docs(mdcode): rewrite the actions guide for a human reader

[libei]

The actions guide passed the house-style checklist — sentence-case headings
within the length caps, spaced em dashes, no nested lists, no SHAG words — and
still read as machine-written. The checklist was never the failing thing. This
rewrites the prose for a reader who wants to declare an action and does not yet
know how.

What changed

The register. kcmd, push and the judge are the subjects of their own
sentences; the personified "we" is gone, 50 occurrences to 0. That matches the
rest of docs/semantic-model/, which uses "we" nowhere. Contractions stay,
and so do the long compound sentences — chopping those into staccato is itself
a tell.

The opening. The guide starts on what a reader can't do today: an agent
that can ask but not act, reaching writes through hand-written tools the model
never sees. The feature is named after the problem, not before it.

Prose that circled a point instead of stating it. This is the bulk of the
work and the class the review kept landing on — explaining by analogy, by
contrast, by rebuttal, or by abstraction before the reader has the thing
itself. Roughly sixty passages, across every section:

• Sections that opened on a debate now open on the object. "An executor is
  opaque", three sentences on what a reader cannot see, and then "declare the
  blast radius" said twice, becomes "affects names the concepts a call
  writes to."
• Rebuttals to designs nobody proposed are cut. "Asking you to repeat it would
  add a second place to get the answer wrong" says what concept takes
  instead.
• Fourteen X, not Y antitheses become "rather than" or a plain sentence. An
  earlier pass in this branch introduced them; the construction appears
  nowhere in the reference prose it was supposed to match.
• Wh-clefts where a plain verb works, colon splices gluing a claim to its own
  gloss, and prose sentences over forty words that carried more than one idea.
• Abstract restatements of an example already on the page. "A store is typed
  by its backend:" led the field and the three values that immediately follow
  it; the field leads now.

Five factual corrections, each found by reading the source before moving
the text:

• Section 1 claimed a push prunes actions the selected binding can't perform.
  It doesn't — kc_actions.ts publishes an entry for every action, with the
  executor fields absent when nothing is wired up. The true behaviour is now
  in section 5, where publishing is described.
• on_violation read as though omitting it were legal with a default. It's a
  push error (validate.ts:520), and "an unmarked constraint rejects" is the
  reason the key is required, not a fallback.
• "The gap is loud... and it's still a gap" conceded its own concession. A
  refusal being easy to spot and the rule still going unchecked are two
  separate facts; both are stated.
• A rule variant that does not exist was described in the present tense.
• The judge's read wrap is bounded per statement, but bounds neither which
  tables a read reaches nor a function that writes. Both limits are now
  stated where a reader would look for them.

What did not change

• Every fenced block is byte-identical to main. All 564 lines of YAML,
  SQL, diagrams and captured CLI output, verified after every batch by
  diffing extracted fences.
• One heading changed, ### What an entity-typed parameter adds to
  ### Parameters typed by an entity, because the original argued with
  hand-written tool schemas instead of saying what the thing is. It carries no
  inbound anchor. The five inbound anchors from reference.md,
  model_spec.md and the agent demo README all still resolve, as does every
  internal one.
• No status note or stated limitation was weakened. The three "Status:"
  notes and "What is not modeled yet" carry the same content; five of them
  now carry more.