The answer's metric names identify the document

[torstei]

--using was the only route to a unit when the window held no !meta — which
meant typing the name of a document whose metrics are already listed in the
answer. facts_for looks it up instead.

Three sources, in order of authority: the document using named (an
instruction), then a !meta in the window (the producer's own word about these
numbers), then the documents installed here. So using becomes the
override rather than the normal path.

Live, with no !meta and no --using — the inferred requests reaches the
right axis and splits the plot, which is the unit-decides-the-axis rule working
off inference alone:

                    http_bytes, http_requests
 4000 +--------------------------------------------+ 4
 2500 |-+........:..........:.........:.....##.+-| 2.5
 1500 |-+........:..........:.........:...##..count (requests)
    0 +------------------------------------------+ 1
    12:01      12:02      12:03     12:04      12:05
         http_bytes *******  http_requests (right) #######

⚠ Inference degrades to NO unit, never a wrong one. Where two documents
declare one metric in different units, both are named and the unit is left
blank. The unit decides which axis a series lands on, so guessing between ms
and s would not mislabel a plot — it would regroup it.

That is not hypothetical. The shipped set today:

http_bytes     timberfs-apache-combined=B  timberfs-nginx-combined=B    <-- several
http_latency   timberfs-nginx-combined=s
http_requests  timberfs-apache-combined=requests  timberfs-nginx-combined=requests

The two overlapping metrics agree, so the ambiguity is currently harmless — and
checkable rather than assumed. But http_latency is nginx's alone at s
while apache logs %{ms}T in milliseconds, so the first time
timberfs-apache-combined grows a latency metric (the field is already in the
log format it parses) a mistyped using would silently re-axis the plot.

Two things now said that were silent: a using naming a document that
declares none of the answer's metrics, and a using that disagrees with a
!meta.

Nothing new in the bark, nothing that travels, no dependency on the producing
host being alive — the only requirement is the one this rests on anyway: the
local documents resemble the server's. When they don't it degrades to no units,
which is exactly today's behaviour.

The harness bug this turned up

tests/timbergraph/test-timbergraph's main() called fn() unguarded, so one
test raising aborted the whole run and every test after it never ran — the
failure mode a suite exists to prevent. It hid the first draft of this PR's own
test (a KeyError that read as a clean pass in a truncated check). Now caught
per test, as the timbersh suite already does. Verified by sabotaging
declaring():

FAIL test_the_documents_installed_here_identify_the_metric: KeyError: 'unit'
timbergraph: 1 failure(s) of 16 test(s)

New in timbergraph.py: declaring(metric) and facts_for(metrics, markers, using), which replaces the three-line facts/units dance duplicated in
timbergraph's main() and timbersh's graph.