docs: run it and read the output, and what that caveat costs

The handoff recorded two greps — comment markers and `unwrap_or` fallbacks —
as the way defects get found. Neither found anything today. Every serious
finding since the tag came from running the thing and reading what it said,
then checking each claim against the machine with a second tool.

That method is now written down with the evidence: a report asserting "No
SSD detected" at 0.85 on a host with three NVMe drives, a desktop tower
classified as a GamingLaptop, a Ryzen 9 9900X dated to 2018, a 24GB card
read as 4GB. None of those are visible in a diff, and none would be caught
by a test asserting that a report is produced. They are only visible when
someone reads the sentences and asks whether each is true here.

Three rules alongside it. Check ground truth in a second tool before
believing a fix — the one plausible reading that went unchecked this session
was the 0.61-second boot duration, which was two timestamps of the same
instant subtracted from each other. A fallback is how a module fails
completely and still looks like it works — `unwrap_or(0)` hid a parse that
returned `None` for every disk on every Windows machine since the module was
written. And watch for a guess placed in front of a reading, which happened
three times in one day: chassis inferred from a battery while Linux read
DMI, a battery weighted above a read chassis in the classifier, and Secure
Boot bound to a reader needing Administrator while the value sat in the
registry.

Open work now opens with what the `hardware_ai` audit does not cover. Every
conclusion corrected was verifiably wrong on this desktop and each fix was
checked against a second source, which is not the same as being right in
general. The classifier weights were rebalanced against a single data point:
the direction is defensible, the numbers are not established. The entry says
what would settle it — run `full_analysis()` on a real laptop and read the
report; if it does not say Laptop, the weights are now wrong in the other
direction. `adapter_vram_bytes` has likewise only met NVIDIA and AMD
adapters in one box.

Commit table for the six since the tag, and 7.0.0 item 6 for the historical
metrics removed in 4dd5cce.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>