000 — CCO's reference regime	N = 8,192 — …
arena tag	N=12000	N=8192
quickstart	python -m eval --n 12000 --pairs 3	--n 8192

The KPI was right for the stale N, which is exactly why it had to move with it.
The neighbouring N=1,000 → 2 GFLOP and N=128,000 → 4.2 PFLOP cards are already
correct decimal-N (the #102/#230 convention) and are untouched.

Same class as the landing-page repair in #128.

Changes

• index/index.html — the three stale claims above (+4/−4).
• tests/test_landing_page_reference_regime.py (new) — CPU-only: re-derives the
  page's advertised N, quickstart command and KPI from strategy/matmul's --n
  default (asserting both engines agree), so the front page cannot silently drift
  from the shipped default again.

Validation

$ python -m pytest tests/test_landing_page_reference_regime.py -v
  4 passed

$ python -m pytest tests/ strategy/tests/ eval/tests/ -q     # CI-equivalent (no torch → GPU tests skip)
  315 passed, 24 skipped

No protected paths touched — index/ is outside the guard regex / CODEOWNERS set
(eval/, docs/, .github/, dashboard/).

Result

Documentation-consistency fix on the landing page; no code path changed, so no
accuracy / latency / VRAM delta.

metric	value
accuracy	unchanged (static page + consistency test only)
time complexity	unchanged
latency	N/A
VRAM usage	N/A

[zeokin]

Approved. The landing page had stale 12000 reference-regime claims despite the shipped 8192 default. This corrects the associated FLOP display and pins the page to both CLI defaults with regression tests. Merged-state tests, strategy smoke, and the full CPU-safe suite pass; required GitHub checks are green.