docs: measured stock-vs-tuned benchmarks (hashrate + efficiency, mining live)

[VijitSingh97]

What

Adds a Benchmarks page + a README "Does it actually help?" highlight answering the obvious question — how much does RigForge's tuning actually move the needle? — with numbers measured mining live on a real 7800X3D (not synthetic --bench).

Configuration	H/s	Watts	H/s/W
Stock XMRig (no HugePages/MSR)	10,416	86.8	120.1
RigForge — system tuning only	10,756	83.6	128.7
RigForge — tuned	10,779	83.5	129.2
Stock → tuned	+3.5%	−3.8%	+7.6%

How it was measured

Each config ran as its own XMRig process mining to the live pool; hashrate (XMRig API, 60 s avg) + CPU-package power (Intel RAPL) sampled every 20 s over 5-minute windows, rotated across multiple rounds (~45 samples/config). Variance was negligible (hashrate within ~0.1%, power within ~0.3%).

Getting a truthful baseline took real care, and the PR is deliberately honest about it:

• The prefetcher MSRs persist, so a naive "bare" run inherits the previous tuned mod — I rebooted to firmware-default MSRs and reset them before every bare window.
• Ubuntu 24.04's Transparent HugePages quietly help the stock run, which is why the gap is ~3.5% on this modern box and not the 20–30% older write-ups quote. That caveat is in the doc, not hidden.
• Efficiency and performance tuning produced the identical config — because RandomX pins this chip at ~84 W in any all-core setup, so there's no hashrate-for-watts trade-off to make. RigForge measured that and correctly didn't invent one.

Notable finding

Stock XMRig draws more power (86.8 W) for fewer hashes — without HugePages the CPU stalls on memory (TLB walks). RigForge is faster and cooler, so the efficiency gain (+7.6%) exceeds the raw-speed gain (+3.5%).

Scope

Documentation only — no code change. New docs/benchmarks.md, README highlight + docs-table/index links, CHANGELOG. make lint clean. The measurement rig has been restored to normal tuned mining.

🤖 Generated with Claude Code