feat(agent-coding-contract): start explanations from a zero-knowledge floor (v4.0.2)

[cheneeheng]

What

explain-until-understood now starts from the assumption that the reader knows nothing about the
subject, and ships as v4.0.2 (ceh-agent-coding-contract v2.9.2 → v2.9.3).

• New Who you are explaining to section — zero-knowledge default, a stated floor the reader can
  raise, and an evidence rule for raising it.
• New Plain language section — concept before name, every term of art defined at first use, the
  shorter word, one idea per sentence, a bounded analogy.
• Step 2 picks foundations for the reader's floor rather than for the subject.
• The escalation ladder gained a routing question separating a word miss, a stated-foundation miss,
  and an assembly miss.
• primitives renamed to foundations throughout; frontmatter description compacted ~20% with no
  facts dropped.

Why

The skill had seven steps about what to say and nothing about who is listening, so the model
filled the gap with its default reader — a peer engineer — and every other rule calibrated to that.
Step 2 was the one place a floor could have been set, but it scoped its primitives to the subject,
which selects the top-layer domain concepts and silently treats everything beneath them as known.
The only adjacent jargon rule banned reusing a term introduced earlier in the session, which permits
every term of art the reader has never seen.

The ladder had a matching gap: one dimension, representation form. A reader lost on a word was
routed to numbered steps — the same unknown vocabulary, renumbered.

How

Two commits: the skill change (375d4df) and the release entry (ac8759d). The release commit
rides in this PR rather than a separate chore/release-v4.0.2 branch, matching how v4.0.1 landed in
#77.

Testing

Three simulated runs against the revised text — a one-sentence ask, a jargon-dense subject (the
ceh-advisor failure-watch hook), and a follow-up miss. They caught three defects, all fixed before
the release commit:

• the floor line was unconditional and consumed half of a one-sentence answer;
• the two foundation-miss routes contradicted each other;
• step 2's comprehension checkpoint told the model to stop mid-reply, against the skill's own
  one-reply structure.

python tools/validate-plugins/validate.py green.

Behavioural lift is unmeasured. The runs were self-simulated by the author of the change, so
they establish that the text is followable and self-consistent — not that a cold model explains
differently. A blind A/B against the old text, or /ceh-evaluation on the skill, would close that.

Generated with Claude Code by CEH