Condense SPEC.md

[BY571]

456 → 402 lines, 3,599 → 3,188 words.

A smaller cut than the README's 48%, and deliberately so. Most of SPEC's length is
evidence and recorded decisions — the prior-art table, the 2019 survey finding that no
research ontology models hypothesis death, the seven-hypotheses-six-died dogfood result —
and cutting those would delete the reason the code is shaped the way it is. What came out
is duplication and staleness.

What was actually redundant

§8 stated the same ~2,340-token measurement twice — once under "The measurement" and
again in the paragraph explaining the deletion. Each was written at a different moment
(demotion, then removal) and each appended rather than edited, so the section had become
three layers of sediment arguing one point. It now reads as one argument.

§5's table had two columns that paraphrased each other — no_signal was glossed as
"the effect is not there" and "a signal that wasn't there" — and it re-printed two rule
blocks that §6 already carries in full. It keeps the vocabulary and the argument for why
the core must not enforce it; §6 keeps the rules.

§6 merged two example rules that differed only in which key they demonstrated. The key
table — the section's actual reference value — is untouched.

§4's derivation subsection and §8.1 lost restatements, not content: both had grown a
paragraph re-making the point the paragraph above already made.

Two stale things fixed in passing

• §11's open question 4 asked whether a closed cause vocabulary should become a rule
  primitive. require_field_one_of shipped some time ago and a second graph did want it —
  struck through and answered, like question 2 beside it.
• §3's node-type paragraph had accumulated a run-on sentence across two edits.

Verified

285 passed. test_docs.py still extracts and executes every node and rules block from
SPEC against the real example graph — the merge of two rule examples into one is why the
count moved from 286.